Skip to content
jsonW Paper GitHub

Measurements

Numbers measured and published in the paper, with the method and the machine. Comparisons hold within the same run.

“JavaScript Object Notation (JSON) is a lightweight, text-based, language-independent data interchange format.”

  1. 1
    Takes 6.6 times less space than uncompressed JSONL, on 100,000 synthetic records.Source: paper
  2. 2
    Reading a whole column was 2,310 times faster than in JSONL (1.59 ms), median of 10 runs.Source: paper
  3. 3
    Writing 100,000 records took 131 ms, against 15,211 ms for JSONL with the same reader: 114 times less.Source: paper
  4. 4
    Reading the whole file took 78 ms, 46 times less than JSONL.Source: paper
  5. 5
    With 8 cores, reading the whole file took 9.0 ms: 419 times less than JSONL.Source: paper
  6. 6
    In the row form, one test record takes 1,073 bytes, against 5,157 bytes in JSON: 4.8 times less.Source: paper
  7. 7
    The full INPI trademark dataset (12,527.4 MB of CSV) became 2,579.2 MB in .jsw: 4.9 times smaller.Source: paper
  8. 8
    At maximum level, the same dataset became 1,938.1 MB: 6.5 times smaller than the CSV.Source: paper
  9. 9
    The INPI dataset in .jsw was 12.8 times smaller than in XML (24,804.7 MB).Source: paper
  10. 10
    At maximum level, 1.3.11 wrote the INPI dataset 16.5% smaller than 1.2.1 (1,938.1 vs 2,321.8 MB).Source: paper
  11. 11
    Counting the 2,343,607 INPI records in force took 43 ms, against 54 ms for Parquet read by DuckDB, both with 8 cores.Source: paper
  12. 12
    The same count took 375 ms with DuckDB reading the CSV: the .jsw was 8.7 times faster.Source: paper
  13. 13
    The same count with grep took 519 ms on the CSV and 3,017 ms on the XML.Source: paper
  14. 14
    Finding the 10 rulings of one case among 24.98 million took 207 ms; grep on the CSV took 6,690 ms in the same run.Source: paper
  15. 15
    For INPI class 35, jsonW counted 1,076,928; grep on the CSV counted 1,076,929, because the text ",35," appears in another column.Source: paper
  16. 16
    On one million INPI rows, the .jsw took 51.2 MB: 2.8 times smaller than CSV, 1.5 times smaller than XLSX and 9.8 times smaller than XML.Source: paper
  17. 17
    For the same query on one million rows, the .jsw took 119 ms and XLSX 9,364 ms: 79 times less.Source: paper
  18. 18
    The INPI conversion checked every typed value: numbers and dates had to turn back into the same CSV text, or it stopped.Source: paper
  19. 19
    The 7 INPI files, with 55,278,212 rows, became .jsw with nothing cut and the full history kept.Source: paper
  20. 20
    The case number column shrank from 53.0 to 25.7 MB with the fixed-width digits of 1.3.11.Source: paper
  21. 21
    The in-force count went from 677 ms (1.2.0) to 44 ms (1.3.11, default level) over five measured and published runs.Source: paper
  22. 22
    A filter found 9 rows out of 100,000 reading only 833, in 1.7 ms, because the zones skipped 102 of 103 blocks.Source: paper
  23. 23
    Opening a 79.1 MB remote file downloaded only 8.7 KB: the header and the footer.Source: paper
  24. 24
    Reading one row from the middle of that remote file downloaded 0.98% of it; a filter, 0.84%.Source: paper
  25. 25
    The paper publishes the earlier runs and the numbers that changed, instead of showing only the best.Source: paper
  26. 26
    Anyone can measure on their own machine with wjsw medir, which writes and reads the same set as the paper.Source: paper
  27. 27
    The paper's example files are published at jsonw.org/paper/exemplos.Source: paper
  28. 28
    Every table's numbers come with the machine, the date, the reader and the number of runs.Source: paper

Costs and time

Arithmetic based on the measurements above. Where a number is derived, the calculation is in the text.

  1. 29
    Storage: for INPI, the .jsw at maximum level takes 15.5% of the CSV (1,938.1 of 12,527.4 MB). For storage billed per GB, the bill drops in the same proportion.Source: paper
  2. 30
    At the ratio measured on INPI, 1 TB of similar CSV becomes about 155 GB of .jsw at maximum level.Source: paper
  3. 31
    Egress: reading one row of a remote file downloaded 0.98% of it, so paid traffic drops by the same measure.Source: paper
  4. 32
    The file can live in Cloudflare R2 or any S3 and be read straight from the bucket, with no application server in between.Source: README
  5. 33
    No separate vector database: cosine runs on the file itself.Source: paper
  6. 34
    No separate search engine for term search.Source: paper
  7. 35
    No separate permission layer: the view comes from the schema and is applied by whoever serves the file.Source: paper
  8. 36
    No separate audit table: the history of edits and deletions stays in the file.Source: paper
  9. 37
    License: free use in your product, on as many servers as you want.Source: README
  10. 38
    No dependencies: the core uses only C and the standard library.Source: README
  11. 39
    CPU: reading only the requested column avoids decoding the rest of the record on every query.Source: paper
  12. 40
    Memory: the .jsw is read with mmap, and several processes share the same file in memory.Source: docs/JSONW.md
  13. 41
    A vq8 vector takes a quarter of a vf32 (8 vs 32 bits per number), and a vq4 an eighth.Source: docs/JSONW.md
  14. 42
    A vf16 vector takes half of a vf32.Source: docs/JSONW.md
  15. 43
    Fewer conversions: the same model serves messages, files and text, and moving between forms loses nothing.Source: paper
  16. 44
    Less rework: a cut write loses only the last block, not the whole load.Source: paper
  17. 45
    No splitting a dataset across several spreadsheets: .jsw has no 1,048,576-row limit.Source: paper
  18. 46
    Bringing a Postgres, MySQL or SQLite table with the right types is one command (wcon), with no hand-written schema.Source: AGENTS.md
  19. 47
    Converting the 12.53 GB of INPI CSV took 177 s on one core.Source: paper
  20. 48
    Similar-spelling search needs no phonetic search service: it is a filter operator.Source: Newsroom
  21. 49
    The compression level is chosen per file (0, 1 or 2): a file queried all the time uncompressed, an archive file at maximum.Source: docs/JSONW.md
  22. 50
    wjsw regravar moves a .jsw to the new version's codecs without going through the CSV again.Source: paper
  23. 51
    Parallel reading by block uses the machine's cores with no configuration.Source: README

Artificial intelligence

What changes for whoever trains, queries or orchestrates models.

  1. 52
    The embedding and the data live in the same file, so a semantic search returns the whole row with no second lookup.Source: paper
  2. 53
    The JSON scoring request is the same in the tool and in the ABI, so an agent can call ranking as a tool.Source: paper
  3. 54
    The lexicon (reconhecer) works as a router: it sends each request to the right specialist without training a classifier.Source: paper
  4. 55
    The deterministic top-k makes RAG evaluation reproducible: the same query returns the same list.Source: paper
  5. 56
    f32, f16, bf16, u8, i8, i16 and i32 tensors are format types, ready for model weights.Source: README
  6. 57
    Anything a model writes as valid JSON is already valid jsonW-text.Source: paper
  7. 58
    The schema fits in one line, short enough to go inside a prompt.Source: paper
  8. 59
    Audio, text and video signals measure and do not decide: the model gets numbers and decides.Source: paper
  9. 60
    The permission view applies to the agent too: it sees only what its credential allows.Source: paper
  10. 61
    MFCC comes out as a vf32[13] vector and can go straight to cosine, to find similar speech.Source: Newsroom
  11. 62
    Each video scene has a vf32[64] shape vector and a dHash, to find similar scenes with cosine.Source: Newsroom
  12. 63
    The mosaic of each scene's frames gives the model one image per scene instead of the whole video.Source: Newsroom
  13. 64
    The columnar file read through mmap serves for training, without one copy per process.Source: docs/JSONW.md
  14. 65
    AGENTS.md has the rules for a coding agent to use jsonW without mistakes.Source: AGENTS.md
  15. 66
    llms.txt summarizes the format in the standard for language models.Source: README
  16. 67
    The paper has a text version for AI in all six languages.Source: paper
  17. 68
    The text-signal lexicons were written by hand in Portuguese, English, Spanish and Chinese.Source: Newsroom
  18. 69
    Chinese is matched character by character, without depending on space segmentation.Source: Newsroom
  19. 70
    The ~~ operator lets an AI find the entity the user misspelled, with no extra model.Source: Newsroom
  20. 71
    reconhecer returns the original span of the text, with start and end, so the model can quote it.Source: paper
  21. 72
    vq8 and vq4 vectors let millions of embeddings fit in the memory of an ordinary machine.Source: docs/JSONW.md
  22. 73
    The jsonW website exposes tools via WebMCP for agents in the browser.Source: Newsroom
  23. 74
    The same cosine query runs over vf32, vf16, vq8 or vq4 without changing the request.Source: paper
  24. 75
    IDF weights are stored in each row's vector, so term ranking does not depend on an updated external index.Source: paper
  25. 76
    The Python and Node SDKs accept an object or dictionary in gravar, straight from a model's output.Source: AGENTS.md
  26. 77
    Signals come out in the jsonw_sinal_esquema schema, so each measurement reaches the model with a type.Source: Newsroom
  27. 78
    A measurement that does not exist comes out as null instead of an invented number, such as the pitch of a segment with no voice.Source: Newsroom
  28. 79
    reconhecer adds up the weights per signal, and each intent's total comes out ready for a router to decide.Source: paper

Security and privacy

What the format guarantees and what it does not promise.

  1. 80
    The scope comes from the server after authentication, never from a header sent by the client.Source: paper
  2. 81
    Asking for a hidden column through campos= or onde= returns 403, so nothing leaks by inference.Source: paper
  3. 82
    Without the scope, the column comes out null and cannot be used even in a filter.Source: paper
  4. 83
    A text tenant column closes the view: nothing comes out, instead of everything.Source: paper
  5. 84
    The credential is checked in constant time, which closes timing attacks.Source: paper
  6. 85
    A nonce and a sequence number prevent a recorded request from being replayed.Source: paper
  7. 86
    The reader treats every file as hostile: it checks every index and limit before using it.Source: paper
  8. 87
    The reader is fuzz-tested and rejects nesting deeper than 1,000 levels.Source: paper
  9. 88
    Quotes, commas, slashes and any byte inside a text are data and never break reading.Source: paper
  10. 89
    The library and the SDKs never access the network.Source: README
  11. 90
    The install count sends only version and platform, once, and is turned off with JSONW_SEM_PING=1.Source: README
  12. 91
    wcon reads the password from a file, never from the command line or the URL.Source: AGENTS.md
  13. 92
    Atomic publishing prevents a reader from getting a half-written file.Source: paper
  14. 93
    The +pessoal column hides personal data from anyone without the scope, which helps with LGPD and GDPR compliance.Source: paper
  15. 94
    Custody and credentials use only 256-bit hashes and HMAC, which remain secure against quantum computers.Source: paper
  16. 95
    Hidden data never leaves the server: the view is applied before answering.Source: paper
  17. 96
    Version 1.2.1 fixed an example configuration that served the raw file, and the paper says so.Source: paper
  18. 97
    SHA-256 can be switched to BLAKE3 when rewriting, and conferir reports which hash the file uses.Source: paper
  19. 98
    The view also works in wjsw ver with WJSW_ESCOPOS, to check what each credential sees before publishing.Source: paper
  20. 99
    wjsw-cgi never serves the raw .jsw: every request goes through the view.Source: AGENTS.md
  21. 100
    Using a Leitor after it is closed returns an error instead of reading freed memory.Source: AGENTS.md

Data types

The types JSON does not have, checked on write.

  1. 101
    decimal[2] stores 12.50 as 12.50, without going through floating point.Source: docs/JSONW.md
  2. 102
    dinheiro (money) stores the ISO 4217 currency in every value, so 10 BRL and 10 USD never become the same number.Source: docs/JSONW.md
  3. 103
    dinheiro accepts "R$ 1.234,56" and stores 1234.56 BRL.Source: docs/JSONW.md
  4. 104
    dinheiro accepts US$ and stores the currency USD.Source: docs/JSONW.md
  5. 105
    dinheiro accepts € and stores the currency EUR.Source: docs/JSONW.md
  6. 106
    dinheiro accepts £, ¥ and ₿ and stores GBP, JPY and BTC.Source: docs/JSONW.md
  7. 107
    medida (measure) stores the unit in every value or converts to the column's unit.Source: docs/JSONW.md
  8. 108
    codigo (code) stores the exact bytes: company IDs, keys, plates and SKUs are never normalized or trimmed.Source: docs/JSONW.md
  9. 109
    texto stores full Unicode, with accents, emoji and \0.Source: docs/JSONW.md
  10. 110
    texto+nfc normalizes Unicode when you ask.Source: docs/JSONW.md
  11. 111
    tempo stores microseconds UTC.Source: docs/JSONW.md
  12. 112
    uuid is a type, not just any text.Source: docs/JSONW.md
  13. 113
    url is validated on write, up to 8 KiB.Source: docs/JSONW.md
  14. 114
    dic stores each distinct text once per file.Source: docs/JSONW.md
  15. 115
    bits[n] stores binary vectors.Source: docs/JSONW.md
  16. 116
    ref points to a row of another file.Source: docs/JSONW.md
  17. 117
    lista<t>, obj{...} and mapa<k,v> nest up to 1,000 levels.Source: docs/JSONW.md
  18. 118
    qualquer accepts open JSON where the schema cannot be fixed.Source: docs/JSONW.md
  19. 119
    A ? after the type accepts null; without it, null is rejected.Source: docs/JSONW.md
  20. 120
    A new field comes in as null, and old files are read by new code and vice versa.Source: paper
  21. 121
    A name never changes type, so nobody reads a field with the wrong type.Source: paper
  22. 122
    comparar compares by value: -0 = 0, 3 = 3.0 and 1.50 = 1.500.Source: README
  23. 123
    Reals never depend on the machine's locale.Source: README
  24. 124
    An integer real comes out in JSON as 12.0, so a typed reader gets a real.Source: AGENTS.md

Units of measure

Every line below was tested with wjsw 1.4.1 on 2026-11-15: "1 unit" written into a medida column of that quantity.

  1. 125
    Measure in mm (length): a medida[m] column accepts "1 mm" and stores 0.001 m; another quantity is rejected.Source: test with wjsw 1.3.11
  2. 126
    Measure in cm (length): a medida[m] column accepts "1 cm" and stores 0.01 m; another quantity is rejected.Source: test with wjsw 1.3.11
  3. 127
    Measure in m (length): a medida[m] column accepts "1 m" and stores 1 m; another quantity is rejected.Source: test with wjsw 1.3.11
  4. 128
    Measure in km (length): a medida[m] column accepts "1 km" and stores 1000 m; another quantity is rejected.Source: test with wjsw 1.3.11
  5. 129
    Measure in µm (length): a medida[m] column accepts "1 µm" and stores 0.000001 m; another quantity is rejected.Source: test with wjsw 1.3.11
  6. 130
    Measure in nm (length): a medida[m] column accepts "1 nm" and stores 0.000000001 m; another quantity is rejected.Source: test with wjsw 1.3.11
  7. 131
    Measure in mi (length): a medida[m] column accepts "1 mi" and stores 1609 m; another quantity is rejected.Source: test with wjsw 1.3.11
  8. 132
    Measure in in (length): a medida[m] column accepts "1 in" and stores 0 m; another quantity is rejected.Source: test with wjsw 1.3.11
  9. 133
    Measure in ft (length): a medida[m] column accepts "1 ft" and stores 0 m; another quantity is rejected.Source: test with wjsw 1.3.11
  10. 134
    Measure in yd (length): a medida[m] column accepts "1 yd" and stores 1 m; another quantity is rejected.Source: test with wjsw 1.3.11
  11. 135
    Measure in nmi (length): a medida[m] column accepts "1 nmi" and stores 1852 m; another quantity is rejected.Source: test with wjsw 1.3.11
  12. 136
    Measure in mg (mass): a medida[g] column accepts "1 mg" and stores 0.001 g; another quantity is rejected.Source: test with wjsw 1.3.11
  13. 137
    Measure in g (mass): a medida[g] column accepts "1 g" and stores 1 g; another quantity is rejected.Source: test with wjsw 1.3.11
  14. 138
    Measure in kg (mass): a medida[g] column accepts "1 kg" and stores 1000 g; another quantity is rejected.Source: test with wjsw 1.3.11
  15. 139
    Measure in t (mass): a medida[g] column accepts "1 t" and stores 1000000 g; another quantity is rejected.Source: test with wjsw 1.3.11
  16. 140
    Measure in lb (mass): a medida[g] column accepts "1 lb" and stores 454 g; another quantity is rejected.Source: test with wjsw 1.3.11
  17. 141
    Measure in oz (mass): a medida[g] column accepts "1 oz" and stores 28 g; another quantity is rejected.Source: test with wjsw 1.3.11
  18. 142
    Measure in ct (mass): a medida[g] column accepts "1 ct" and stores 0 g; another quantity is rejected.Source: test with wjsw 1.3.11
  19. 143
    Measure in mL (volume): a medida[L] column accepts "1 mL" and stores 0.001 L; another quantity is rejected.Source: test with wjsw 1.3.11
  20. 144
    Measure in L (volume): a medida[L] column accepts "1 L" and stores 1 L; another quantity is rejected.Source: test with wjsw 1.3.11
  21. 145
    Measure in m3 (volume): a medida[L] column accepts "1 m3" and stores 1000 L; another quantity is rejected.Source: test with wjsw 1.3.11
  22. 146
    Measure in cm3 (volume): a medida[L] column accepts "1 cm3" and stores 0.001 L; another quantity is rejected.Source: test with wjsw 1.3.11
  23. 147
    Measure in hL (volume): a medida[L] column accepts "1 hL" and stores 100 L; another quantity is rejected.Source: test with wjsw 1.3.11
  24. 148
    Measure in gal (volume): a medida[L] column accepts "1 gal" and stores 4 L; another quantity is rejected.Source: test with wjsw 1.3.11
  25. 149
    Measure in J (energy): a medida[J] column accepts "1 J" and stores 1 J; another quantity is rejected.Source: test with wjsw 1.3.11
  26. 150
    Measure in kJ (energy): a medida[J] column accepts "1 kJ" and stores 1000 J; another quantity is rejected.Source: test with wjsw 1.3.11
  27. 151
    Measure in MJ (energy): a medida[J] column accepts "1 MJ" and stores 1000000 J; another quantity is rejected.Source: test with wjsw 1.3.11
  28. 152
    Measure in GJ (energy): a medida[J] column accepts "1 GJ" and stores 1000000000 J; another quantity is rejected.Source: test with wjsw 1.3.11
  29. 153
    Measure in TJ (energy): a medida[J] column accepts "1 TJ" and stores 1000000000000 J; another quantity is rejected.Source: test with wjsw 1.3.11
  30. 154
    Measure in Wh (energy): a medida[J] column accepts "1 Wh" and stores 3600 J; another quantity is rejected.Source: test with wjsw 1.3.11
  31. 155
    Measure in kWh (energy): a medida[J] column accepts "1 kWh" and stores 3600000 J; another quantity is rejected.Source: test with wjsw 1.3.11
  32. 156
    Measure in MWh (energy): a medida[J] column accepts "1 MWh" and stores 3600000000 J; another quantity is rejected.Source: test with wjsw 1.3.11
  33. 157
    Measure in GWh (energy): a medida[J] column accepts "1 GWh" and stores 3600000000000 J; another quantity is rejected.Source: test with wjsw 1.3.11
  34. 158
    Measure in cal (energy): a medida[J] column accepts "1 cal" and stores 4 J; another quantity is rejected.Source: test with wjsw 1.3.11
  35. 159
    Measure in kcal (energy): a medida[J] column accepts "1 kcal" and stores 4184 J; another quantity is rejected.Source: test with wjsw 1.3.11
  36. 160
    Measure in BTU (energy): a medida[J] column accepts "1 BTU" and stores 1055 J; another quantity is rejected.Source: test with wjsw 1.3.11
  37. 161
    Measure in s (time): a medida[s] column accepts "1 s" and stores 1 s; another quantity is rejected.Source: test with wjsw 1.3.11
  38. 162
    Measure in ms (time): a medida[s] column accepts "1 ms" and stores 0.001 s; another quantity is rejected.Source: test with wjsw 1.3.11
  39. 163
    Measure in us (time): a medida[s] column accepts "1 us" and stores 0.000001 s; another quantity is rejected.Source: test with wjsw 1.3.11
  40. 164
    Measure in min (time): a medida[s] column accepts "1 min" and stores 60 s; another quantity is rejected.Source: test with wjsw 1.3.11
  41. 165
    Measure in h (time): a medida[s] column accepts "1 h" and stores 3600 s; another quantity is rejected.Source: test with wjsw 1.3.11
  42. 166
    Measure in d (time): a medida[s] column accepts "1 d" and stores 86400 s; another quantity is rejected.Source: test with wjsw 1.3.11
  43. 167
    Measure in sem (time): a medida[s] column accepts "1 sem" and stores 604800 s; another quantity is rejected.Source: test with wjsw 1.3.11
  44. 168
    Measure in m2 (area): a medida[m2] column accepts "1 m2" and stores 1 m2; another quantity is rejected.Source: test with wjsw 1.3.11
  45. 169
    Measure in cm2 (area): a medida[m2] column accepts "1 cm2" and stores 0.0001 m2; another quantity is rejected.Source: test with wjsw 1.3.11
  46. 170
    Measure in km2 (area): a medida[m2] column accepts "1 km2" and stores 1000000 m2; another quantity is rejected.Source: test with wjsw 1.3.11
  47. 171
    Measure in ha (area): a medida[m2] column accepts "1 ha" and stores 10000 m2; another quantity is rejected.Source: test with wjsw 1.3.11
  48. 172
    Measure in ft2 (area): a medida[m2] column accepts "1 ft2" and stores 0 m2; another quantity is rejected.Source: test with wjsw 1.3.11
  49. 173
    Measure in alq (area): a medida[m2] column accepts "1 alq" and stores 24200 m2; another quantity is rejected.Source: test with wjsw 1.3.11
  50. 174
    Measure in B (data): a medida[B] column accepts "1 B" and stores 1 B; another quantity is rejected.Source: test with wjsw 1.3.11
  51. 175
    Measure in KB (data): a medida[B] column accepts "1 KB" and stores 1000 B; another quantity is rejected.Source: test with wjsw 1.3.11
  52. 176
    Measure in MB (data): a medida[B] column accepts "1 MB" and stores 1000000 B; another quantity is rejected.Source: test with wjsw 1.3.11
  53. 177
    Measure in GB (data): a medida[B] column accepts "1 GB" and stores 1000000000 B; another quantity is rejected.Source: test with wjsw 1.3.11
  54. 178
    Measure in TB (data): a medida[B] column accepts "1 TB" and stores 1000000000000 B; another quantity is rejected.Source: test with wjsw 1.3.11
  55. 179
    Measure in KiB (data): a medida[B] column accepts "1 KiB" and stores 1024 B; another quantity is rejected.Source: test with wjsw 1.3.11
  56. 180
    Measure in MiB (data): a medida[B] column accepts "1 MiB" and stores 1048576 B; another quantity is rejected.Source: test with wjsw 1.3.11
  57. 181
    Measure in GiB (data): a medida[B] column accepts "1 GiB" and stores 1073741824 B; another quantity is rejected.Source: test with wjsw 1.3.11
  58. 182
    Measure in TiB (data): a medida[B] column accepts "1 TiB" and stores 1099511627776 B; another quantity is rejected.Source: test with wjsw 1.3.11
  59. 183
    Measure in bit (data): a medida[B] column accepts "1 bit" and stores 0 B; another quantity is rejected.Source: test with wjsw 1.3.11
  60. 184
    Measure in m/s (speed): a medida[m/s] column accepts "1 m/s" and stores 1 m/s; another quantity is rejected.Source: test with wjsw 1.3.11
  61. 185
    Measure in km/h (speed): a medida[m/s] column accepts "1 km/h" and stores 0 m/s; another quantity is rejected.Source: test with wjsw 1.3.11
  62. 186
    Measure in mph (speed): a medida[m/s] column accepts "1 mph" and stores 0 m/s; another quantity is rejected.Source: test with wjsw 1.3.11
  63. 187
    Measure in kn (speed): a medida[m/s] column accepts "1 kn" and stores 1 m/s; another quantity is rejected.Source: test with wjsw 1.3.11
  64. 188
    Measure in ft/s (speed): a medida[m/s] column accepts "1 ft/s" and stores 0 m/s; another quantity is rejected.Source: test with wjsw 1.3.11
  65. 189
    Measure in C (temperature): a medida[C] column accepts "1 C" and stores 1 C; another quantity is rejected.Source: test with wjsw 1.3.11
  66. 190
    Measure in F (temperature): a medida[C] column accepts "1 F" and stores -17 C; another quantity is rejected.Source: test with wjsw 1.3.11
  67. 191
    Measure in K (temperature): a medida[C] column accepts "1 K" and stores -272 C; another quantity is rejected.Source: test with wjsw 1.3.11
  68. 192
    Measure in un (dimensionless): a medida[un] column accepts "1 un" and stores 1 un; another quantity is rejected.Source: test with wjsw 1.3.11
  69. 193
    Measure in % (dimensionless): a medida[un] column accepts "1 %" and stores 0.01 un; another quantity is rejected.Source: test with wjsw 1.3.11
  70. 194
    Measure in permil (dimensionless): a medida[un] column accepts "1 permil" and stores 0.001 un; another quantity is rejected.Source: test with wjsw 1.3.11
  71. 195
    Measure in ppm (dimensionless): a medida[un] column accepts "1 ppm" and stores 0.000001 un; another quantity is rejected.Source: test with wjsw 1.3.11
  72. 196
    Measure in pc (dimensionless): a medida[un] column accepts "1 pc" and stores 1 un; another quantity is rejected.Source: test with wjsw 1.3.11
  73. 197
    Measure in W (power): a medida[W] column accepts "1 W" and stores 1 W; another quantity is rejected.Source: test with wjsw 1.3.11
  74. 198
    Measure in kW (power): a medida[W] column accepts "1 kW" and stores 1000 W; another quantity is rejected.Source: test with wjsw 1.3.11
  75. 199
    Measure in MW (power): a medida[W] column accepts "1 MW" and stores 1000000 W; another quantity is rejected.Source: test with wjsw 1.3.11
  76. 200
    Measure in cv (power): a medida[W] column accepts "1 cv" and stores 735 W; another quantity is rejected.Source: test with wjsw 1.3.11
  77. 201
    Measure in hp (power): a medida[W] column accepts "1 hp" and stores 746 W; another quantity is rejected.Source: test with wjsw 1.3.11
  78. 202
    Measure in Pa (pressure): a medida[Pa] column accepts "1 Pa" and stores 1 Pa; another quantity is rejected.Source: test with wjsw 1.3.11
  79. 203
    Measure in kPa (pressure): a medida[Pa] column accepts "1 kPa" and stores 1000 Pa; another quantity is rejected.Source: test with wjsw 1.3.11
  80. 204
    Measure in bar (pressure): a medida[Pa] column accepts "1 bar" and stores 100000 Pa; another quantity is rejected.Source: test with wjsw 1.3.11
  81. 205
    Measure in atm (pressure): a medida[Pa] column accepts "1 atm" and stores 101325 Pa; another quantity is rejected.Source: test with wjsw 1.3.11
  82. 206
    Measure in psi (pressure): a medida[Pa] column accepts "1 psi" and stores 6895 Pa; another quantity is rejected.Source: test with wjsw 1.3.11
  83. 207
    Measure in mmHg (pressure): a medida[Pa] column accepts "1 mmHg" and stores 133 Pa; another quantity is rejected.Source: test with wjsw 1.3.11
  84. 208
    Measure in gCO2e (carbon): a medida[kgCO2e] column accepts "1 gCO2e" and stores 0.001 kgCO2e; another quantity is rejected.Source: test with wjsw 1.3.11
  85. 209
    Measure in kgCO2e (carbon): a medida[kgCO2e] column accepts "1 kgCO2e" and stores 1 kgCO2e; another quantity is rejected.Source: test with wjsw 1.3.11
  86. 210
    Measure in tCO2e (carbon): a medida[kgCO2e] column accepts "1 tCO2e" and stores 1000 kgCO2e; another quantity is rejected.Source: test with wjsw 1.3.11
  87. 211
    Measure in kg/m3 (density): a medida[kg/m3] column accepts "1 kg/m3" and stores 1 kg/m3; another quantity is rejected.Source: test with wjsw 1.3.11
  88. 212
    Measure in g/cm3 (density): a medida[kg/m3] column accepts "1 g/cm3" and stores 1000 kg/m3; another quantity is rejected.Source: test with wjsw 1.3.11
  89. 213
    Measure in g/L (density): a medida[kg/m3] column accepts "1 g/L" and stores 1 kg/m3; another quantity is rejected.Source: test with wjsw 1.3.11

Comparisons by format

Each comparison holds for the format itself, with no library on top, and cites the other format's official specification.

jsonW vs JSON

What jsonW has in the format itself and JSON does not, with no library on top.

  1. 214
    Money with currency: jsonW stores the ISO 4217 currency in every value (12.50 BRL), so reais and dollars never become the same number; JSON has no money type in the format itself.Source: JSON specification · paper
  2. 215
    Measure with unit: a medida[kg] column converts g, t or lb on write and rejects another quantity; JSON neither stores the unit nor converts.Source: JSON specification · paper
  3. 216
    Permissions by column, row and cell: whoever serves the file hides the +pessoal column and other tenants' rows before answering; JSON has no permission view in the format.Source: JSON specification · paper
  4. 217
    Tenant in the file: another customer's blocks are never even opened on read; JSON does not separate customers in the format itself.Source: JSON specification · paper
  5. 218
    A chain of SHA-256 or BLAKE3 hashes covering the whole file: one changed byte fails conferir, with no key at all; JSON has no such chain.Source: JSON specification · paper
  6. 219
    Semantic search inside the file, with cosine over vf32, vf16, vq8 or vq4; with JSON, you need a separate vector database.Source: JSON specification · paper
  7. 220
    Term search with IDF weights inside the file; with JSON, you need a separate search engine.Source: JSON specification · paper
  8. 221
    Top k with deterministic order (ties go to the lower row) in the format itself; JSON has no ranking.Source: JSON specification · paper
  9. 222
    A lexicon classifier (Aho-Corasick) that finds every term of a text in one pass; JSON has nothing like it.Source: JSON specification · paper
  10. 223
    Multilinear grid: bilinear, trilinear and higher interpolation straight from the file; with JSON, interpolation lives in application code.Source: JSON specification · paper
  11. 224
    Similar spelling and sound: "Kokakola" finds "Coca-Cola" with the ~~ operator, in pt, en, es and zh; JSON has no similar-spelling search.Source: JSON specification · paper
  12. 225
    Audio signals (pitch, pauses, jitter, shimmer, HNR, formants, MFCC) measured and written as one row of the format; JSON does not measure audio.Source: JSON specification · paper
  13. 226
    Text signals (negation, polarity, passive voice, pronouns, emoji, laughter) in pt, en, es and zh; JSON does not measure text.Source: JSON specification · paper
  14. 227
    Video in scenes: cuts, light, motion, colors and each segment's audio come out as rows of the format; JSON does not analyze video.Source: JSON specification · paper
  15. 228
    History in the file itself: writing the same key edits and _apagado: true deletes, without losing the previous version; JSON does not keep the edit history in the file.Source: JSON specification · paper
  16. 229
    URL validated on write (scheme://, mailto:, tel:, urn:); JSON accepts any text where an address should be.Source: JSON specification · paper
  17. 230
    A media type with hash, MIME type, size, duration, width and height; JSON has no media type.Source: JSON specification · paper
  18. 231
    Three lossless forms in the same model: binary row for real time, columnar file and text that accepts any JSON; JSON does not combine all three.Source: JSON specification · paper
  19. 232
    A credential signed with HMAC-SHA256, with tenant, scopes and expiry, checked in constant time and protected against replay; JSON defines no credential.Source: JSON specification · paper
  20. 233
    The schema inside the file, in one line (id:i64,nome:texto?,preco:dinheiro); JSON does not carry the schema with the data.Source: JSON specification · paper
  21. 234
    Reading one column without the others: in the paper's test, reading one column was 2,310 times faster than in JSONL; JSON forces you to read the whole record.Source: JSON specification · paper
  22. 235
    Exact decimal (decimal[2]): 12.50 never goes through floating point; JSON has no exact decimal.Source: JSON specification · paper
  23. 236
    Date and time with their own type (tempo, microseconds UTC); in JSON, a date has no type and every program reads it its own way.Source: JSON specification · paper
  24. 237
    Zones: minimum and maximum per block let a filter skip whole blocks (102 of 103 in a paper test); JSON forces you to read everything to filter.Source: JSON specification · paper
  25. 238
    Dictionary: repeated text (city, status, category) stored once per file; JSON repeats the text in every record.Source: JSON specification · paper
  26. 239
    Built-in per-block compression, at three levels, without losing access to each block; JSON has no compression in the format itself.Source: JSON specification · paper
  27. 240
    Partial remote reading via HTTP Range: reading one row of a 79.1 MB file downloaded 0.98% of it; JSON must be downloaded whole to find one row.Source: JSON specification · paper
  28. 241
    CRC32C on every block to catch accidental damage; JSON has no checksum.Source: JSON specification · paper
  29. 242
    A cut write loses only the last block, and wjsw reparar rebuilds the end; a JSON file cut in the middle becomes unreadable.Source: JSON specification · paper
  30. 243
    16, 8 and 4-bit vectors (vf16, vq8, vq4) as a type, ready for cosine; JSON has no half-precision or quantized vector.Source: JSON specification · paper
  31. 244
    Numbers with comma or dot: input accepts 1.234,56 and 1,234.56, and canonical output uses a dot; JSON accepts one syntax and leaves the separator to each program.Source: JSON specification · paper

jsonW vs JSONL

What jsonW has in the format itself and JSONL does not, with no library on top.

  1. 245
    Money with currency: jsonW stores the ISO 4217 currency in every value (12.50 BRL), so reais and dollars never become the same number; JSONL has no money type in the format itself.Source: JSONL specification · paper
  2. 246
    Measure with unit: a medida[kg] column converts g, t or lb on write and rejects another quantity; JSONL neither stores the unit nor converts.Source: JSONL specification · paper
  3. 247
    Permissions by column, row and cell: whoever serves the file hides the +pessoal column and other tenants' rows before answering; JSONL has no permission view in the format.Source: JSONL specification · paper
  4. 248
    Tenant in the file: another customer's blocks are never even opened on read; JSONL does not separate customers in the format itself.Source: JSONL specification · paper
  5. 249
    A chain of SHA-256 or BLAKE3 hashes covering the whole file: one changed byte fails conferir, with no key at all; JSONL has no such chain.Source: JSONL specification · paper
  6. 250
    Semantic search inside the file, with cosine over vf32, vf16, vq8 or vq4; with JSONL, you need a separate vector database.Source: JSONL specification · paper
  7. 251
    Term search with IDF weights inside the file; with JSONL, you need a separate search engine.Source: JSONL specification · paper
  8. 252
    Top k with deterministic order (ties go to the lower row) in the format itself; JSONL has no ranking.Source: JSONL specification · paper
  9. 253
    A lexicon classifier (Aho-Corasick) that finds every term of a text in one pass; JSONL has nothing like it.Source: JSONL specification · paper
  10. 254
    Multilinear grid: bilinear, trilinear and higher interpolation straight from the file; with JSONL, interpolation lives in application code.Source: JSONL specification · paper
  11. 255
    Similar spelling and sound: "Kokakola" finds "Coca-Cola" with the ~~ operator, in pt, en, es and zh; JSONL has no similar-spelling search.Source: JSONL specification · paper
  12. 256
    Audio signals (pitch, pauses, jitter, shimmer, HNR, formants, MFCC) measured and written as one row of the format; JSONL does not measure audio.Source: JSONL specification · paper
  13. 257
    Text signals (negation, polarity, passive voice, pronouns, emoji, laughter) in pt, en, es and zh; JSONL does not measure text.Source: JSONL specification · paper
  14. 258
    Video in scenes: cuts, light, motion, colors and each segment's audio come out as rows of the format; JSONL does not analyze video.Source: JSONL specification · paper
  15. 259
    History in the file itself: writing the same key edits and _apagado: true deletes, without losing the previous version; JSONL does not keep the edit history in the file.Source: JSONL specification · paper
  16. 260
    URL validated on write (scheme://, mailto:, tel:, urn:); JSONL accepts any text where an address should be.Source: JSONL specification · paper
  17. 261
    A media type with hash, MIME type, size, duration, width and height; JSONL has no media type.Source: JSONL specification · paper
  18. 262
    Three lossless forms in the same model: binary row for real time, columnar file and text that accepts any JSON; JSONL does not combine all three.Source: JSONL specification · paper
  19. 263
    A credential signed with HMAC-SHA256, with tenant, scopes and expiry, checked in constant time and protected against replay; JSONL defines no credential.Source: JSONL specification · paper
  20. 264
    The schema inside the file, in one line (id:i64,nome:texto?,preco:dinheiro); JSONL does not carry the schema with the data.Source: JSONL specification · paper
  21. 265
    Reading one column without the others: in the paper's test, reading one column was 2,310 times faster than in JSONL; JSONL forces you to read the whole record.Source: JSONL specification · paper
  22. 266
    Exact decimal (decimal[2]): 12.50 never goes through floating point; JSONL has no exact decimal.Source: JSONL specification · paper
  23. 267
    Date and time with their own type (tempo, microseconds UTC); in JSONL, a date has no type and every program reads it its own way.Source: JSONL specification · paper
  24. 268
    Zones: minimum and maximum per block let a filter skip whole blocks (102 of 103 in a paper test); JSONL forces you to read everything to filter.Source: JSONL specification · paper
  25. 269
    Dictionary: repeated text (city, status, category) stored once per file; JSONL repeats the text in every record.Source: JSONL specification · paper
  26. 270
    Built-in per-block compression, at three levels, without losing access to each block; JSONL has no compression in the format itself.Source: JSONL specification · paper
  27. 271
    Partial remote reading via HTTP Range: reading one row of a 79.1 MB file downloaded 0.98% of it; JSONL must be downloaded whole to find one row.Source: JSONL specification · paper
  28. 272
    CRC32C on every block to catch accidental damage; JSONL has no checksum.Source: JSONL specification · paper
  29. 273
    16, 8 and 4-bit vectors (vf16, vq8, vq4) as a type, ready for cosine; JSONL has no half-precision or quantized vector.Source: JSONL specification · paper
  30. 274
    Numbers with comma or dot: input accepts 1.234,56 and 1,234.56, and canonical output uses a dot; JSONL accepts one syntax and leaves the separator to each program.Source: JSONL specification · paper

jsonW vs CSV

What jsonW has in the format itself and CSV does not, with no library on top.

  1. 275
    Money with currency: jsonW stores the ISO 4217 currency in every value (12.50 BRL), so reais and dollars never become the same number; CSV has no money type in the format itself.Source: CSV specification · paper
  2. 276
    Measure with unit: a medida[kg] column converts g, t or lb on write and rejects another quantity; CSV neither stores the unit nor converts.Source: CSV specification · paper
  3. 277
    Permissions by column, row and cell: whoever serves the file hides the +pessoal column and other tenants' rows before answering; CSV has no permission view in the format.Source: CSV specification · paper
  4. 278
    Tenant in the file: another customer's blocks are never even opened on read; CSV does not separate customers in the format itself.Source: CSV specification · paper
  5. 279
    A chain of SHA-256 or BLAKE3 hashes covering the whole file: one changed byte fails conferir, with no key at all; CSV has no such chain.Source: CSV specification · paper
  6. 280
    Semantic search inside the file, with cosine over vf32, vf16, vq8 or vq4; with CSV, you need a separate vector database.Source: CSV specification · paper
  7. 281
    Term search with IDF weights inside the file; with CSV, you need a separate search engine.Source: CSV specification · paper
  8. 282
    Top k with deterministic order (ties go to the lower row) in the format itself; CSV has no ranking.Source: CSV specification · paper
  9. 283
    A lexicon classifier (Aho-Corasick) that finds every term of a text in one pass; CSV has nothing like it.Source: CSV specification · paper
  10. 284
    Multilinear grid: bilinear, trilinear and higher interpolation straight from the file; with CSV, interpolation lives in application code.Source: CSV specification · paper
  11. 285
    Similar spelling and sound: "Kokakola" finds "Coca-Cola" with the ~~ operator, in pt, en, es and zh; CSV has no similar-spelling search.Source: CSV specification · paper
  12. 286
    Audio signals (pitch, pauses, jitter, shimmer, HNR, formants, MFCC) measured and written as one row of the format; CSV does not measure audio.Source: CSV specification · paper
  13. 287
    Text signals (negation, polarity, passive voice, pronouns, emoji, laughter) in pt, en, es and zh; CSV does not measure text.Source: CSV specification · paper
  14. 288
    Video in scenes: cuts, light, motion, colors and each segment's audio come out as rows of the format; CSV does not analyze video.Source: CSV specification · paper
  15. 289
    History in the file itself: writing the same key edits and _apagado: true deletes, without losing the previous version; CSV does not keep the edit history in the file.Source: CSV specification · paper
  16. 290
    URL validated on write (scheme://, mailto:, tel:, urn:); CSV accepts any text where an address should be.Source: CSV specification · paper
  17. 291
    A media type with hash, MIME type, size, duration, width and height; CSV has no media type.Source: CSV specification · paper
  18. 292
    Three lossless forms in the same model: binary row for real time, columnar file and text that accepts any JSON; CSV does not combine all three.Source: CSV specification · paper
  19. 293
    A credential signed with HMAC-SHA256, with tenant, scopes and expiry, checked in constant time and protected against replay; CSV defines no credential.Source: CSV specification · paper
  20. 294
    The schema inside the file, in one line (id:i64,nome:texto?,preco:dinheiro); CSV does not carry the schema with the data.Source: CSV specification · paper
  21. 295
    Reading one column without the others: in the paper's test, reading one column was 2,310 times faster than in JSONL; CSV forces you to read the whole record.Source: CSV specification · paper
  22. 296
    Exact decimal (decimal[2]): 12.50 never goes through floating point; CSV has no exact decimal.Source: CSV specification · paper
  23. 297
    Date and time with their own type (tempo, microseconds UTC); in CSV, a date has no type and every program reads it its own way.Source: CSV specification · paper
  24. 298
    The type comes from the schema, not from guessing: "0123" in a codigo column stays "0123"; in CSV, the reader guesses the type and may drop the zero or turn "NO" into false.Source: CSV specification · paper
  25. 299
    Zones: minimum and maximum per block let a filter skip whole blocks (102 of 103 in a paper test); CSV forces you to read everything to filter.Source: CSV specification · paper
  26. 300
    Dictionary: repeated text (city, status, category) stored once per file; CSV repeats the text in every record.Source: CSV specification · paper
  27. 301
    Built-in per-block compression, at three levels, without losing access to each block; CSV has no compression in the format itself.Source: CSV specification · paper
  28. 302
    Partial remote reading via HTTP Range: reading one row of a 79.1 MB file downloaded 0.98% of it; CSV must be downloaded whole to find one row.Source: CSV specification · paper
  29. 303
    CRC32C on every block to catch accidental damage; CSV has no checksum.Source: CSV specification · paper
  30. 304
    16, 8 and 4-bit vectors (vf16, vq8, vq4) as a type, ready for cosine; CSV has no half-precision or quantized vector.Source: CSV specification · paper
  31. 305
    Typed nested lists, objects and maps (lista<t>, obj{...}, mapa<k,v>); CSV has only flat rows and columns.Source: CSV specification · paper
  32. 306
    Any valid JSON is already valid jsonW-text, with no conversion; going from JSON to CSV requires a conversion.Source: CSV specification · paper
  33. 307
    Numbers with comma or dot: input accepts 1.234,56 and 1,234.56, and canonical output uses a dot; CSV accepts one syntax and leaves the separator to each program.Source: CSV specification · paper

jsonW vs TSV

What jsonW has in the format itself and TSV does not, with no library on top.

  1. 308
    Money with currency: jsonW stores the ISO 4217 currency in every value (12.50 BRL), so reais and dollars never become the same number; TSV has no money type in the format itself.Source: TSV specification · paper
  2. 309
    Measure with unit: a medida[kg] column converts g, t or lb on write and rejects another quantity; TSV neither stores the unit nor converts.Source: TSV specification · paper
  3. 310
    Permissions by column, row and cell: whoever serves the file hides the +pessoal column and other tenants' rows before answering; TSV has no permission view in the format.Source: TSV specification · paper
  4. 311
    Tenant in the file: another customer's blocks are never even opened on read; TSV does not separate customers in the format itself.Source: TSV specification · paper
  5. 312
    A chain of SHA-256 or BLAKE3 hashes covering the whole file: one changed byte fails conferir, with no key at all; TSV has no such chain.Source: TSV specification · paper
  6. 313
    Semantic search inside the file, with cosine over vf32, vf16, vq8 or vq4; with TSV, you need a separate vector database.Source: TSV specification · paper
  7. 314
    Term search with IDF weights inside the file; with TSV, you need a separate search engine.Source: TSV specification · paper
  8. 315
    Top k with deterministic order (ties go to the lower row) in the format itself; TSV has no ranking.Source: TSV specification · paper
  9. 316
    A lexicon classifier (Aho-Corasick) that finds every term of a text in one pass; TSV has nothing like it.Source: TSV specification · paper
  10. 317
    Multilinear grid: bilinear, trilinear and higher interpolation straight from the file; with TSV, interpolation lives in application code.Source: TSV specification · paper
  11. 318
    Similar spelling and sound: "Kokakola" finds "Coca-Cola" with the ~~ operator, in pt, en, es and zh; TSV has no similar-spelling search.Source: TSV specification · paper
  12. 319
    Audio signals (pitch, pauses, jitter, shimmer, HNR, formants, MFCC) measured and written as one row of the format; TSV does not measure audio.Source: TSV specification · paper
  13. 320
    Text signals (negation, polarity, passive voice, pronouns, emoji, laughter) in pt, en, es and zh; TSV does not measure text.Source: TSV specification · paper
  14. 321
    Video in scenes: cuts, light, motion, colors and each segment's audio come out as rows of the format; TSV does not analyze video.Source: TSV specification · paper
  15. 322
    History in the file itself: writing the same key edits and _apagado: true deletes, without losing the previous version; TSV does not keep the edit history in the file.Source: TSV specification · paper
  16. 323
    URL validated on write (scheme://, mailto:, tel:, urn:); TSV accepts any text where an address should be.Source: TSV specification · paper
  17. 324
    A media type with hash, MIME type, size, duration, width and height; TSV has no media type.Source: TSV specification · paper
  18. 325
    Three lossless forms in the same model: binary row for real time, columnar file and text that accepts any JSON; TSV does not combine all three.Source: TSV specification · paper
  19. 326
    A credential signed with HMAC-SHA256, with tenant, scopes and expiry, checked in constant time and protected against replay; TSV defines no credential.Source: TSV specification · paper
  20. 327
    The schema inside the file, in one line (id:i64,nome:texto?,preco:dinheiro); TSV does not carry the schema with the data.Source: TSV specification · paper
  21. 328
    Reading one column without the others: in the paper's test, reading one column was 2,310 times faster than in JSONL; TSV forces you to read the whole record.Source: TSV specification · paper
  22. 329
    Exact decimal (decimal[2]): 12.50 never goes through floating point; TSV has no exact decimal.Source: TSV specification · paper
  23. 330
    Date and time with their own type (tempo, microseconds UTC); in TSV, a date has no type and every program reads it its own way.Source: TSV specification · paper
  24. 331
    The type comes from the schema, not from guessing: "0123" in a codigo column stays "0123"; in TSV, the reader guesses the type and may drop the zero or turn "NO" into false.Source: TSV specification · paper
  25. 332
    Zones: minimum and maximum per block let a filter skip whole blocks (102 of 103 in a paper test); TSV forces you to read everything to filter.Source: TSV specification · paper
  26. 333
    Dictionary: repeated text (city, status, category) stored once per file; TSV repeats the text in every record.Source: TSV specification · paper
  27. 334
    Built-in per-block compression, at three levels, without losing access to each block; TSV has no compression in the format itself.Source: TSV specification · paper
  28. 335
    Partial remote reading via HTTP Range: reading one row of a 79.1 MB file downloaded 0.98% of it; TSV must be downloaded whole to find one row.Source: TSV specification · paper
  29. 336
    CRC32C on every block to catch accidental damage; TSV has no checksum.Source: TSV specification · paper
  30. 337
    16, 8 and 4-bit vectors (vf16, vq8, vq4) as a type, ready for cosine; TSV has no half-precision or quantized vector.Source: TSV specification · paper
  31. 338
    Typed nested lists, objects and maps (lista<t>, obj{...}, mapa<k,v>); TSV has only flat rows and columns.Source: TSV specification · paper
  32. 339
    Any valid JSON is already valid jsonW-text, with no conversion; going from JSON to TSV requires a conversion.Source: TSV specification · paper
  33. 340
    Numbers with comma or dot: input accepts 1.234,56 and 1,234.56, and canonical output uses a dot; TSV accepts one syntax and leaves the separator to each program.Source: TSV specification · paper

jsonW vs XML

What jsonW has in the format itself and XML does not, with no library on top.

  1. 341
    Money with currency: jsonW stores the ISO 4217 currency in every value (12.50 BRL), so reais and dollars never become the same number; XML has no money type in the format itself.Source: XML specification · paper
  2. 342
    Measure with unit: a medida[kg] column converts g, t or lb on write and rejects another quantity; XML neither stores the unit nor converts.Source: XML specification · paper
  3. 343
    Permissions by column, row and cell: whoever serves the file hides the +pessoal column and other tenants' rows before answering; XML has no permission view in the format.Source: XML specification · paper
  4. 344
    Tenant in the file: another customer's blocks are never even opened on read; XML does not separate customers in the format itself.Source: XML specification · paper
  5. 345
    A chain of SHA-256 or BLAKE3 hashes covering the whole file: one changed byte fails conferir, with no key at all; XML has no such chain.Source: XML specification · paper
  6. 346
    Semantic search inside the file, with cosine over vf32, vf16, vq8 or vq4; with XML, you need a separate vector database.Source: XML specification · paper
  7. 347
    Term search with IDF weights inside the file; with XML, you need a separate search engine.Source: XML specification · paper
  8. 348
    Top k with deterministic order (ties go to the lower row) in the format itself; XML has no ranking.Source: XML specification · paper
  9. 349
    A lexicon classifier (Aho-Corasick) that finds every term of a text in one pass; XML has nothing like it.Source: XML specification · paper
  10. 350
    Multilinear grid: bilinear, trilinear and higher interpolation straight from the file; with XML, interpolation lives in application code.Source: XML specification · paper
  11. 351
    Similar spelling and sound: "Kokakola" finds "Coca-Cola" with the ~~ operator, in pt, en, es and zh; XML has no similar-spelling search.Source: XML specification · paper
  12. 352
    Audio signals (pitch, pauses, jitter, shimmer, HNR, formants, MFCC) measured and written as one row of the format; XML does not measure audio.Source: XML specification · paper
  13. 353
    Text signals (negation, polarity, passive voice, pronouns, emoji, laughter) in pt, en, es and zh; XML does not measure text.Source: XML specification · paper
  14. 354
    Video in scenes: cuts, light, motion, colors and each segment's audio come out as rows of the format; XML does not analyze video.Source: XML specification · paper
  15. 355
    History in the file itself: writing the same key edits and _apagado: true deletes, without losing the previous version; XML does not keep the edit history in the file.Source: XML specification · paper
  16. 356
    URL validated on write (scheme://, mailto:, tel:, urn:); XML accepts any text where an address should be.Source: XML specification · paper
  17. 357
    A media type with hash, MIME type, size, duration, width and height; XML has no media type.Source: XML specification · paper
  18. 358
    Three lossless forms in the same model: binary row for real time, columnar file and text that accepts any JSON; XML does not combine all three.Source: XML specification · paper
  19. 359
    A credential signed with HMAC-SHA256, with tenant, scopes and expiry, checked in constant time and protected against replay; XML defines no credential.Source: XML specification · paper
  20. 360
    The schema inside the file, in one line (id:i64,nome:texto?,preco:dinheiro); XML does not carry the schema with the data.Source: XML specification · paper
  21. 361
    Reading one column without the others: in the paper's test, reading one column was 2,310 times faster than in JSONL; XML forces you to read the whole record.Source: XML specification · paper
  22. 362
    Exact decimal (decimal[2]): 12.50 never goes through floating point; XML has no exact decimal.Source: XML specification · paper
  23. 363
    Date and time with their own type (tempo, microseconds UTC); in XML, a date has no type and every program reads it its own way.Source: XML specification · paper
  24. 364
    Zones: minimum and maximum per block let a filter skip whole blocks (102 of 103 in a paper test); XML forces you to read everything to filter.Source: XML specification · paper
  25. 365
    Dictionary: repeated text (city, status, category) stored once per file; XML repeats the text in every record.Source: XML specification · paper
  26. 366
    Built-in per-block compression, at three levels, without losing access to each block; XML has no compression in the format itself.Source: XML specification · paper
  27. 367
    Partial remote reading via HTTP Range: reading one row of a 79.1 MB file downloaded 0.98% of it; XML must be downloaded whole to find one row.Source: XML specification · paper
  28. 368
    CRC32C on every block to catch accidental damage; XML has no checksum.Source: XML specification · paper
  29. 369
    A cut write loses only the last block, and wjsw reparar rebuilds the end; a XML file cut in the middle becomes unreadable.Source: XML specification · paper
  30. 370
    16, 8 and 4-bit vectors (vf16, vq8, vq4) as a type, ready for cosine; XML has no half-precision or quantized vector.Source: XML specification · paper
  31. 371
    Any valid JSON is already valid jsonW-text, with no conversion; going from JSON to XML requires a conversion.Source: XML specification · paper
  32. 372
    Numbers with comma or dot: input accepts 1.234,56 and 1,234.56, and canonical output uses a dot; XML accepts one syntax and leaves the separator to each program.Source: XML specification · paper

jsonW vs XLSX

What jsonW has in the format itself and XLSX does not, with no library on top.

  1. 373
    Money with currency: jsonW stores the ISO 4217 currency in every value (12.50 BRL), so reais and dollars never become the same number; XLSX has no money type in the format itself.Source: XLSX specification · paper
  2. 374
    Measure with unit: a medida[kg] column converts g, t or lb on write and rejects another quantity; XLSX neither stores the unit nor converts.Source: XLSX specification · paper
  3. 375
    Permissions by column, row and cell: whoever serves the file hides the +pessoal column and other tenants' rows before answering; XLSX has no permission view in the format.Source: XLSX specification · paper
  4. 376
    Tenant in the file: another customer's blocks are never even opened on read; XLSX does not separate customers in the format itself.Source: XLSX specification · paper
  5. 377
    A chain of SHA-256 or BLAKE3 hashes covering the whole file: one changed byte fails conferir, with no key at all; XLSX has no such chain.Source: XLSX specification · paper
  6. 378
    Semantic search inside the file, with cosine over vf32, vf16, vq8 or vq4; with XLSX, you need a separate vector database.Source: XLSX specification · paper
  7. 379
    Term search with IDF weights inside the file; with XLSX, you need a separate search engine.Source: XLSX specification · paper
  8. 380
    Top k with deterministic order (ties go to the lower row) in the format itself; XLSX has no ranking.Source: XLSX specification · paper
  9. 381
    A lexicon classifier (Aho-Corasick) that finds every term of a text in one pass; XLSX has nothing like it.Source: XLSX specification · paper
  10. 382
    Multilinear grid: bilinear, trilinear and higher interpolation straight from the file; with XLSX, interpolation lives in application code.Source: XLSX specification · paper
  11. 383
    Similar spelling and sound: "Kokakola" finds "Coca-Cola" with the ~~ operator, in pt, en, es and zh; XLSX has no similar-spelling search.Source: XLSX specification · paper
  12. 384
    Audio signals (pitch, pauses, jitter, shimmer, HNR, formants, MFCC) measured and written as one row of the format; XLSX does not measure audio.Source: XLSX specification · paper
  13. 385
    Text signals (negation, polarity, passive voice, pronouns, emoji, laughter) in pt, en, es and zh; XLSX does not measure text.Source: XLSX specification · paper
  14. 386
    Video in scenes: cuts, light, motion, colors and each segment's audio come out as rows of the format; XLSX does not analyze video.Source: XLSX specification · paper
  15. 387
    History in the file itself: writing the same key edits and _apagado: true deletes, without losing the previous version; XLSX does not keep the edit history in the file.Source: XLSX specification · paper
  16. 388
    URL validated on write (scheme://, mailto:, tel:, urn:); XLSX accepts any text where an address should be.Source: XLSX specification · paper
  17. 389
    A media type with hash, MIME type, size, duration, width and height; XLSX has no media type.Source: XLSX specification · paper
  18. 390
    Three lossless forms in the same model: binary row for real time, columnar file and text that accepts any JSON; XLSX does not combine all three.Source: XLSX specification · paper
  19. 391
    A credential signed with HMAC-SHA256, with tenant, scopes and expiry, checked in constant time and protected against replay; XLSX defines no credential.Source: XLSX specification · paper
  20. 392
    The schema inside the file, in one line (id:i64,nome:texto?,preco:dinheiro); XLSX does not carry the schema with the data.Source: XLSX specification · paper
  21. 393
    Reading one column without the others: in the paper's test, reading one column was 2,310 times faster than in JSONL; XLSX forces you to read the whole record.Source: XLSX specification · paper
  22. 394
    Exact decimal (decimal[2]): 12.50 never goes through floating point; XLSX has no exact decimal.Source: XLSX specification · paper
  23. 395
    Zones: minimum and maximum per block let a filter skip whole blocks (102 of 103 in a paper test); XLSX forces you to read everything to filter.Source: XLSX specification · paper
  24. 396
    Partial remote reading via HTTP Range: reading one row of a 79.1 MB file downloaded 0.98% of it; XLSX must be downloaded whole to find one row.Source: XLSX specification · paper
  25. 397
    A cut write loses only the last block, and wjsw reparar rebuilds the end; a XLSX file cut in the middle becomes unreadable.Source: XLSX specification · paper
  26. 398
    16, 8 and 4-bit vectors (vf16, vq8, vq4) as a type, ready for cosine; XLSX has no half-precision or quantized vector.Source: XLSX specification · paper
  27. 399
    No row limit per file: the INPI dataset has 24,978,976 rulings in one file; a XLSX sheet stops at 1,048,576 rows.Source: XLSX specification · paper
  28. 400
    Typed nested lists, objects and maps (lista<t>, obj{...}, mapa<k,v>); XLSX has only flat rows and columns.Source: XLSX specification · paper
  29. 401
    Any valid JSON is already valid jsonW-text, with no conversion; going from JSON to XLSX requires a conversion.Source: XLSX specification · paper

jsonW vs XLS

What jsonW has in the format itself and XLS does not, with no library on top.

  1. 402
    Money with currency: jsonW stores the ISO 4217 currency in every value (12.50 BRL), so reais and dollars never become the same number; XLS has no money type in the format itself.Source: XLS specification · paper
  2. 403
    Measure with unit: a medida[kg] column converts g, t or lb on write and rejects another quantity; XLS neither stores the unit nor converts.Source: XLS specification · paper
  3. 404
    Permissions by column, row and cell: whoever serves the file hides the +pessoal column and other tenants' rows before answering; XLS has no permission view in the format.Source: XLS specification · paper
  4. 405
    Tenant in the file: another customer's blocks are never even opened on read; XLS does not separate customers in the format itself.Source: XLS specification · paper
  5. 406
    A chain of SHA-256 or BLAKE3 hashes covering the whole file: one changed byte fails conferir, with no key at all; XLS has no such chain.Source: XLS specification · paper
  6. 407
    Semantic search inside the file, with cosine over vf32, vf16, vq8 or vq4; with XLS, you need a separate vector database.Source: XLS specification · paper
  7. 408
    Term search with IDF weights inside the file; with XLS, you need a separate search engine.Source: XLS specification · paper
  8. 409
    Top k with deterministic order (ties go to the lower row) in the format itself; XLS has no ranking.Source: XLS specification · paper
  9. 410
    A lexicon classifier (Aho-Corasick) that finds every term of a text in one pass; XLS has nothing like it.Source: XLS specification · paper
  10. 411
    Multilinear grid: bilinear, trilinear and higher interpolation straight from the file; with XLS, interpolation lives in application code.Source: XLS specification · paper
  11. 412
    Similar spelling and sound: "Kokakola" finds "Coca-Cola" with the ~~ operator, in pt, en, es and zh; XLS has no similar-spelling search.Source: XLS specification · paper
  12. 413
    Audio signals (pitch, pauses, jitter, shimmer, HNR, formants, MFCC) measured and written as one row of the format; XLS does not measure audio.Source: XLS specification · paper
  13. 414
    Text signals (negation, polarity, passive voice, pronouns, emoji, laughter) in pt, en, es and zh; XLS does not measure text.Source: XLS specification · paper
  14. 415
    Video in scenes: cuts, light, motion, colors and each segment's audio come out as rows of the format; XLS does not analyze video.Source: XLS specification · paper
  15. 416
    History in the file itself: writing the same key edits and _apagado: true deletes, without losing the previous version; XLS does not keep the edit history in the file.Source: XLS specification · paper
  16. 417
    URL validated on write (scheme://, mailto:, tel:, urn:); XLS accepts any text where an address should be.Source: XLS specification · paper
  17. 418
    A media type with hash, MIME type, size, duration, width and height; XLS has no media type.Source: XLS specification · paper
  18. 419
    Three lossless forms in the same model: binary row for real time, columnar file and text that accepts any JSON; XLS does not combine all three.Source: XLS specification · paper
  19. 420
    A credential signed with HMAC-SHA256, with tenant, scopes and expiry, checked in constant time and protected against replay; XLS defines no credential.Source: XLS specification · paper
  20. 421
    The schema inside the file, in one line (id:i64,nome:texto?,preco:dinheiro); XLS does not carry the schema with the data.Source: XLS specification · paper
  21. 422
    Reading one column without the others: in the paper's test, reading one column was 2,310 times faster than in JSONL; XLS forces you to read the whole record.Source: XLS specification · paper
  22. 423
    Exact decimal (decimal[2]): 12.50 never goes through floating point; XLS has no exact decimal.Source: XLS specification · paper
  23. 424
    Zones: minimum and maximum per block let a filter skip whole blocks (102 of 103 in a paper test); XLS forces you to read everything to filter.Source: XLS specification · paper
  24. 425
    Built-in per-block compression, at three levels, without losing access to each block; XLS has no compression in the format itself.Source: XLS specification · paper
  25. 426
    Partial remote reading via HTTP Range: reading one row of a 79.1 MB file downloaded 0.98% of it; XLS must be downloaded whole to find one row.Source: XLS specification · paper
  26. 427
    16, 8 and 4-bit vectors (vf16, vq8, vq4) as a type, ready for cosine; XLS has no half-precision or quantized vector.Source: XLS specification · paper
  27. 428
    No row limit per file: the INPI dataset has 24,978,976 rulings in one file; a XLS sheet stops at 65,536 rows.Source: XLS specification · paper
  28. 429
    Typed nested lists, objects and maps (lista<t>, obj{...}, mapa<k,v>); XLS has only flat rows and columns.Source: XLS specification · paper
  29. 430
    Any valid JSON is already valid jsonW-text, with no conversion; going from JSON to XLS requires a conversion.Source: XLS specification · paper

jsonW vs ODS

What jsonW has in the format itself and ODS does not, with no library on top.

  1. 431
    Measure with unit: a medida[kg] column converts g, t or lb on write and rejects another quantity; ODS neither stores the unit nor converts.Source: ODS specification · paper
  2. 432
    Permissions by column, row and cell: whoever serves the file hides the +pessoal column and other tenants' rows before answering; ODS has no permission view in the format.Source: ODS specification · paper
  3. 433
    Tenant in the file: another customer's blocks are never even opened on read; ODS does not separate customers in the format itself.Source: ODS specification · paper
  4. 434
    A chain of SHA-256 or BLAKE3 hashes covering the whole file: one changed byte fails conferir, with no key at all; ODS has no such chain.Source: ODS specification · paper
  5. 435
    Semantic search inside the file, with cosine over vf32, vf16, vq8 or vq4; with ODS, you need a separate vector database.Source: ODS specification · paper
  6. 436
    Term search with IDF weights inside the file; with ODS, you need a separate search engine.Source: ODS specification · paper
  7. 437
    Top k with deterministic order (ties go to the lower row) in the format itself; ODS has no ranking.Source: ODS specification · paper
  8. 438
    A lexicon classifier (Aho-Corasick) that finds every term of a text in one pass; ODS has nothing like it.Source: ODS specification · paper
  9. 439
    Multilinear grid: bilinear, trilinear and higher interpolation straight from the file; with ODS, interpolation lives in application code.Source: ODS specification · paper
  10. 440
    Similar spelling and sound: "Kokakola" finds "Coca-Cola" with the ~~ operator, in pt, en, es and zh; ODS has no similar-spelling search.Source: ODS specification · paper
  11. 441
    Audio signals (pitch, pauses, jitter, shimmer, HNR, formants, MFCC) measured and written as one row of the format; ODS does not measure audio.Source: ODS specification · paper
  12. 442
    Text signals (negation, polarity, passive voice, pronouns, emoji, laughter) in pt, en, es and zh; ODS does not measure text.Source: ODS specification · paper
  13. 443
    Video in scenes: cuts, light, motion, colors and each segment's audio come out as rows of the format; ODS does not analyze video.Source: ODS specification · paper
  14. 444
    History in the file itself: writing the same key edits and _apagado: true deletes, without losing the previous version; ODS does not keep the edit history in the file.Source: ODS specification · paper
  15. 445
    URL validated on write (scheme://, mailto:, tel:, urn:); ODS accepts any text where an address should be.Source: ODS specification · paper
  16. 446
    A media type with hash, MIME type, size, duration, width and height; ODS has no media type.Source: ODS specification · paper
  17. 447
    Three lossless forms in the same model: binary row for real time, columnar file and text that accepts any JSON; ODS does not combine all three.Source: ODS specification · paper
  18. 448
    A credential signed with HMAC-SHA256, with tenant, scopes and expiry, checked in constant time and protected against replay; ODS defines no credential.Source: ODS specification · paper
  19. 449
    The schema inside the file, in one line (id:i64,nome:texto?,preco:dinheiro); ODS does not carry the schema with the data.Source: ODS specification · paper
  20. 450
    Reading one column without the others: in the paper's test, reading one column was 2,310 times faster than in JSONL; ODS forces you to read the whole record.Source: ODS specification · paper
  21. 451
    Zones: minimum and maximum per block let a filter skip whole blocks (102 of 103 in a paper test); ODS forces you to read everything to filter.Source: ODS specification · paper
  22. 452
    Partial remote reading via HTTP Range: reading one row of a 79.1 MB file downloaded 0.98% of it; ODS must be downloaded whole to find one row.Source: ODS specification · paper
  23. 453
    A cut write loses only the last block, and wjsw reparar rebuilds the end; a ODS file cut in the middle becomes unreadable.Source: ODS specification · paper
  24. 454
    16, 8 and 4-bit vectors (vf16, vq8, vq4) as a type, ready for cosine; ODS has no half-precision or quantized vector.Source: ODS specification · paper
  25. 455
    Typed nested lists, objects and maps (lista<t>, obj{...}, mapa<k,v>); ODS has only flat rows and columns.Source: ODS specification · paper
  26. 456
    Any valid JSON is already valid jsonW-text, with no conversion; going from JSON to ODS requires a conversion.Source: ODS specification · paper

jsonW vs YAML

What jsonW has in the format itself and YAML does not, with no library on top.

  1. 457
    Money with currency: jsonW stores the ISO 4217 currency in every value (12.50 BRL), so reais and dollars never become the same number; YAML has no money type in the format itself.Source: YAML specification · paper
  2. 458
    Measure with unit: a medida[kg] column converts g, t or lb on write and rejects another quantity; YAML neither stores the unit nor converts.Source: YAML specification · paper
  3. 459
    Permissions by column, row and cell: whoever serves the file hides the +pessoal column and other tenants' rows before answering; YAML has no permission view in the format.Source: YAML specification · paper
  4. 460
    Tenant in the file: another customer's blocks are never even opened on read; YAML does not separate customers in the format itself.Source: YAML specification · paper
  5. 461
    A chain of SHA-256 or BLAKE3 hashes covering the whole file: one changed byte fails conferir, with no key at all; YAML has no such chain.Source: YAML specification · paper
  6. 462
    Semantic search inside the file, with cosine over vf32, vf16, vq8 or vq4; with YAML, you need a separate vector database.Source: YAML specification · paper
  7. 463
    Term search with IDF weights inside the file; with YAML, you need a separate search engine.Source: YAML specification · paper
  8. 464
    Top k with deterministic order (ties go to the lower row) in the format itself; YAML has no ranking.Source: YAML specification · paper
  9. 465
    A lexicon classifier (Aho-Corasick) that finds every term of a text in one pass; YAML has nothing like it.Source: YAML specification · paper
  10. 466
    Multilinear grid: bilinear, trilinear and higher interpolation straight from the file; with YAML, interpolation lives in application code.Source: YAML specification · paper
  11. 467
    Similar spelling and sound: "Kokakola" finds "Coca-Cola" with the ~~ operator, in pt, en, es and zh; YAML has no similar-spelling search.Source: YAML specification · paper
  12. 468
    Audio signals (pitch, pauses, jitter, shimmer, HNR, formants, MFCC) measured and written as one row of the format; YAML does not measure audio.Source: YAML specification · paper
  13. 469
    Text signals (negation, polarity, passive voice, pronouns, emoji, laughter) in pt, en, es and zh; YAML does not measure text.Source: YAML specification · paper
  14. 470
    Video in scenes: cuts, light, motion, colors and each segment's audio come out as rows of the format; YAML does not analyze video.Source: YAML specification · paper
  15. 471
    History in the file itself: writing the same key edits and _apagado: true deletes, without losing the previous version; YAML does not keep the edit history in the file.Source: YAML specification · paper
  16. 472
    URL validated on write (scheme://, mailto:, tel:, urn:); YAML accepts any text where an address should be.Source: YAML specification · paper
  17. 473
    A media type with hash, MIME type, size, duration, width and height; YAML has no media type.Source: YAML specification · paper
  18. 474
    Three lossless forms in the same model: binary row for real time, columnar file and text that accepts any JSON; YAML does not combine all three.Source: YAML specification · paper
  19. 475
    A credential signed with HMAC-SHA256, with tenant, scopes and expiry, checked in constant time and protected against replay; YAML defines no credential.Source: YAML specification · paper
  20. 476
    The schema inside the file, in one line (id:i64,nome:texto?,preco:dinheiro); YAML does not carry the schema with the data.Source: YAML specification · paper
  21. 477
    Reading one column without the others: in the paper's test, reading one column was 2,310 times faster than in JSONL; YAML forces you to read the whole record.Source: YAML specification · paper
  22. 478
    Exact decimal (decimal[2]): 12.50 never goes through floating point; YAML has no exact decimal.Source: YAML specification · paper
  23. 479
    The type comes from the schema, not from guessing: "0123" in a codigo column stays "0123"; in YAML, the reader guesses the type and may drop the zero or turn "NO" into false.Source: YAML specification · paper
  24. 480
    Zones: minimum and maximum per block let a filter skip whole blocks (102 of 103 in a paper test); YAML forces you to read everything to filter.Source: YAML specification · paper
  25. 481
    Built-in per-block compression, at three levels, without losing access to each block; YAML has no compression in the format itself.Source: YAML specification · paper
  26. 482
    Partial remote reading via HTTP Range: reading one row of a 79.1 MB file downloaded 0.98% of it; YAML must be downloaded whole to find one row.Source: YAML specification · paper
  27. 483
    CRC32C on every block to catch accidental damage; YAML has no checksum.Source: YAML specification · paper
  28. 484
    16, 8 and 4-bit vectors (vf16, vq8, vq4) as a type, ready for cosine; YAML has no half-precision or quantized vector.Source: YAML specification · paper
  29. 485
    Numbers with comma or dot: input accepts 1.234,56 and 1,234.56, and canonical output uses a dot; YAML accepts one syntax and leaves the separator to each program.Source: YAML specification · paper

jsonW vs TOML

What jsonW has in the format itself and TOML does not, with no library on top.

  1. 486
    Money with currency: jsonW stores the ISO 4217 currency in every value (12.50 BRL), so reais and dollars never become the same number; TOML has no money type in the format itself.Source: TOML specification · paper
  2. 487
    Measure with unit: a medida[kg] column converts g, t or lb on write and rejects another quantity; TOML neither stores the unit nor converts.Source: TOML specification · paper
  3. 488
    Permissions by column, row and cell: whoever serves the file hides the +pessoal column and other tenants' rows before answering; TOML has no permission view in the format.Source: TOML specification · paper
  4. 489
    Tenant in the file: another customer's blocks are never even opened on read; TOML does not separate customers in the format itself.Source: TOML specification · paper
  5. 490
    A chain of SHA-256 or BLAKE3 hashes covering the whole file: one changed byte fails conferir, with no key at all; TOML has no such chain.Source: TOML specification · paper
  6. 491
    Semantic search inside the file, with cosine over vf32, vf16, vq8 or vq4; with TOML, you need a separate vector database.Source: TOML specification · paper
  7. 492
    Term search with IDF weights inside the file; with TOML, you need a separate search engine.Source: TOML specification · paper
  8. 493
    Top k with deterministic order (ties go to the lower row) in the format itself; TOML has no ranking.Source: TOML specification · paper
  9. 494
    A lexicon classifier (Aho-Corasick) that finds every term of a text in one pass; TOML has nothing like it.Source: TOML specification · paper
  10. 495
    Multilinear grid: bilinear, trilinear and higher interpolation straight from the file; with TOML, interpolation lives in application code.Source: TOML specification · paper
  11. 496
    Similar spelling and sound: "Kokakola" finds "Coca-Cola" with the ~~ operator, in pt, en, es and zh; TOML has no similar-spelling search.Source: TOML specification · paper
  12. 497
    Audio signals (pitch, pauses, jitter, shimmer, HNR, formants, MFCC) measured and written as one row of the format; TOML does not measure audio.Source: TOML specification · paper
  13. 498
    Text signals (negation, polarity, passive voice, pronouns, emoji, laughter) in pt, en, es and zh; TOML does not measure text.Source: TOML specification · paper
  14. 499
    Video in scenes: cuts, light, motion, colors and each segment's audio come out as rows of the format; TOML does not analyze video.Source: TOML specification · paper
  15. 500
    History in the file itself: writing the same key edits and _apagado: true deletes, without losing the previous version; TOML does not keep the edit history in the file.Source: TOML specification · paper
  16. 501
    URL validated on write (scheme://, mailto:, tel:, urn:); TOML accepts any text where an address should be.Source: TOML specification · paper
  17. 502
    A media type with hash, MIME type, size, duration, width and height; TOML has no media type.Source: TOML specification · paper
  18. 503
    Three lossless forms in the same model: binary row for real time, columnar file and text that accepts any JSON; TOML does not combine all three.Source: TOML specification · paper
  19. 504
    A credential signed with HMAC-SHA256, with tenant, scopes and expiry, checked in constant time and protected against replay; TOML defines no credential.Source: TOML specification · paper
  20. 505
    The schema inside the file, in one line (id:i64,nome:texto?,preco:dinheiro); TOML does not carry the schema with the data.Source: TOML specification · paper
  21. 506
    Reading one column without the others: in the paper's test, reading one column was 2,310 times faster than in JSONL; TOML forces you to read the whole record.Source: TOML specification · paper
  22. 507
    Exact decimal (decimal[2]): 12.50 never goes through floating point; TOML has no exact decimal.Source: TOML specification · paper
  23. 508
    Zones: minimum and maximum per block let a filter skip whole blocks (102 of 103 in a paper test); TOML forces you to read everything to filter.Source: TOML specification · paper
  24. 509
    Dictionary: repeated text (city, status, category) stored once per file; TOML repeats the text in every record.Source: TOML specification · paper
  25. 510
    Built-in per-block compression, at three levels, without losing access to each block; TOML has no compression in the format itself.Source: TOML specification · paper
  26. 511
    Partial remote reading via HTTP Range: reading one row of a 79.1 MB file downloaded 0.98% of it; TOML must be downloaded whole to find one row.Source: TOML specification · paper
  27. 512
    CRC32C on every block to catch accidental damage; TOML has no checksum.Source: TOML specification · paper
  28. 513
    16, 8 and 4-bit vectors (vf16, vq8, vq4) as a type, ready for cosine; TOML has no half-precision or quantized vector.Source: TOML specification · paper
  29. 514
    Any valid JSON is already valid jsonW-text, with no conversion; going from JSON to TOML requires a conversion.Source: TOML specification · paper
  30. 515
    Numbers with comma or dot: input accepts 1.234,56 and 1,234.56, and canonical output uses a dot; TOML accepts one syntax and leaves the separator to each program.Source: TOML specification · paper

jsonW vs MessagePack

What jsonW has in the format itself and MessagePack does not, with no library on top.

  1. 516
    Money with currency: jsonW stores the ISO 4217 currency in every value (12.50 BRL), so reais and dollars never become the same number; MessagePack has no money type in the format itself.Source: MessagePack specification · paper
  2. 517
    Measure with unit: a medida[kg] column converts g, t or lb on write and rejects another quantity; MessagePack neither stores the unit nor converts.Source: MessagePack specification · paper
  3. 518
    Permissions by column, row and cell: whoever serves the file hides the +pessoal column and other tenants' rows before answering; MessagePack has no permission view in the format.Source: MessagePack specification · paper
  4. 519
    Tenant in the file: another customer's blocks are never even opened on read; MessagePack does not separate customers in the format itself.Source: MessagePack specification · paper
  5. 520
    A chain of SHA-256 or BLAKE3 hashes covering the whole file: one changed byte fails conferir, with no key at all; MessagePack has no such chain.Source: MessagePack specification · paper
  6. 521
    Semantic search inside the file, with cosine over vf32, vf16, vq8 or vq4; with MessagePack, you need a separate vector database.Source: MessagePack specification · paper
  7. 522
    Term search with IDF weights inside the file; with MessagePack, you need a separate search engine.Source: MessagePack specification · paper
  8. 523
    Top k with deterministic order (ties go to the lower row) in the format itself; MessagePack has no ranking.Source: MessagePack specification · paper
  9. 524
    A lexicon classifier (Aho-Corasick) that finds every term of a text in one pass; MessagePack has nothing like it.Source: MessagePack specification · paper
  10. 525
    Multilinear grid: bilinear, trilinear and higher interpolation straight from the file; with MessagePack, interpolation lives in application code.Source: MessagePack specification · paper
  11. 526
    Similar spelling and sound: "Kokakola" finds "Coca-Cola" with the ~~ operator, in pt, en, es and zh; MessagePack has no similar-spelling search.Source: MessagePack specification · paper
  12. 527
    Audio signals (pitch, pauses, jitter, shimmer, HNR, formants, MFCC) measured and written as one row of the format; MessagePack does not measure audio.Source: MessagePack specification · paper
  13. 528
    Text signals (negation, polarity, passive voice, pronouns, emoji, laughter) in pt, en, es and zh; MessagePack does not measure text.Source: MessagePack specification · paper
  14. 529
    Video in scenes: cuts, light, motion, colors and each segment's audio come out as rows of the format; MessagePack does not analyze video.Source: MessagePack specification · paper
  15. 530
    History in the file itself: writing the same key edits and _apagado: true deletes, without losing the previous version; MessagePack does not keep the edit history in the file.Source: MessagePack specification · paper
  16. 531
    URL validated on write (scheme://, mailto:, tel:, urn:); MessagePack accepts any text where an address should be.Source: MessagePack specification · paper
  17. 532
    A media type with hash, MIME type, size, duration, width and height; MessagePack has no media type.Source: MessagePack specification · paper
  18. 533
    Three lossless forms in the same model: binary row for real time, columnar file and text that accepts any JSON; MessagePack does not combine all three.Source: MessagePack specification · paper
  19. 534
    A credential signed with HMAC-SHA256, with tenant, scopes and expiry, checked in constant time and protected against replay; MessagePack defines no credential.Source: MessagePack specification · paper
  20. 535
    The schema inside the file, in one line (id:i64,nome:texto?,preco:dinheiro); MessagePack does not carry the schema with the data.Source: MessagePack specification · paper
  21. 536
    Reading one column without the others: in the paper's test, reading one column was 2,310 times faster than in JSONL; MessagePack forces you to read the whole record.Source: MessagePack specification · paper
  22. 537
    Exact decimal (decimal[2]): 12.50 never goes through floating point; MessagePack has no exact decimal.Source: MessagePack specification · paper
  23. 538
    Zones: minimum and maximum per block let a filter skip whole blocks (102 of 103 in a paper test); MessagePack forces you to read everything to filter.Source: MessagePack specification · paper
  24. 539
    Dictionary: repeated text (city, status, category) stored once per file; MessagePack repeats the text in every record.Source: MessagePack specification · paper
  25. 540
    Built-in per-block compression, at three levels, without losing access to each block; MessagePack has no compression in the format itself.Source: MessagePack specification · paper
  26. 541
    Partial remote reading via HTTP Range: reading one row of a 79.1 MB file downloaded 0.98% of it; MessagePack must be downloaded whole to find one row.Source: MessagePack specification · paper
  27. 542
    CRC32C on every block to catch accidental damage; MessagePack has no checksum.Source: MessagePack specification · paper
  28. 543
    16, 8 and 4-bit vectors (vf16, vq8, vq4) as a type, ready for cosine; MessagePack has no half-precision or quantized vector.Source: MessagePack specification · paper
  29. 544
    Any valid JSON is already valid jsonW-text, with no conversion; going from JSON to MessagePack requires a conversion.Source: MessagePack specification · paper

jsonW vs CBOR

What jsonW has in the format itself and CBOR does not, with no library on top.

  1. 545
    Money with currency: jsonW stores the ISO 4217 currency in every value (12.50 BRL), so reais and dollars never become the same number; CBOR has no money type in the format itself.Source: CBOR specification · paper
  2. 546
    Measure with unit: a medida[kg] column converts g, t or lb on write and rejects another quantity; CBOR neither stores the unit nor converts.Source: CBOR specification · paper
  3. 547
    Permissions by column, row and cell: whoever serves the file hides the +pessoal column and other tenants' rows before answering; CBOR has no permission view in the format.Source: CBOR specification · paper
  4. 548
    Tenant in the file: another customer's blocks are never even opened on read; CBOR does not separate customers in the format itself.Source: CBOR specification · paper
  5. 549
    A chain of SHA-256 or BLAKE3 hashes covering the whole file: one changed byte fails conferir, with no key at all; CBOR has no such chain.Source: CBOR specification · paper
  6. 550
    Semantic search inside the file, with cosine over vf32, vf16, vq8 or vq4; with CBOR, you need a separate vector database.Source: CBOR specification · paper
  7. 551
    Term search with IDF weights inside the file; with CBOR, you need a separate search engine.Source: CBOR specification · paper
  8. 552
    Top k with deterministic order (ties go to the lower row) in the format itself; CBOR has no ranking.Source: CBOR specification · paper
  9. 553
    A lexicon classifier (Aho-Corasick) that finds every term of a text in one pass; CBOR has nothing like it.Source: CBOR specification · paper
  10. 554
    Multilinear grid: bilinear, trilinear and higher interpolation straight from the file; with CBOR, interpolation lives in application code.Source: CBOR specification · paper
  11. 555
    Similar spelling and sound: "Kokakola" finds "Coca-Cola" with the ~~ operator, in pt, en, es and zh; CBOR has no similar-spelling search.Source: CBOR specification · paper
  12. 556
    Audio signals (pitch, pauses, jitter, shimmer, HNR, formants, MFCC) measured and written as one row of the format; CBOR does not measure audio.Source: CBOR specification · paper
  13. 557
    Text signals (negation, polarity, passive voice, pronouns, emoji, laughter) in pt, en, es and zh; CBOR does not measure text.Source: CBOR specification · paper
  14. 558
    Video in scenes: cuts, light, motion, colors and each segment's audio come out as rows of the format; CBOR does not analyze video.Source: CBOR specification · paper
  15. 559
    History in the file itself: writing the same key edits and _apagado: true deletes, without losing the previous version; CBOR does not keep the edit history in the file.Source: CBOR specification · paper
  16. 560
    URL validated on write (scheme://, mailto:, tel:, urn:); CBOR accepts any text where an address should be.Source: CBOR specification · paper
  17. 561
    A media type with hash, MIME type, size, duration, width and height; CBOR has no media type.Source: CBOR specification · paper
  18. 562
    Three lossless forms in the same model: binary row for real time, columnar file and text that accepts any JSON; CBOR does not combine all three.Source: CBOR specification · paper
  19. 563
    A credential signed with HMAC-SHA256, with tenant, scopes and expiry, checked in constant time and protected against replay; CBOR defines no credential.Source: CBOR specification · paper
  20. 564
    The schema inside the file, in one line (id:i64,nome:texto?,preco:dinheiro); CBOR does not carry the schema with the data.Source: CBOR specification · paper
  21. 565
    Reading one column without the others: in the paper's test, reading one column was 2,310 times faster than in JSONL; CBOR forces you to read the whole record.Source: CBOR specification · paper
  22. 566
    Zones: minimum and maximum per block let a filter skip whole blocks (102 of 103 in a paper test); CBOR forces you to read everything to filter.Source: CBOR specification · paper
  23. 567
    Built-in per-block compression, at three levels, without losing access to each block; CBOR has no compression in the format itself.Source: CBOR specification · paper
  24. 568
    Partial remote reading via HTTP Range: reading one row of a 79.1 MB file downloaded 0.98% of it; CBOR must be downloaded whole to find one row.Source: CBOR specification · paper
  25. 569
    CRC32C on every block to catch accidental damage; CBOR has no checksum.Source: CBOR specification · paper
  26. 570
    Any valid JSON is already valid jsonW-text, with no conversion; going from JSON to CBOR requires a conversion.Source: CBOR specification · paper

jsonW vs BSON

What jsonW has in the format itself and BSON does not, with no library on top.

  1. 571
    Money with currency: jsonW stores the ISO 4217 currency in every value (12.50 BRL), so reais and dollars never become the same number; BSON has no money type in the format itself.Source: BSON specification · paper
  2. 572
    Measure with unit: a medida[kg] column converts g, t or lb on write and rejects another quantity; BSON neither stores the unit nor converts.Source: BSON specification · paper
  3. 573
    Permissions by column, row and cell: whoever serves the file hides the +pessoal column and other tenants' rows before answering; BSON has no permission view in the format.Source: BSON specification · paper
  4. 574
    Tenant in the file: another customer's blocks are never even opened on read; BSON does not separate customers in the format itself.Source: BSON specification · paper
  5. 575
    A chain of SHA-256 or BLAKE3 hashes covering the whole file: one changed byte fails conferir, with no key at all; BSON has no such chain.Source: BSON specification · paper
  6. 576
    Semantic search inside the file, with cosine over vf32, vf16, vq8 or vq4; with BSON, you need a separate vector database.Source: BSON specification · paper
  7. 577
    Term search with IDF weights inside the file; with BSON, you need a separate search engine.Source: BSON specification · paper
  8. 578
    Top k with deterministic order (ties go to the lower row) in the format itself; BSON has no ranking.Source: BSON specification · paper
  9. 579
    A lexicon classifier (Aho-Corasick) that finds every term of a text in one pass; BSON has nothing like it.Source: BSON specification · paper
  10. 580
    Multilinear grid: bilinear, trilinear and higher interpolation straight from the file; with BSON, interpolation lives in application code.Source: BSON specification · paper
  11. 581
    Similar spelling and sound: "Kokakola" finds "Coca-Cola" with the ~~ operator, in pt, en, es and zh; BSON has no similar-spelling search.Source: BSON specification · paper
  12. 582
    Audio signals (pitch, pauses, jitter, shimmer, HNR, formants, MFCC) measured and written as one row of the format; BSON does not measure audio.Source: BSON specification · paper
  13. 583
    Text signals (negation, polarity, passive voice, pronouns, emoji, laughter) in pt, en, es and zh; BSON does not measure text.Source: BSON specification · paper
  14. 584
    Video in scenes: cuts, light, motion, colors and each segment's audio come out as rows of the format; BSON does not analyze video.Source: BSON specification · paper
  15. 585
    History in the file itself: writing the same key edits and _apagado: true deletes, without losing the previous version; BSON does not keep the edit history in the file.Source: BSON specification · paper
  16. 586
    URL validated on write (scheme://, mailto:, tel:, urn:); BSON accepts any text where an address should be.Source: BSON specification · paper
  17. 587
    A media type with hash, MIME type, size, duration, width and height; BSON has no media type.Source: BSON specification · paper
  18. 588
    Three lossless forms in the same model: binary row for real time, columnar file and text that accepts any JSON; BSON does not combine all three.Source: BSON specification · paper
  19. 589
    A credential signed with HMAC-SHA256, with tenant, scopes and expiry, checked in constant time and protected against replay; BSON defines no credential.Source: BSON specification · paper
  20. 590
    The schema inside the file, in one line (id:i64,nome:texto?,preco:dinheiro); BSON does not carry the schema with the data.Source: BSON specification · paper
  21. 591
    Reading one column without the others: in the paper's test, reading one column was 2,310 times faster than in JSONL; BSON forces you to read the whole record.Source: BSON specification · paper
  22. 592
    Zones: minimum and maximum per block let a filter skip whole blocks (102 of 103 in a paper test); BSON forces you to read everything to filter.Source: BSON specification · paper
  23. 593
    Dictionary: repeated text (city, status, category) stored once per file; BSON repeats the text in every record.Source: BSON specification · paper
  24. 594
    Built-in per-block compression, at three levels, without losing access to each block; BSON has no compression in the format itself.Source: BSON specification · paper
  25. 595
    Partial remote reading via HTTP Range: reading one row of a 79.1 MB file downloaded 0.98% of it; BSON must be downloaded whole to find one row.Source: BSON specification · paper
  26. 596
    CRC32C on every block to catch accidental damage; BSON has no checksum.Source: BSON specification · paper
  27. 597
    Any valid JSON is already valid jsonW-text, with no conversion; going from JSON to BSON requires a conversion.Source: BSON specification · paper

jsonW vs Protobuf

What jsonW has in the format itself and Protobuf does not, with no library on top.

  1. 598
    Measure with unit: a medida[kg] column converts g, t or lb on write and rejects another quantity; Protobuf neither stores the unit nor converts.Source: Protobuf specification · paper
  2. 599
    Permissions by column, row and cell: whoever serves the file hides the +pessoal column and other tenants' rows before answering; Protobuf has no permission view in the format.Source: Protobuf specification · paper
  3. 600
    Tenant in the file: another customer's blocks are never even opened on read; Protobuf does not separate customers in the format itself.Source: Protobuf specification · paper
  4. 601
    A chain of SHA-256 or BLAKE3 hashes covering the whole file: one changed byte fails conferir, with no key at all; Protobuf has no such chain.Source: Protobuf specification · paper
  5. 602
    Semantic search inside the file, with cosine over vf32, vf16, vq8 or vq4; with Protobuf, you need a separate vector database.Source: Protobuf specification · paper
  6. 603
    Term search with IDF weights inside the file; with Protobuf, you need a separate search engine.Source: Protobuf specification · paper
  7. 604
    Top k with deterministic order (ties go to the lower row) in the format itself; Protobuf has no ranking.Source: Protobuf specification · paper
  8. 605
    A lexicon classifier (Aho-Corasick) that finds every term of a text in one pass; Protobuf has nothing like it.Source: Protobuf specification · paper
  9. 606
    Multilinear grid: bilinear, trilinear and higher interpolation straight from the file; with Protobuf, interpolation lives in application code.Source: Protobuf specification · paper
  10. 607
    Similar spelling and sound: "Kokakola" finds "Coca-Cola" with the ~~ operator, in pt, en, es and zh; Protobuf has no similar-spelling search.Source: Protobuf specification · paper
  11. 608
    Audio signals (pitch, pauses, jitter, shimmer, HNR, formants, MFCC) measured and written as one row of the format; Protobuf does not measure audio.Source: Protobuf specification · paper
  12. 609
    Text signals (negation, polarity, passive voice, pronouns, emoji, laughter) in pt, en, es and zh; Protobuf does not measure text.Source: Protobuf specification · paper
  13. 610
    Video in scenes: cuts, light, motion, colors and each segment's audio come out as rows of the format; Protobuf does not analyze video.Source: Protobuf specification · paper
  14. 611
    History in the file itself: writing the same key edits and _apagado: true deletes, without losing the previous version; Protobuf does not keep the edit history in the file.Source: Protobuf specification · paper
  15. 612
    URL validated on write (scheme://, mailto:, tel:, urn:); Protobuf accepts any text where an address should be.Source: Protobuf specification · paper
  16. 613
    A media type with hash, MIME type, size, duration, width and height; Protobuf has no media type.Source: Protobuf specification · paper
  17. 614
    Three lossless forms in the same model: binary row for real time, columnar file and text that accepts any JSON; Protobuf does not combine all three.Source: Protobuf specification · paper
  18. 615
    A credential signed with HMAC-SHA256, with tenant, scopes and expiry, checked in constant time and protected against replay; Protobuf defines no credential.Source: Protobuf specification · paper
  19. 616
    The schema inside the file, in one line (id:i64,nome:texto?,preco:dinheiro); Protobuf does not carry the schema with the data.Source: Protobuf specification · paper
  20. 617
    Reading one column without the others: in the paper's test, reading one column was 2,310 times faster than in JSONL; Protobuf forces you to read the whole record.Source: Protobuf specification · paper
  21. 618
    Exact decimal (decimal[2]): 12.50 never goes through floating point; Protobuf has no exact decimal.Source: Protobuf specification · paper
  22. 619
    Zones: minimum and maximum per block let a filter skip whole blocks (102 of 103 in a paper test); Protobuf forces you to read everything to filter.Source: Protobuf specification · paper
  23. 620
    Dictionary: repeated text (city, status, category) stored once per file; Protobuf repeats the text in every record.Source: Protobuf specification · paper
  24. 621
    Built-in per-block compression, at three levels, without losing access to each block; Protobuf has no compression in the format itself.Source: Protobuf specification · paper
  25. 622
    Partial remote reading via HTTP Range: reading one row of a 79.1 MB file downloaded 0.98% of it; Protobuf must be downloaded whole to find one row.Source: Protobuf specification · paper
  26. 623
    CRC32C on every block to catch accidental damage; Protobuf has no checksum.Source: Protobuf specification · paper
  27. 624
    A cut write loses only the last block, and wjsw reparar rebuilds the end; a Protobuf file cut in the middle becomes unreadable.Source: Protobuf specification · paper
  28. 625
    16, 8 and 4-bit vectors (vf16, vq8, vq4) as a type, ready for cosine; Protobuf has no half-precision or quantized vector.Source: Protobuf specification · paper
  29. 626
    Any valid JSON is already valid jsonW-text, with no conversion; going from JSON to Protobuf requires a conversion.Source: Protobuf specification · paper

jsonW vs Thrift

What jsonW has in the format itself and Thrift does not, with no library on top.

  1. 627
    Money with currency: jsonW stores the ISO 4217 currency in every value (12.50 BRL), so reais and dollars never become the same number; Thrift has no money type in the format itself.Source: Thrift specification · paper
  2. 628
    Measure with unit: a medida[kg] column converts g, t or lb on write and rejects another quantity; Thrift neither stores the unit nor converts.Source: Thrift specification · paper
  3. 629
    Permissions by column, row and cell: whoever serves the file hides the +pessoal column and other tenants' rows before answering; Thrift has no permission view in the format.Source: Thrift specification · paper
  4. 630
    Tenant in the file: another customer's blocks are never even opened on read; Thrift does not separate customers in the format itself.Source: Thrift specification · paper
  5. 631
    A chain of SHA-256 or BLAKE3 hashes covering the whole file: one changed byte fails conferir, with no key at all; Thrift has no such chain.Source: Thrift specification · paper
  6. 632
    Semantic search inside the file, with cosine over vf32, vf16, vq8 or vq4; with Thrift, you need a separate vector database.Source: Thrift specification · paper
  7. 633
    Term search with IDF weights inside the file; with Thrift, you need a separate search engine.Source: Thrift specification · paper
  8. 634
    Top k with deterministic order (ties go to the lower row) in the format itself; Thrift has no ranking.Source: Thrift specification · paper
  9. 635
    A lexicon classifier (Aho-Corasick) that finds every term of a text in one pass; Thrift has nothing like it.Source: Thrift specification · paper
  10. 636
    Multilinear grid: bilinear, trilinear and higher interpolation straight from the file; with Thrift, interpolation lives in application code.Source: Thrift specification · paper
  11. 637
    Similar spelling and sound: "Kokakola" finds "Coca-Cola" with the ~~ operator, in pt, en, es and zh; Thrift has no similar-spelling search.Source: Thrift specification · paper
  12. 638
    Audio signals (pitch, pauses, jitter, shimmer, HNR, formants, MFCC) measured and written as one row of the format; Thrift does not measure audio.Source: Thrift specification · paper
  13. 639
    Text signals (negation, polarity, passive voice, pronouns, emoji, laughter) in pt, en, es and zh; Thrift does not measure text.Source: Thrift specification · paper
  14. 640
    Video in scenes: cuts, light, motion, colors and each segment's audio come out as rows of the format; Thrift does not analyze video.Source: Thrift specification · paper
  15. 641
    History in the file itself: writing the same key edits and _apagado: true deletes, without losing the previous version; Thrift does not keep the edit history in the file.Source: Thrift specification · paper
  16. 642
    URL validated on write (scheme://, mailto:, tel:, urn:); Thrift accepts any text where an address should be.Source: Thrift specification · paper
  17. 643
    A media type with hash, MIME type, size, duration, width and height; Thrift has no media type.Source: Thrift specification · paper
  18. 644
    Three lossless forms in the same model: binary row for real time, columnar file and text that accepts any JSON; Thrift does not combine all three.Source: Thrift specification · paper
  19. 645
    A credential signed with HMAC-SHA256, with tenant, scopes and expiry, checked in constant time and protected against replay; Thrift defines no credential.Source: Thrift specification · paper
  20. 646
    The schema inside the file, in one line (id:i64,nome:texto?,preco:dinheiro); Thrift does not carry the schema with the data.Source: Thrift specification · paper
  21. 647
    Reading one column without the others: in the paper's test, reading one column was 2,310 times faster than in JSONL; Thrift forces you to read the whole record.Source: Thrift specification · paper
  22. 648
    Exact decimal (decimal[2]): 12.50 never goes through floating point; Thrift has no exact decimal.Source: Thrift specification · paper
  23. 649
    Date and time with their own type (tempo, microseconds UTC); in Thrift, a date has no type and every program reads it its own way.Source: Thrift specification · paper
  24. 650
    Zones: minimum and maximum per block let a filter skip whole blocks (102 of 103 in a paper test); Thrift forces you to read everything to filter.Source: Thrift specification · paper
  25. 651
    Dictionary: repeated text (city, status, category) stored once per file; Thrift repeats the text in every record.Source: Thrift specification · paper
  26. 652
    Built-in per-block compression, at three levels, without losing access to each block; Thrift has no compression in the format itself.Source: Thrift specification · paper
  27. 653
    Partial remote reading via HTTP Range: reading one row of a 79.1 MB file downloaded 0.98% of it; Thrift must be downloaded whole to find one row.Source: Thrift specification · paper
  28. 654
    CRC32C on every block to catch accidental damage; Thrift has no checksum.Source: Thrift specification · paper
  29. 655
    A cut write loses only the last block, and wjsw reparar rebuilds the end; a Thrift file cut in the middle becomes unreadable.Source: Thrift specification · paper
  30. 656
    16, 8 and 4-bit vectors (vf16, vq8, vq4) as a type, ready for cosine; Thrift has no half-precision or quantized vector.Source: Thrift specification · paper
  31. 657
    Any valid JSON is already valid jsonW-text, with no conversion; going from JSON to Thrift requires a conversion.Source: Thrift specification · paper

jsonW vs FlatBuffers

What jsonW has in the format itself and FlatBuffers does not, with no library on top.

  1. 658
    Money with currency: jsonW stores the ISO 4217 currency in every value (12.50 BRL), so reais and dollars never become the same number; FlatBuffers has no money type in the format itself.Source: FlatBuffers specification · paper
  2. 659
    Measure with unit: a medida[kg] column converts g, t or lb on write and rejects another quantity; FlatBuffers neither stores the unit nor converts.Source: FlatBuffers specification · paper
  3. 660
    Permissions by column, row and cell: whoever serves the file hides the +pessoal column and other tenants' rows before answering; FlatBuffers has no permission view in the format.Source: FlatBuffers specification · paper
  4. 661
    Tenant in the file: another customer's blocks are never even opened on read; FlatBuffers does not separate customers in the format itself.Source: FlatBuffers specification · paper
  5. 662
    A chain of SHA-256 or BLAKE3 hashes covering the whole file: one changed byte fails conferir, with no key at all; FlatBuffers has no such chain.Source: FlatBuffers specification · paper
  6. 663
    Semantic search inside the file, with cosine over vf32, vf16, vq8 or vq4; with FlatBuffers, you need a separate vector database.Source: FlatBuffers specification · paper
  7. 664
    Term search with IDF weights inside the file; with FlatBuffers, you need a separate search engine.Source: FlatBuffers specification · paper
  8. 665
    Top k with deterministic order (ties go to the lower row) in the format itself; FlatBuffers has no ranking.Source: FlatBuffers specification · paper
  9. 666
    A lexicon classifier (Aho-Corasick) that finds every term of a text in one pass; FlatBuffers has nothing like it.Source: FlatBuffers specification · paper
  10. 667
    Multilinear grid: bilinear, trilinear and higher interpolation straight from the file; with FlatBuffers, interpolation lives in application code.Source: FlatBuffers specification · paper
  11. 668
    Similar spelling and sound: "Kokakola" finds "Coca-Cola" with the ~~ operator, in pt, en, es and zh; FlatBuffers has no similar-spelling search.Source: FlatBuffers specification · paper
  12. 669
    Audio signals (pitch, pauses, jitter, shimmer, HNR, formants, MFCC) measured and written as one row of the format; FlatBuffers does not measure audio.Source: FlatBuffers specification · paper
  13. 670
    Text signals (negation, polarity, passive voice, pronouns, emoji, laughter) in pt, en, es and zh; FlatBuffers does not measure text.Source: FlatBuffers specification · paper
  14. 671
    Video in scenes: cuts, light, motion, colors and each segment's audio come out as rows of the format; FlatBuffers does not analyze video.Source: FlatBuffers specification · paper
  15. 672
    History in the file itself: writing the same key edits and _apagado: true deletes, without losing the previous version; FlatBuffers does not keep the edit history in the file.Source: FlatBuffers specification · paper
  16. 673
    URL validated on write (scheme://, mailto:, tel:, urn:); FlatBuffers accepts any text where an address should be.Source: FlatBuffers specification · paper
  17. 674
    A media type with hash, MIME type, size, duration, width and height; FlatBuffers has no media type.Source: FlatBuffers specification · paper
  18. 675
    Three lossless forms in the same model: binary row for real time, columnar file and text that accepts any JSON; FlatBuffers does not combine all three.Source: FlatBuffers specification · paper
  19. 676
    A credential signed with HMAC-SHA256, with tenant, scopes and expiry, checked in constant time and protected against replay; FlatBuffers defines no credential.Source: FlatBuffers specification · paper
  20. 677
    The schema inside the file, in one line (id:i64,nome:texto?,preco:dinheiro); FlatBuffers does not carry the schema with the data.Source: FlatBuffers specification · paper
  21. 678
    Exact decimal (decimal[2]): 12.50 never goes through floating point; FlatBuffers has no exact decimal.Source: FlatBuffers specification · paper
  22. 679
    Date and time with their own type (tempo, microseconds UTC); in FlatBuffers, a date has no type and every program reads it its own way.Source: FlatBuffers specification · paper
  23. 680
    Zones: minimum and maximum per block let a filter skip whole blocks (102 of 103 in a paper test); FlatBuffers forces you to read everything to filter.Source: FlatBuffers specification · paper
  24. 681
    Built-in per-block compression, at three levels, without losing access to each block; FlatBuffers has no compression in the format itself.Source: FlatBuffers specification · paper
  25. 682
    CRC32C on every block to catch accidental damage; FlatBuffers has no checksum.Source: FlatBuffers specification · paper
  26. 683
    16, 8 and 4-bit vectors (vf16, vq8, vq4) as a type, ready for cosine; FlatBuffers has no half-precision or quantized vector.Source: FlatBuffers specification · paper
  27. 684
    Any valid JSON is already valid jsonW-text, with no conversion; going from JSON to FlatBuffers requires a conversion.Source: FlatBuffers specification · paper

jsonW vs Avro

What jsonW has in the format itself and Avro does not, with no library on top.

  1. 685
    Money with currency: jsonW stores the ISO 4217 currency in every value (12.50 BRL), so reais and dollars never become the same number; Avro has no money type in the format itself.Source: Avro specification · paper
  2. 686
    Measure with unit: a medida[kg] column converts g, t or lb on write and rejects another quantity; Avro neither stores the unit nor converts.Source: Avro specification · paper
  3. 687
    Permissions by column, row and cell: whoever serves the file hides the +pessoal column and other tenants' rows before answering; Avro has no permission view in the format.Source: Avro specification · paper
  4. 688
    Tenant in the file: another customer's blocks are never even opened on read; Avro does not separate customers in the format itself.Source: Avro specification · paper
  5. 689
    A chain of SHA-256 or BLAKE3 hashes covering the whole file: one changed byte fails conferir, with no key at all; Avro has no such chain.Source: Avro specification · paper
  6. 690
    Semantic search inside the file, with cosine over vf32, vf16, vq8 or vq4; with Avro, you need a separate vector database.Source: Avro specification · paper
  7. 691
    Term search with IDF weights inside the file; with Avro, you need a separate search engine.Source: Avro specification · paper
  8. 692
    Top k with deterministic order (ties go to the lower row) in the format itself; Avro has no ranking.Source: Avro specification · paper
  9. 693
    A lexicon classifier (Aho-Corasick) that finds every term of a text in one pass; Avro has nothing like it.Source: Avro specification · paper
  10. 694
    Multilinear grid: bilinear, trilinear and higher interpolation straight from the file; with Avro, interpolation lives in application code.Source: Avro specification · paper
  11. 695
    Similar spelling and sound: "Kokakola" finds "Coca-Cola" with the ~~ operator, in pt, en, es and zh; Avro has no similar-spelling search.Source: Avro specification · paper
  12. 696
    Audio signals (pitch, pauses, jitter, shimmer, HNR, formants, MFCC) measured and written as one row of the format; Avro does not measure audio.Source: Avro specification · paper
  13. 697
    Text signals (negation, polarity, passive voice, pronouns, emoji, laughter) in pt, en, es and zh; Avro does not measure text.Source: Avro specification · paper
  14. 698
    Video in scenes: cuts, light, motion, colors and each segment's audio come out as rows of the format; Avro does not analyze video.Source: Avro specification · paper
  15. 699
    History in the file itself: writing the same key edits and _apagado: true deletes, without losing the previous version; Avro does not keep the edit history in the file.Source: Avro specification · paper
  16. 700
    URL validated on write (scheme://, mailto:, tel:, urn:); Avro accepts any text where an address should be.Source: Avro specification · paper
  17. 701
    A media type with hash, MIME type, size, duration, width and height; Avro has no media type.Source: Avro specification · paper
  18. 702
    Three lossless forms in the same model: binary row for real time, columnar file and text that accepts any JSON; Avro does not combine all three.Source: Avro specification · paper
  19. 703
    A credential signed with HMAC-SHA256, with tenant, scopes and expiry, checked in constant time and protected against replay; Avro defines no credential.Source: Avro specification · paper
  20. 704
    Reading one column without the others: in the paper's test, reading one column was 2,310 times faster than in JSONL; Avro forces you to read the whole record.Source: Avro specification · paper
  21. 705
    Zones: minimum and maximum per block let a filter skip whole blocks (102 of 103 in a paper test); Avro forces you to read everything to filter.Source: Avro specification · paper
  22. 706
    Partial remote reading via HTTP Range: reading one row of a 79.1 MB file downloaded 0.98% of it; Avro must be downloaded whole to find one row.Source: Avro specification · paper
  23. 707
    16, 8 and 4-bit vectors (vf16, vq8, vq4) as a type, ready for cosine; Avro has no half-precision or quantized vector.Source: Avro specification · paper
  24. 708
    Any valid JSON is already valid jsonW-text, with no conversion; going from JSON to Avro requires a conversion.Source: Avro specification · paper

jsonW vs Parquet

What jsonW has in the format itself and Parquet does not, with no library on top.

  1. 709
    Money with currency: jsonW stores the ISO 4217 currency in every value (12.50 BRL), so reais and dollars never become the same number; Parquet has no money type in the format itself.Source: Parquet specification · paper
  2. 710
    Measure with unit: a medida[kg] column converts g, t or lb on write and rejects another quantity; Parquet neither stores the unit nor converts.Source: Parquet specification · paper
  3. 711
    Permissions by column, row and cell: whoever serves the file hides the +pessoal column and other tenants' rows before answering; Parquet has no permission view in the format.Source: Parquet specification · paper
  4. 712
    Tenant in the file: another customer's blocks are never even opened on read; Parquet does not separate customers in the format itself.Source: Parquet specification · paper
  5. 713
    A chain of SHA-256 or BLAKE3 hashes covering the whole file: one changed byte fails conferir, with no key at all; Parquet has no such chain.Source: Parquet specification · paper
  6. 714
    Semantic search inside the file, with cosine over vf32, vf16, vq8 or vq4; with Parquet, you need a separate vector database.Source: Parquet specification · paper
  7. 715
    Term search with IDF weights inside the file; with Parquet, you need a separate search engine.Source: Parquet specification · paper
  8. 716
    Top k with deterministic order (ties go to the lower row) in the format itself; Parquet has no ranking.Source: Parquet specification · paper
  9. 717
    A lexicon classifier (Aho-Corasick) that finds every term of a text in one pass; Parquet has nothing like it.Source: Parquet specification · paper
  10. 718
    Multilinear grid: bilinear, trilinear and higher interpolation straight from the file; with Parquet, interpolation lives in application code.Source: Parquet specification · paper
  11. 719
    Similar spelling and sound: "Kokakola" finds "Coca-Cola" with the ~~ operator, in pt, en, es and zh; Parquet has no similar-spelling search.Source: Parquet specification · paper
  12. 720
    Audio signals (pitch, pauses, jitter, shimmer, HNR, formants, MFCC) measured and written as one row of the format; Parquet does not measure audio.Source: Parquet specification · paper
  13. 721
    Text signals (negation, polarity, passive voice, pronouns, emoji, laughter) in pt, en, es and zh; Parquet does not measure text.Source: Parquet specification · paper
  14. 722
    Video in scenes: cuts, light, motion, colors and each segment's audio come out as rows of the format; Parquet does not analyze video.Source: Parquet specification · paper
  15. 723
    History in the file itself: writing the same key edits and _apagado: true deletes, without losing the previous version; Parquet does not keep the edit history in the file.Source: Parquet specification · paper
  16. 724
    URL validated on write (scheme://, mailto:, tel:, urn:); Parquet accepts any text where an address should be.Source: Parquet specification · paper
  17. 725
    A media type with hash, MIME type, size, duration, width and height; Parquet has no media type.Source: Parquet specification · paper
  18. 726
    Three lossless forms in the same model: binary row for real time, columnar file and text that accepts any JSON; Parquet does not combine all three.Source: Parquet specification · paper
  19. 727
    A credential signed with HMAC-SHA256, with tenant, scopes and expiry, checked in constant time and protected against replay; Parquet defines no credential.Source: Parquet specification · paper
  20. 728
    A cut write loses only the last block, and wjsw reparar rebuilds the end; a Parquet file cut in the middle becomes unreadable.Source: Parquet specification · paper
  21. 729
    Any valid JSON is already valid jsonW-text, with no conversion; going from JSON to Parquet requires a conversion.Source: Parquet specification · paper

jsonW vs Arrow

What jsonW has in the format itself and Arrow does not, with no library on top.

  1. 730
    Money with currency: jsonW stores the ISO 4217 currency in every value (12.50 BRL), so reais and dollars never become the same number; Arrow has no money type in the format itself.Source: Arrow specification · paper
  2. 731
    Measure with unit: a medida[kg] column converts g, t or lb on write and rejects another quantity; Arrow neither stores the unit nor converts.Source: Arrow specification · paper
  3. 732
    Permissions by column, row and cell: whoever serves the file hides the +pessoal column and other tenants' rows before answering; Arrow has no permission view in the format.Source: Arrow specification · paper
  4. 733
    Tenant in the file: another customer's blocks are never even opened on read; Arrow does not separate customers in the format itself.Source: Arrow specification · paper
  5. 734
    A chain of SHA-256 or BLAKE3 hashes covering the whole file: one changed byte fails conferir, with no key at all; Arrow has no such chain.Source: Arrow specification · paper
  6. 735
    Semantic search inside the file, with cosine over vf32, vf16, vq8 or vq4; with Arrow, you need a separate vector database.Source: Arrow specification · paper
  7. 736
    Term search with IDF weights inside the file; with Arrow, you need a separate search engine.Source: Arrow specification · paper
  8. 737
    Top k with deterministic order (ties go to the lower row) in the format itself; Arrow has no ranking.Source: Arrow specification · paper
  9. 738
    A lexicon classifier (Aho-Corasick) that finds every term of a text in one pass; Arrow has nothing like it.Source: Arrow specification · paper
  10. 739
    Multilinear grid: bilinear, trilinear and higher interpolation straight from the file; with Arrow, interpolation lives in application code.Source: Arrow specification · paper
  11. 740
    Similar spelling and sound: "Kokakola" finds "Coca-Cola" with the ~~ operator, in pt, en, es and zh; Arrow has no similar-spelling search.Source: Arrow specification · paper
  12. 741
    Audio signals (pitch, pauses, jitter, shimmer, HNR, formants, MFCC) measured and written as one row of the format; Arrow does not measure audio.Source: Arrow specification · paper
  13. 742
    Text signals (negation, polarity, passive voice, pronouns, emoji, laughter) in pt, en, es and zh; Arrow does not measure text.Source: Arrow specification · paper
  14. 743
    Video in scenes: cuts, light, motion, colors and each segment's audio come out as rows of the format; Arrow does not analyze video.Source: Arrow specification · paper
  15. 744
    History in the file itself: writing the same key edits and _apagado: true deletes, without losing the previous version; Arrow does not keep the edit history in the file.Source: Arrow specification · paper
  16. 745
    URL validated on write (scheme://, mailto:, tel:, urn:); Arrow accepts any text where an address should be.Source: Arrow specification · paper
  17. 746
    A media type with hash, MIME type, size, duration, width and height; Arrow has no media type.Source: Arrow specification · paper
  18. 747
    Three lossless forms in the same model: binary row for real time, columnar file and text that accepts any JSON; Arrow does not combine all three.Source: Arrow specification · paper
  19. 748
    A credential signed with HMAC-SHA256, with tenant, scopes and expiry, checked in constant time and protected against replay; Arrow defines no credential.Source: Arrow specification · paper
  20. 749
    CRC32C on every block to catch accidental damage; Arrow has no checksum.Source: Arrow specification · paper
  21. 750
    A cut write loses only the last block, and wjsw reparar rebuilds the end; a Arrow file cut in the middle becomes unreadable.Source: Arrow specification · paper
  22. 751
    Any valid JSON is already valid jsonW-text, with no conversion; going from JSON to Arrow requires a conversion.Source: Arrow specification · paper

jsonW vs ORC

What jsonW has in the format itself and ORC does not, with no library on top.

  1. 752
    Money with currency: jsonW stores the ISO 4217 currency in every value (12.50 BRL), so reais and dollars never become the same number; ORC has no money type in the format itself.Source: ORC specification · paper
  2. 753
    Measure with unit: a medida[kg] column converts g, t or lb on write and rejects another quantity; ORC neither stores the unit nor converts.Source: ORC specification · paper
  3. 754
    Permissions by column, row and cell: whoever serves the file hides the +pessoal column and other tenants' rows before answering; ORC has no permission view in the format.Source: ORC specification · paper
  4. 755
    Tenant in the file: another customer's blocks are never even opened on read; ORC does not separate customers in the format itself.Source: ORC specification · paper
  5. 756
    A chain of SHA-256 or BLAKE3 hashes covering the whole file: one changed byte fails conferir, with no key at all; ORC has no such chain.Source: ORC specification · paper
  6. 757
    Semantic search inside the file, with cosine over vf32, vf16, vq8 or vq4; with ORC, you need a separate vector database.Source: ORC specification · paper
  7. 758
    Term search with IDF weights inside the file; with ORC, you need a separate search engine.Source: ORC specification · paper
  8. 759
    Top k with deterministic order (ties go to the lower row) in the format itself; ORC has no ranking.Source: ORC specification · paper
  9. 760
    A lexicon classifier (Aho-Corasick) that finds every term of a text in one pass; ORC has nothing like it.Source: ORC specification · paper
  10. 761
    Multilinear grid: bilinear, trilinear and higher interpolation straight from the file; with ORC, interpolation lives in application code.Source: ORC specification · paper
  11. 762
    Similar spelling and sound: "Kokakola" finds "Coca-Cola" with the ~~ operator, in pt, en, es and zh; ORC has no similar-spelling search.Source: ORC specification · paper
  12. 763
    Audio signals (pitch, pauses, jitter, shimmer, HNR, formants, MFCC) measured and written as one row of the format; ORC does not measure audio.Source: ORC specification · paper
  13. 764
    Text signals (negation, polarity, passive voice, pronouns, emoji, laughter) in pt, en, es and zh; ORC does not measure text.Source: ORC specification · paper
  14. 765
    Video in scenes: cuts, light, motion, colors and each segment's audio come out as rows of the format; ORC does not analyze video.Source: ORC specification · paper
  15. 766
    History in the file itself: writing the same key edits and _apagado: true deletes, without losing the previous version; ORC does not keep the edit history in the file.Source: ORC specification · paper
  16. 767
    URL validated on write (scheme://, mailto:, tel:, urn:); ORC accepts any text where an address should be.Source: ORC specification · paper
  17. 768
    A media type with hash, MIME type, size, duration, width and height; ORC has no media type.Source: ORC specification · paper
  18. 769
    Three lossless forms in the same model: binary row for real time, columnar file and text that accepts any JSON; ORC does not combine all three.Source: ORC specification · paper
  19. 770
    A credential signed with HMAC-SHA256, with tenant, scopes and expiry, checked in constant time and protected against replay; ORC defines no credential.Source: ORC specification · paper
  20. 771
    A cut write loses only the last block, and wjsw reparar rebuilds the end; a ORC file cut in the middle becomes unreadable.Source: ORC specification · paper
  21. 772
    16, 8 and 4-bit vectors (vf16, vq8, vq4) as a type, ready for cosine; ORC has no half-precision or quantized vector.Source: ORC specification · paper
  22. 773
    Any valid JSON is already valid jsonW-text, with no conversion; going from JSON to ORC requires a conversion.Source: ORC specification · paper

jsonW vs SQLite

What jsonW has in the format itself and SQLite does not, with no library on top.

  1. 774
    Money with currency: jsonW stores the ISO 4217 currency in every value (12.50 BRL), so reais and dollars never become the same number; SQLite has no money type in the format itself.Source: SQLite specification · paper
  2. 775
    Measure with unit: a medida[kg] column converts g, t or lb on write and rejects another quantity; SQLite neither stores the unit nor converts.Source: SQLite specification · paper
  3. 776
    Permissions by column, row and cell: whoever serves the file hides the +pessoal column and other tenants' rows before answering; SQLite has no permission view in the format.Source: SQLite specification · paper
  4. 777
    Tenant in the file: another customer's blocks are never even opened on read; SQLite does not separate customers in the format itself.Source: SQLite specification · paper
  5. 778
    A chain of SHA-256 or BLAKE3 hashes covering the whole file: one changed byte fails conferir, with no key at all; SQLite has no such chain.Source: SQLite specification · paper
  6. 779
    Semantic search inside the file, with cosine over vf32, vf16, vq8 or vq4; with SQLite, you need a separate vector database.Source: SQLite specification · paper
  7. 780
    A lexicon classifier (Aho-Corasick) that finds every term of a text in one pass; SQLite has nothing like it.Source: SQLite specification · paper
  8. 781
    Multilinear grid: bilinear, trilinear and higher interpolation straight from the file; with SQLite, interpolation lives in application code.Source: SQLite specification · paper
  9. 782
    Audio signals (pitch, pauses, jitter, shimmer, HNR, formants, MFCC) measured and written as one row of the format; SQLite does not measure audio.Source: SQLite specification · paper
  10. 783
    Text signals (negation, polarity, passive voice, pronouns, emoji, laughter) in pt, en, es and zh; SQLite does not measure text.Source: SQLite specification · paper
  11. 784
    Video in scenes: cuts, light, motion, colors and each segment's audio come out as rows of the format; SQLite does not analyze video.Source: SQLite specification · paper
  12. 785
    History in the file itself: writing the same key edits and _apagado: true deletes, without losing the previous version; SQLite does not keep the edit history in the file.Source: SQLite specification · paper
  13. 786
    URL validated on write (scheme://, mailto:, tel:, urn:); SQLite accepts any text where an address should be.Source: SQLite specification · paper
  14. 787
    A media type with hash, MIME type, size, duration, width and height; SQLite has no media type.Source: SQLite specification · paper
  15. 788
    Three lossless forms in the same model: binary row for real time, columnar file and text that accepts any JSON; SQLite does not combine all three.Source: SQLite specification · paper
  16. 789
    A credential signed with HMAC-SHA256, with tenant, scopes and expiry, checked in constant time and protected against replay; SQLite defines no credential.Source: SQLite specification · paper
  17. 790
    Reading one column without the others: in the paper's test, reading one column was 2,310 times faster than in JSONL; SQLite forces you to read the whole record.Source: SQLite specification · paper
  18. 791
    Exact decimal (decimal[2]): 12.50 never goes through floating point; SQLite has no exact decimal.Source: SQLite specification · paper
  19. 792
    Built-in per-block compression, at three levels, without losing access to each block; SQLite has no compression in the format itself.Source: SQLite specification · paper
  20. 793
    CRC32C on every block to catch accidental damage; SQLite has no checksum.Source: SQLite specification · paper
  21. 794
    16, 8 and 4-bit vectors (vf16, vq8, vq4) as a type, ready for cosine; SQLite has no half-precision or quantized vector.Source: SQLite specification · paper
  22. 795
    Any valid JSON is already valid jsonW-text, with no conversion; going from JSON to SQLite requires a conversion.Source: SQLite specification · paper

Tool, databases, servers and cloud

Commands of the wjsw tool and of the programs that come in the package.

  1. 796
    wjsw de-texto converts JSON, JSONL or jsonW-text into .jsw with the schema you write.Source: docs/JSONW.md
  2. 797
    wjsw de-json converts JSON or JSONL into .jsw with an inferred schema.Source: AGENTS.md
  3. 798
    wjsw ver shows the rows as jsonW-text, already with the credential's view.Source: docs/JSONW.md
  4. 799
    wjsw ver file 100 50 shows 50 rows starting at row 100.Source: README
  5. 800
    wjsw info shows schema, blocks, size, hash and zones.Source: docs/JSONW.md
  6. 801
    wjsw conferir checks the whole custody chain.Source: docs/JSONW.md
  7. 802
    wjsw reparar rebuilds the end of a cut file.Source: docs/JSONW.md
  8. 803
    wjsw para-texto --json returns the JSON without loss.Source: docs/JSONW.md
  9. 804
    wjsw filtrar takes several conditions separated by semicolons.Source: docs/JSONW.md
  10. 805
    wjsw filtrar compares measures by quantity: "peso > 500" matches "0,5 kg".Source: AGENTS.md
  11. 806
    wjsw calcular gives count, exact sum, minimum, maximum and average, with --onde and --por.Source: docs/JSONW.md
  12. 807
    wjsw comparar shows what is only in a, only in b, what changed and what is equal.Source: docs/JSONW.md
  13. 808
    wjsw avaliar scores by weights, cosine, sparse or grid and returns the k best.Source: docs/JSONW.md
  14. 809
    wjsw escolher returns the k rows with the highest value in a column.Source: docs/JSONW.md
  15. 810
    wjsw reconhecer finds the lexicon's terms in a text and adds up per signal.Source: docs/JSONW.md
  16. 811
    wjsw medir measures writing and reading on your machine.Source: docs/JSONW.md
  17. 812
    wjsw regravar switches the hash to BLAKE3 or re-encodes with the new version's codecs.Source: paper
  18. 813
    wcon pg brings a PostgreSQL query result into .jsw, with the schema taken from the types.Source: AGENTS.md
  19. 814
    wcon pg-gravar writes the .jsw back to a PostgreSQL table, in one transaction.Source: AGENTS.md
  20. 815
    In Postgres, timestamptz, timestamp and date become tempo and come back unchanged.Source: AGENTS.md
  21. 816
    In Postgres, numeric becomes an exact decimal.Source: AGENTS.md
  22. 817
    wcon mysql brings a MySQL query into .jsw.Source: AGENTS.md
  23. 818
    wcon sqlite and wcon sqlite-gravar move data from SQLite to .jsw and back.Source: AGENTS.md
  24. 819
    wcon redis brings Redis keys by name pattern.Source: AGENTS.md
  25. 820
    wcon mongo brings a MongoDB collection with a filter.Source: AGENTS.md
  26. 821
    wremoto info opens a remote file downloading only the header and footer.Source: paper
  27. 822
    wremoto ver reads a remote row downloading only its block.Source: paper
  28. 823
    wremoto filtrar downloads only the blocks that can pass the filter.Source: paper
  29. 824
    On R2 and S3, every wremoto request is signed with AWS SigV4.Source: paper
  30. 825
    wjsw-cgi runs on Apache.Source: paper
  31. 826
    wjsw-cgi runs on nginx.Source: paper
  32. 827
    wjsw-cgi runs on LiteSpeed.Source: paper
  33. 828
    wjsw-cgi runs on Caddy.Source: paper
  34. 829
    wjsw-cgi runs on IIS.Source: paper
  35. 830
    wjsw-cgi answers in JSON or NDJSON with ?formato=, besides jsonW.Source: AGENTS.md
  36. 831
    A file with no marked column and no tenant is served as is, with Range 206 and CORS.Source: AGENTS.md
  37. 832
    Ready-made binaries for macOS arm64.Source: README
  38. 833
    Ready-made binaries for Linux x86_64 (glibc 2.34+: Ubuntu 22.04+, Debian 12+, RHEL 9+).Source: README
  39. 834
    Ready-made binaries for Windows x86_64.Source: README
  40. 835
    npm install @wortictecnologia/jsonw brings the prebuilt addon and the jsonw command.Source: AGENTS.md
  41. 836
    wjsw ver --json shows the same rows as strict JSON.Source: AGENTS.md
  42. 837
    wjsw avaliar --json returns one JSON per line, ready for another tool.Source: docs/JSONW.md
  43. 838
    WJSW_ESCOPOS sets the scopes of a command-line session, to test the view.Source: AGENTS.md

SDKs in 10 languages

Names are the same in every language, and all of them pass the same test, byte for byte.

  1. 839
    C SDK: the Leitor (reader) opens a .jsw file and returns every row, with the schema and the fields.Source: README
  2. 840
    C SDK: json(i) and texto(i) return row i as JSON or jsonW-text, without loss.Source: README
  3. 841
    C SDK: campo(i, name) reads one field of a row without decoding the rest.Source: README
  4. 842
    C SDK: conferir() checks the CRC32C and the hash of every block in the file.Source: README
  5. 843
    C SDK: avaliar(request, k) scores every row and returns the k best, with the same result as the tool.Source: README
  6. 844
    C SDK: the Escritor (writer) writes with compression 0, 1 or 2 and a SHA-256 or BLAKE3 hash.Source: README
  7. 845
    C SDK: fechar() publishes the file atomically: it only appears complete.Source: README
  8. 846
    C SDK: abandonar() gives up the write without touching the file that already existed.Source: README
  9. 847
    C SDK: the Reconhecedor builds the lexicon automaton and finds the terms in a text.Source: README
  10. 848
    C SDK: escolher(scores, k) returns the k highest scores in jsonW's deterministic order.Source: README
  11. 849
    C++ SDK: the Leitor (reader) opens a .jsw file and returns every row, with the schema and the fields.Source: README
  12. 850
    C++ SDK: json(i) and texto(i) return row i as JSON or jsonW-text, without loss.Source: README
  13. 851
    C++ SDK: campo(i, name) reads one field of a row without decoding the rest.Source: README
  14. 852
    C++ SDK: conferir() checks the CRC32C and the hash of every block in the file.Source: README
  15. 853
    C++ SDK: avaliar(request, k) scores every row and returns the k best, with the same result as the tool.Source: README
  16. 854
    C++ SDK: the Escritor (writer) writes with compression 0, 1 or 2 and a SHA-256 or BLAKE3 hash.Source: README
  17. 855
    C++ SDK: fechar() publishes the file atomically: it only appears complete.Source: README
  18. 856
    C++ SDK: abandonar() gives up the write without touching the file that already existed.Source: README
  19. 857
    C++ SDK: the Reconhecedor builds the lexicon automaton and finds the terms in a text.Source: README
  20. 858
    C++ SDK: escolher(scores, k) returns the k highest scores in jsonW's deterministic order.Source: README
  21. 859
    Python SDK: the Leitor (reader) opens a .jsw file and returns every row, with the schema and the fields.Source: README
  22. 860
    Python SDK: json(i) and texto(i) return row i as JSON or jsonW-text, without loss.Source: README
  23. 861
    Python SDK: campo(i, name) reads one field of a row without decoding the rest.Source: README
  24. 862
    Python SDK: conferir() checks the CRC32C and the hash of every block in the file.Source: README
  25. 863
    Python SDK: avaliar(request, k) scores every row and returns the k best, with the same result as the tool.Source: README
  26. 864
    Python SDK: the Escritor (writer) writes with compression 0, 1 or 2 and a SHA-256 or BLAKE3 hash.Source: README
  27. 865
    Python SDK: fechar() publishes the file atomically: it only appears complete.Source: README
  28. 866
    Python SDK: abandonar() gives up the write without touching the file that already existed.Source: README
  29. 867
    Python SDK: the Reconhecedor builds the lexicon automaton and finds the terms in a text.Source: README
  30. 868
    Python SDK: escolher(scores, k) returns the k highest scores in jsonW's deterministic order.Source: README
  31. 869
    Node.js SDK: the Leitor (reader) opens a .jsw file and returns every row, with the schema and the fields.Source: README
  32. 870
    Node.js SDK: json(i) and texto(i) return row i as JSON or jsonW-text, without loss.Source: README
  33. 871
    Node.js SDK: campo(i, name) reads one field of a row without decoding the rest.Source: README
  34. 872
    Node.js SDK: conferir() checks the CRC32C and the hash of every block in the file.Source: README
  35. 873
    Node.js SDK: avaliar(request, k) scores every row and returns the k best, with the same result as the tool.Source: README
  36. 874
    Node.js SDK: the Escritor (writer) writes with compression 0, 1 or 2 and a SHA-256 or BLAKE3 hash.Source: README
  37. 875
    Node.js SDK: fechar() publishes the file atomically: it only appears complete.Source: README
  38. 876
    Node.js SDK: abandonar() gives up the write without touching the file that already existed.Source: README
  39. 877
    Node.js SDK: the Reconhecedor builds the lexicon automaton and finds the terms in a text.Source: README
  40. 878
    Node.js SDK: escolher(scores, k) returns the k highest scores in jsonW's deterministic order.Source: README
  41. 879
    Go SDK: the Leitor (reader) opens a .jsw file and returns every row, with the schema and the fields.Source: README
  42. 880
    Go SDK: json(i) and texto(i) return row i as JSON or jsonW-text, without loss.Source: README
  43. 881
    Go SDK: campo(i, name) reads one field of a row without decoding the rest.Source: README
  44. 882
    Go SDK: conferir() checks the CRC32C and the hash of every block in the file.Source: README
  45. 883
    Go SDK: avaliar(request, k) scores every row and returns the k best, with the same result as the tool.Source: README
  46. 884
    Go SDK: the Escritor (writer) writes with compression 0, 1 or 2 and a SHA-256 or BLAKE3 hash.Source: README
  47. 885
    Go SDK: fechar() publishes the file atomically: it only appears complete.Source: README
  48. 886
    Go SDK: abandonar() gives up the write without touching the file that already existed.Source: README
  49. 887
    Go SDK: the Reconhecedor builds the lexicon automaton and finds the terms in a text.Source: README
  50. 888
    Go SDK: escolher(scores, k) returns the k highest scores in jsonW's deterministic order.Source: README
  51. 889
    Rust SDK: the Leitor (reader) opens a .jsw file and returns every row, with the schema and the fields.Source: README
  52. 890
    Rust SDK: json(i) and texto(i) return row i as JSON or jsonW-text, without loss.Source: README
  53. 891
    Rust SDK: campo(i, name) reads one field of a row without decoding the rest.Source: README
  54. 892
    Rust SDK: conferir() checks the CRC32C and the hash of every block in the file.Source: README
  55. 893
    Rust SDK: avaliar(request, k) scores every row and returns the k best, with the same result as the tool.Source: README
  56. 894
    Rust SDK: the Escritor (writer) writes with compression 0, 1 or 2 and a SHA-256 or BLAKE3 hash.Source: README
  57. 895
    Rust SDK: fechar() publishes the file atomically: it only appears complete.Source: README
  58. 896
    Rust SDK: abandonar() gives up the write without touching the file that already existed.Source: README
  59. 897
    Rust SDK: the Reconhecedor builds the lexicon automaton and finds the terms in a text.Source: README
  60. 898
    Rust SDK: escolher(scores, k) returns the k highest scores in jsonW's deterministic order.Source: README
  61. 899
    Java SDK: the Leitor (reader) opens a .jsw file and returns every row, with the schema and the fields.Source: README
  62. 900
    Java SDK: json(i) and texto(i) return row i as JSON or jsonW-text, without loss.Source: README
  63. 901
    Java SDK: campo(i, name) reads one field of a row without decoding the rest.Source: README
  64. 902
    Java SDK: conferir() checks the CRC32C and the hash of every block in the file.Source: README
  65. 903
    Java SDK: avaliar(request, k) scores every row and returns the k best, with the same result as the tool.Source: README
  66. 904
    Java SDK: the Escritor (writer) writes with compression 0, 1 or 2 and a SHA-256 or BLAKE3 hash.Source: README
  67. 905
    Java SDK: fechar() publishes the file atomically: it only appears complete.Source: README
  68. 906
    Java SDK: abandonar() gives up the write without touching the file that already existed.Source: README
  69. 907
    Java SDK: the Reconhecedor builds the lexicon automaton and finds the terms in a text.Source: README
  70. 908
    Java SDK: escolher(scores, k) returns the k highest scores in jsonW's deterministic order.Source: README
  71. 909
    Kotlin SDK: the Leitor (reader) opens a .jsw file and returns every row, with the schema and the fields.Source: README
  72. 910
    Kotlin SDK: json(i) and texto(i) return row i as JSON or jsonW-text, without loss.Source: README
  73. 911
    Kotlin SDK: campo(i, name) reads one field of a row without decoding the rest.Source: README
  74. 912
    Kotlin SDK: conferir() checks the CRC32C and the hash of every block in the file.Source: README
  75. 913
    Kotlin SDK: avaliar(request, k) scores every row and returns the k best, with the same result as the tool.Source: README
  76. 914
    Kotlin SDK: the Escritor (writer) writes with compression 0, 1 or 2 and a SHA-256 or BLAKE3 hash.Source: README
  77. 915
    Kotlin SDK: fechar() publishes the file atomically: it only appears complete.Source: README
  78. 916
    Kotlin SDK: abandonar() gives up the write without touching the file that already existed.Source: README
  79. 917
    Kotlin SDK: the Reconhecedor builds the lexicon automaton and finds the terms in a text.Source: README
  80. 918
    Kotlin SDK: escolher(scores, k) returns the k highest scores in jsonW's deterministic order.Source: README
  81. 919
    Swift SDK: the Leitor (reader) opens a .jsw file and returns every row, with the schema and the fields.Source: README
  82. 920
    Swift SDK: json(i) and texto(i) return row i as JSON or jsonW-text, without loss.Source: README
  83. 921
    Swift SDK: campo(i, name) reads one field of a row without decoding the rest.Source: README
  84. 922
    Swift SDK: conferir() checks the CRC32C and the hash of every block in the file.Source: README
  85. 923
    Swift SDK: avaliar(request, k) scores every row and returns the k best, with the same result as the tool.Source: README
  86. 924
    Swift SDK: the Escritor (writer) writes with compression 0, 1 or 2 and a SHA-256 or BLAKE3 hash.Source: README
  87. 925
    Swift SDK: fechar() publishes the file atomically: it only appears complete.Source: README
  88. 926
    Swift SDK: abandonar() gives up the write without touching the file that already existed.Source: README
  89. 927
    Swift SDK: the Reconhecedor builds the lexicon automaton and finds the terms in a text.Source: README
  90. 928
    Swift SDK: escolher(scores, k) returns the k highest scores in jsonW's deterministic order.Source: README
  91. 929
    PHP SDK: the Leitor (reader) opens a .jsw file and returns every row, with the schema and the fields.Source: README
  92. 930
    PHP SDK: json(i) and texto(i) return row i as JSON or jsonW-text, without loss.Source: README
  93. 931
    PHP SDK: campo(i, name) reads one field of a row without decoding the rest.Source: README
  94. 932
    PHP SDK: conferir() checks the CRC32C and the hash of every block in the file.Source: README
  95. 933
    PHP SDK: avaliar(request, k) scores every row and returns the k best, with the same result as the tool.Source: README
  96. 934
    PHP SDK: the Escritor (writer) writes with compression 0, 1 or 2 and a SHA-256 or BLAKE3 hash.Source: README
  97. 935
    PHP SDK: fechar() publishes the file atomically: it only appears complete.Source: README
  98. 936
    PHP SDK: abandonar() gives up the write without touching the file that already existed.Source: README
  99. 937
    PHP SDK: the Reconhecedor builds the lexicon automaton and finds the terms in a text.Source: README
  100. 938
    PHP SDK: escolher(scores, k) returns the k highest scores in jsonW's deterministic order.Source: README

By industry

How the features apply in each industry. The examples use fictitious data.

  1. 939
    Clinics: the front desk sees the unit's schedule without the patient ID or the amount; finance sees the amount; nobody sees another unit.Source: paper
  2. 940
    Clinics: blood pressure in mmHg is a measure with a unit, converted to Pa when the column asks for it.Source: test with wjsw 1.3.11
  3. 941
    Pharmacies: doses in mg converted to g without decimal-separator mistakes.Source: test with wjsw 1.3.11
  4. 942
    Accounting: exact decimals avoid the cent lost to floating point, and the currency is stored in every value.Source: docs/JSONW.md
  5. 943
    Accounting: the edit history stays in the file, for audits.Source: paper
  6. 944
    Online stores: a SKU as codigo never loses its leading zero.Source: docs/JSONW.md
  7. 945
    Online stores: weight in g, kg or lb converted to the unit of the shipping column.Source: test with wjsw 1.3.11
  8. 946
    Online stores: similar-spelling search finds the product the customer misspelled.Source: Newsroom
  9. 947
    Online stores: the top k by price and rating in a one-line request.Source: paper
  10. 948
    Logistics: a shipping table by distance and weight as a multilinear grid.Source: paper
  11. 949
    Logistics: distances in km, mi or nmi converted to m.Source: test with wjsw 1.3.11
  12. 950
    Energy: consumption in Wh, kWh, MWh or GWh in the same column, converted to J.Source: test with wjsw 1.3.11
  13. 951
    Climate and ESG: emissions in gCO2e, kgCO2e and tCO2e with conversion on write.Source: test with wjsw 1.3.11
  14. 952
    Legal: the hash chain proves that an evidence file has not changed since it was written.Source: paper
  15. 953
    Legal: the history shows every version of a record, with deletions marked.Source: paper
  16. 954
    Trademarks and patents: "Kokakola" finds "Coca-Cola" in a prior-art search.Source: Newsroom
  17. 955
    Trademarks and patents: the whole INPI dataset fits in 1.9 GB at maximum level, and the in-force count comes back in 43 ms.Source: paper
  18. 956
    Human resources: the salary column gets +escopo(financeiro), and only finance sees it.Source: paper
  19. 957
    Customer service: the lexicon separates purchase intent from churn risk in a message.Source: paper
  20. 958
    Customer service: text signals count negation, all caps, exclamation marks and laughter in each message.Source: Newsroom
  21. 959
    Call centers: audio signals measure pauses, pitch and speech rate in each call.Source: Newsroom
  22. 960
    Media: video becomes scenes with measurements and one mosaic per scene.Source: Newsroom
  23. 961
    Multi-tenant SaaS: one file with a tenant column, and another customer's block is never opened.Source: Newsroom
  24. 962
    Open data: the whole public INPI dataset, 12.53 GB of CSV, became 2.58 GB of .jsw.Source: paper
  25. 963
    Internet of things: readings with microsecond timestamps and append-only writing.Source: paper
  26. 964
    Finance: amounts in several currencies in the same column, each with its own ISO 4217 currency.Source: docs/JSONW.md
  27. 965
    Real estate and agriculture: area in m2, ha, km2, ft2 or alqueire, converted to m2.Source: test with wjsw 1.3.11
  28. 966
    Agriculture: production in t, kg or lb in the same mass column.Source: test with wjsw 1.3.11
  29. 967
    Industry: pressure in bar, psi, kPa or atm converted to Pa.Source: test with wjsw 1.3.11
  30. 968
    Industry: temperature in °C, °F or K in the same column.Source: test with wjsw 1.3.11
  31. 969
    Technology: sizes in KB, MB, KiB or GiB without mixing up 1,000 and 1,024.Source: test with wjsw 1.3.11
  32. 970
    Sports and transport: speed in km/h, mph, knots or m/s.Source: test with wjsw 1.3.11
  33. 971
    Research: experiment data with units and a hash on every block.Source: paper
  34. 972
    Government: the same query with different credentials returns only what each agency may see.Source: paper
  35. 973
    Education: semantic search finds the lesson by topic, not just by word.Source: paper
  36. 974
    Customer databases: tax IDs as codigo, with the leading zero kept.Source: docs/JSONW.md
  37. 975
    Hospitality and travel: nightly rates in several currencies, each value with its own currency.Source: docs/JSONW.md
  38. 976
    Restaurants: ingredient weights in g, kg or oz in the same column.Source: test with wjsw 1.3.11
  39. 977
    Laboratories: concentration in ppm and per mille as a dimensionless measure.Source: test with wjsw 1.3.11
  40. 978
    Construction: area in m2 and volume in m3 or L, converted on write.Source: test with wjsw 1.3.11
  41. 979
    Telecom: traffic in B, MB, GB or TB in the same column.Source: test with wjsw 1.3.11
  42. 980
    Fitness and health: energy spent in kcal or kJ in the same column.Source: test with wjsw 1.3.11

Documentation, AI and search engines

jsonW is documented for people and for AI.

  1. 981
    The technical paper is in Portuguese, English, Spanish, Italian, French and Chinese, with a PDF in each language.Source: paper
  2. 982
    Every Newsroom post has a text version for AI.Source: Newsroom
  3. 983
    The whole jsonW website is in all six languages.Source: README
  4. 984
    The GitHub README is in Portuguese and English.Source: README
  5. 985
    The full usage documentation is in docs/JSONW.md.Source: docs/JSONW.md
  6. 986
    The site's llms.txt is at jsonw.org/llms.txt.Source: README
  7. 987
    The sitemap links the six languages of every page for search engines.Source: Newsroom
  8. 988
    The paper says what does not exist yet, instead of promising it.Source: paper
  9. 989
    The paper has a table comparing jsonW with JSON, MessagePack, CBOR, Protobuf, Parquet and Arrow, with sources.Source: paper
  10. 990
    The paper's examples were run with the public binary and show the real output.Source: paper
  11. 991
    The whole ABI is in a single header, sdk/jsonw.h.Source: README
  12. 992
    The package includes ready-made Apache and nginx configurations for wjsw-cgi.Source: AGENTS.md

Quantum, license and transparency

What already holds today and where it is written.

  1. 993
    Custody and credentials use only SHA-256, BLAKE3 and HMAC-SHA256, which do not depend on factoring or discrete logarithms.Source: paper
  2. 994
    The row form's wire protocol uses the post-quantum X25519MLKEM768 key exchange (ML-KEM, FIPS 203); it is not in the public 1.3.11 package.Source: paper
  3. 995
    With Grover's algorithm, finding a SHA-256 preimage or the key of a 256-bit HMAC still costs about 2^128 operations on a quantum computer.Source: paper
  4. 996
    Sustainable Use License: free use inside products, with no server limit.Source: README
  5. 997
    The documentation can be read, quoted and used freely to build programs that use jsonW.Source: README
  6. 998
    The C ABI (jsonw.h) is stable and the same across all SDKs.Source: README
  7. 999
    jsonW is maintained by Wortic, which runs its agents' index, its models and the edge of its infrastructure on it.Source: README
  8. 1,000
    The paper states which binary each measurement used, with its SHA-256.Source: paper