funkpower
ba8ec81735
A catalogue's worst records are the ones whose file says nothing true about itself: a PDF titled `Untitled` by an author called `Oxford`, which is a publisher, from a filename that was never anything but a filename. What those books do have is pages. The title page carries the title as the publisher set it; the copyright page carries the imprint, the year, the ISBN, and in most English books since the 1970s the cataloguing block, which is a catalogue record a librarian already made, subject headings and all. So this reads the pages. Every field names the page it came from, and a book that printed nothing comes back empty with its fields in `unsettled` rather than filled with something plausible — an extraction that starts at chapter one has no front matter in it, and saying so is the answer. Two pages look like a title page and are not, and both are how a catalogue comes to name the wrong book. A series advertisement lists other titles in the shape a title page uses, over as many as four pages with only the first announcing itself. A translation's cataloguing block prints the original title in brackets above the real one. Both are found and refused. Three things the pages give that no file does: the subject headings a librarian assigned, an ISBN that exists only in print — which turns a book with no identifier into one `bibfetch` can look up — and a title in the case the publisher meant, because a half title prints unbroken what the title page breaks across two lines and sets in capitals. `--prompt` and `--merge` are citefind's: a question about exactly the fields the pages left unsettled, composed here and answered wherever the caller keeps a model, with the advertising taken out of what the model is shown. A field a model answered is marked `model`, so a record always says which of it was read and which was inferred. Measured against 600 books of the home corpus: the cataloguing block answered for 21% of them and agreed with the catalogue 89% of the time, most of the residue being the catalogue's error rather than this one's. Title pages answered for another 41%. The remaining 38% print nothing readable, and are what the prompt is for.
catena
Small programs that each do one part of turning a pile of files into a searchable corpus, and hand the next one text on stdout.
A catena is a commentary assembled out of excerpts drawn from many sources and chained in order, each one carrying the name of who said it and where. That is both what these programs produce and how they run.
Nothing here needs a database, a GPU, a queue or a config file. Each tool is a filter: bytes in, records out, diagnostics on stderr, and an exit code that means something.
Install
cargo install --git https://git.pumpin.online/funkpower/catena <tool>
Each tool installs by name, so you can take one without taking the rest.
The tools
booksniff |
What a file actually is, whatever its name claims |
bookmeta |
The metadata a file already carries |
colophon |
The record a book prints in its own front matter |
tocfind |
A book's own table of contents |
langid |
What language a document is in |
pdfgutter |
Books scanned two pages to a sheet, split back into pages |
pdftriage |
Whether a PDF's text layer is worth reading, or whether it has to be OCRed |
bibclean |
Catalogue records audited and repaired, with a reason for every change |
vocabmerge |
The near-duplicate headings in a catalogue's vocabulary |
seriesof |
The series a book belongs to, and which volume this is |
idnorm |
DOIs, ISSNs, ISBNs and arXiv ids, read and checked |
bibfetch |
What a register says about an identifier |
citefind |
The references in a bibliography or a page of notes |
mdscrub |
Converted markdown in, clean markdown out |
mdsplit |
One markdown file into the works inside it |
chunkmd |
Markdown into passages that carry their page, chapter and role |
cuenorm |
Any subtitle format in, one normalised cue stream out |
cuepack |
Cues into passages worth embedding, keeping the times they were said at |
cueseg |
A timed transcript into topic chapters |
emb |
Lines into vectors, against an embedding server |
rrf |
Several ranked lists into one, without the database |
avplan |
Whether a recording plays, and the one ffmpeg call that fixes it |
Each links to its own page: what it does, when you would use it, worked examples, options and exit codes.
→ Examples — programs that run. A book to passages and references, a recording to chapter marks and a copy that plays on a phone, and a hybrid search fused, in shell, Rust, Python and TypeScript.
→ Recipes — using these tools together. A scanned book to a searchable passage, a transcript to a citation with a timestamp, vectors into pgvector, and hybrid search fused on top.
The libraries underneath
Every tool is a library crate with a thin command on top, so anything a command does can be done in process instead.
corpus-core |
The record types and the stream plumbing everything shares |
textkit |
Token counting, sentence and paragraph splitting |
cuekit |
The cue model, and the three subtitle formats it is read from |
catena-cli |
The parts of a filter every command does the same way |
bindings |
The same libraries from Python, C and JavaScript |
From another language
The tools are filters first, and a subprocess is the right answer most of the time. When the caller is already holding the text, the same libraries are callable in process from three languages:
| Install | Import | |
|---|---|---|
| Python | pip install catena-corpus |
import catena |
| JavaScript | npm install catena-corpus |
import { chunkMarkdown } from "catena-corpus" |
| C | build the capi feature |
#include <catena.h> |
chunks = catena.chunk_markdown(markdown, target=512)
chunks[0]["page"] # 41
Each hands back the same record shapes the commands write — dictionaries in
Python, plain objects in JavaScript, JSON in C — so a stage that was parsing
NDJSON from a subprocess can call the function and get what it was already
parsing. Details in bindings.
A worked pipeline
Converted markdown in, retrievable passages out:
$ pandoc -t commonmark chapter.xhtml | mdscrub | chunkmd --min-tokens 20
{"v":1,"idx":0,"text":"The obligation ran both ways, and contemporaries said so plainly\nwhenever they were asked.4 A lord who could not\nprotect was a lord who could not command…","tokens":56,"page":41,"chapter":"III. The Second Estate","role":"body","start_time":null,"end_time":null,"flags":[]}
{"v":1,"idx":1,"text":"What the rolls do not say, and what has to be inferred, is how often the\nobligation was actually tested…","tokens":143,"page":42,"chapter":"III. The Second Estate","role":"body","start_time":null,"end_time":null,"flags":[]}
A recording instead of a book, with chapter marks a scrubber can use:
$ cuenorm episode.vtt | cueseg --duration 7200
{"v":1,"start":0.0,"end":1190.6,"title":"The Mill Stood At The Northern End","excerpt":"the mill stood at the northern end of the valley where…"}
{"v":1,"start":1190.6,"end":2381.3,"title":"Parliament Dissolved Itself Again","excerpt":"parliament dissolved itself again amid furious objections…"}
And before any of that, whether a scanned PDF is worth reading at all:
$ pdftriage scan.pdf
garbled, needs OCR
1820 characters on the opening pages
1820 characters over 5 sample pages, ascii text score 0.15
letters 0.13, latin-1 0.29
$ for f in *.pdf; do pdftriage --quiet "$f" && ocrmypdf "$f" "ocr/$f"; done
What they promise
Eight things hold for all of them, and together they are what lets any of these go into a pipeline without first reading how it works:
- Each does a single job, with no mode switch that changes what it is.
- Input on stdin, results on stdout, diagnostics on stderr — never mixed.
- Anything list-shaped is NDJSON, one record per line, flushed as produced.
- Exit codes carry the answer:
0yes,1no,2could not tell, and failures in thesysexitsrange, so "no" and "broke" are never the same number. - Nothing reaches the network, a database, a config file or any stored state,
bar two exceptions that say so in
--help. - Nothing is assumed about where it is running: no hostnames, no paths, no private vocabularies built in.
- The same input gives the same output, on every run and every machine.
--jsonis the stable form, and does not change shape without a major version.--manwrites the program's manual page, generated from the same definition--helpis.
The whole of it, with the reasoning: what these programs promise.
The records
Five shapes travel between the tools, versioned, defined in
corpus-core and described in the
records.
$ cargo run -p corpus-core -- schema all
Build
./check.sh
Formatting, lints, tests and documentation, which is the same gate CI runs.
Licence
MPL-2.0, library crates and binaries alike. The notice lives in LICENSE and
the license field of every manifest rather than in a header on every file,
which is what MPL-2.0 Exhibit A allows for.