saucier

Structured procedure extraction from culinary sources.

Reads a cookbook and returns a catalogue. Each preparation carries its names, the language of each name, the parent it derives from, and the source line the claim came from. No model runs. Every output is traceable to the text.

Try it

No GPU, no API key, no database, no network. Both source texts are committed.

$ git clone https://github.com/Alberto-Codes/saucier
$ cd saucier && uv sync
$ uv run saucier parse
escoffier-1909  New and Revised Edition, January 1909 (impression: January 1920)
                transcription of Project Gutenberg 71395
                mothers: bechamel, espagnole, hollandaise, tomato, veloute
                151 sauces, 57 derived, 94 unresolved
escoffier-1907  no edition stated, copyright 1907
                ocr of Internet Archive cu31924000610117
                mothers: bechamel, espagnole, hollandaise, tomato, veloute
                140 sauces, 51 derived, 89 unresolved

The five mothers were not supplied. Escoffier names them in his own text, and the parser reads them out of it.

Neither is the edition supplied. Each source states its own printing history on its own title page, and the id is built from that reading. See ADR-0009.

$ uv run saucier tree espagnole
BROWN SAUCE OR ESPAGNOLE  [espagnole]  derives from brown-roux
├── HALF GLAZE  (en)
│   ├── SAUCE BORDELAISE  (fr)
│   │   └── MARROW SAUCE  (en)
│   ├── BROWN CHAUD-FROID SAUCE  (en)
│   ├── DEVILLED SAUCE  (en)
│   ├── ITALIAN SAUCE  (en)
│   ├── LYONNAISE SAUCE  (en)
│   ├── MADEIRA SAUCE  (en)
│   ├── PIQUANTE SAUCE  (en)
│   └── ROBERT SAUCE  (en)
├── LENTEN ESPAGNOLE  (fr)
│   └── GENEVOISE SAUCE  (en)
├── ORDINARY POIVRADE SAUCE  (en)
└── POIVRADE SAUCE FOR VENISON  (en)

A parent may be any catalogued preparation, not only a mother. Robert states half glaze, half glaze states Espagnole, and Espagnole states brown roux. A mother is what the source names as foundational, and two of the five now state a parent of their own.

The interesting number is 94

94 of 151 preparations state no base in their opening paragraph, or state more than one and choose neither. Escoffier wrote for a reader who already knew, so he did not write it down. That gap is the point: it is the measured bar that anything cleverer has to beat. See why there is no model in this yet.

The catalogue holds 151 entries, not every sauce in the book. An entry gets in when its own heading says "sauce", or when Escoffier filed it in a chapter he titles as sauces. Both tests read the source. Neither guesses.

Derived rose from 50 to 57 in this release, and that seven is three numbers. Twelve sauces gained the parent Escoffier wrote, most of them half glaze. Ten lost one. A newly catalogued butter or half glaze now sits in their opening paragraph beside the old parent, and a resolver may refuse, never rank. Five admitted entries state a parent. saucier show cardinal-sauce prints both names it saw. See ADR-0015.

The parent has a verb

A parent says what a sauce is built from. MORNAY SAUCE now says what is done with it:

$ uv run saucier show mornay
MORNAY SAUCE
entry 91, line 2437, transcription of escoffier-1909
  term  MORNAY SAUCE  [en]  mornay-sauce
  parent  bechamel
  procedure  6 operations, recorded by hand
    Boil      Béchamel Sauce [fr] 1 pint, fumet [fr] 1/4 pint
    Reduce    criterion: by a good quarter (unresolved)
    add       Gruyère [fr] 2 oz., Parmesan [en] 2 oz.
    Put       duration: a few minutes (unresolved), on the fire again
    stirring  instrument: small whisk, criterion: the melting of the cheese (unresolved)
    Finish    butter [en] 2 oz., away from the fire, added by degrees

Every word is quoted from the entry, and the command checks each one against the body before it prints. a few minutes names no number, so its number is unresolved, the same absence a parent records. One preparation is recorded in this release, by hand, once per witness. The scan's copy reads Bdchamel, so its parent is lost where its verb is not. See ADR-0017.

Two witnesses, and a diff that names its causes

The corpus holds two texts of one book. One is proofread and one is a raw scan, so every difference between them has two explanations.

$ uv run saucier diff escoffier-1907 escoffier-1909
...
  11 unmatched, 18 parent-changed, 35 ocr-suspected
  entries read  2679 of escoffier-1907, 2963 of escoffier-1909, a blind spot of 284

QENEVOISE SAUCE is GENEVOISE SAUCE with one letter misread, and the diff says so rather than reporting a removal it cannot support.

It reports no additions at all. The 1907 witness is a scan and 284 of its entries are unread. So a concept found in one witness and not the other is unmatched rather than added. The diff has confirmed no editorial difference between the two printings. Every candidate it has raised has been the scanner or the reader. One difference has been confirmed by hand. Mornay's fumet is of that fish in 1907 and of the fish, poultry, or vegetable in 1909, and ADR-0017 quotes both lines.

One record per line, for whatever comes next

The JSON files under data/ are a snapshot for this program. saucier export writes the same catalogues as JSON Lines, one self-describing record per line. saucier import --check proves the stream rebuilds them and writes nothing:

$ uv run saucier export | uv run saucier import --check
escoffier-1909  151 sauces, 57 derived, 94 unresolved
escoffier-1907  140 sauces, 51 derived, 89 unresolved
2 catalogues and 291 preparations rebuilt. Nothing written.

Every record carries a schema version, a type, and a source-local id. A loader in any language reads one line without this package. Two exports of one data/ directory are byte-identical. A serialization and a database are different decisions, and this release makes only the first. See ADR-0016.

Documentation

The rendered site is https://alberto-codes.github.io/saucier/. Its API reference is generated from the docstrings, so it cannot drift from the code.

• Tutorial — clone it and print a family tree
• How-to — point it at a different book
• Reference — commands and data model
• Explanation — why a parser comes first
• Decisions — the seventeen records, and what each one accepted

Licence

MIT for the code.

corpus/escoffier-1909.txt is Project Gutenberg ebook #71395, public domain, redistributed with its licence wrapper intact. It transcribes the January 1920 impression of the New and Revised Edition of January 1909.

corpus/escoffier-1907.txt is Internet Archive item cu31924000610117, the Cornell University Library copy of the 1907 first printing. Cornell records no known United States copyright restrictions on the text, and a 1907 publication date puts it in the public domain.

No third-party corpus is shipped in the published package.