tabellio

Turn an image of a civil-registry or parish record into validated structured JSON.

tabellio: the Roman / medieval scribe who officially drafted legal acts.

The 'vieux tabellion' Georges Brassens addresses in Supplique pour être enterré sur la plage de Sète (1966):

Trempe dans l'encre bleue du Golfe du Lion,

Trempe, trempe ta plume, ô mon vieux tabellion,

Et de ta plus belle écriture,

Note ce qu'il faudrait qu'il advînt de mon corps…

Brassens - INA archive

This library reads back what that pen wrote.

• BYOK — you pass your own VLM API key; it is never stored, never logged.
• Provider-agnostic — Gemini, OpenAI, NVIDIA NIM, Anthropic, or a local Ollama model. None required at install.
• No embedded model, no training, no storage. Reliability comes from a strict Pydantic schema and post-extraction consistency rules, not from trusting the model.
• No silent resolution. Original spelling is kept, every date carries a qualifier (exact / about / calculated / …), every field a confidence.
• Verbatim transcription too. The full diplomatic text of the act, [?] for an illegible word — as act.transcription in full mode, or on its own with output_mode="transcription".
• GEDCOM 7 out of the box. tabellio.to_gedcom(act) — schema built to map cleanly to GEDCOM.
• Any language or script. Text is transcribed in the language of the act (Latin, French, German, …) and never translated; only the schema's type / role vocabulary is fixed English. The detected language is recorded on act.language.

Example

A parish burial act from 1757, a public-domain register page, examples/sample_act.jpg:

export TABELLIO_PROVIDER=...
export TABELLIO_KEY=...
for mode in full simple transcription; do
    python -m tabellio examples/sample_act.jpg --output "$mode"
done

(The .json files land once the reference run against gemini is done.)

Install

pip install "tabellio[gemini]"     # or [openai] / [nim] / [anthropic] / [ollama] / [all]

Requires Python 3.11+.

Use

import tabellio

act = tabellio.parse(
    "register-page.jpg",
    provider="gemini",
    api_key="...",  # your key, BYOK
    act_type_hint="baptism",  # optional
)

print(act.type)  # ActType.BAPTISM
print(act.date.raw)  # "le douziesme jour de may mil sept cens trois"
print(act.date.iso)  # "1703-05-12"
for p in act.persons:
    print(p.role, p.given.value if p.given else None, p.surname.raw if p.surname else None)
print(act.warnings)  # consistency + low-confidence flags

parse() returns an Act Pydantic model. Unread fields are null with a note — never guessed.

Output modes

for mode in ("full", "simple", "transcription"):
    print(tabellio.parse("register-page.jpg", output_mode=mode))

• output_mode="full" (default) — rich Act: raw spelling, per-field confidence, date qualifier, transcription, validation warnings. act.date is the date of the act; the subject's real birth / death (often a different day — "né la veille", "décédé hier") lands on persons[i].birth_date / death_date with qualifier: "calculated" + a note when deduced. validate warns if a baptism or burial has no such date on its subject.
• output_mode="simple" — a bare ActSummary: type, date (ISO string or null), location, persons (role / given / surname). Nothing else — no confidence, no raw, no qualifiers, no language, no life-event dates, no transcription. Shortest prompt, fewest tokens.
• output_mode="transcription" — a Transcription: {text, language} — the verbatim text of the act ([?] / [illegible] for gaps) and its detected language, nothing else. Still JSON, just two fields.

Hints

tabellio.parse(
    img,
    act_type_hint="burial",  # --hint
    act_language_hint="fr",  # --lang
    context="The deceased is Joseph BOMMAL, buried 21 December 1757.",  # --context
)

All three are optional. act_type_hint / act_language_hint are guesses the model verifies against the image. context is free text of whatever you already know — names, a date, a place. It is used only to disambiguate hard-to-read passages: it never overrides a clear reading, raw stays faithful to the page, and in full mode the model adds a note wherever it followed the image against your context.

GEDCOM

act = tabellio.parse("register-page.jpg")  # the one network call
print(act.model_dump_json(indent=2))  # JSON
print(tabellio.to_gedcom(act))  # a complete GEDCOM 7.0 document

to_gedcom(act) writes one SOUR record for the act, one INDI per person it names, and a FAM for the couple or parent-child link the act states — BIRT / BAPM / MARR / DEAT / BURI events with DATE / PLAC / AGE, ASSO + ROLE for witnesses and godparents, QUAY from confidence. It transcribes one act; it does not merge a tree — your genealogy software does that on import. Output is conformance-checked against the gedcom7 library in the test suite.

CLI: python -m tabellio examples/sample_act.jpg --format gedcom.

The schema is GEDCOM-mappable by design: GenDate.qualifier (about/before/after/calculated → ABT/BEF/AFT/CAL), GenDate.calendar (julian, french_republican), Person.name_particle (SPFX), Person.name_suffix (NSFX).

Configuration

Three optional environment variables, each a fallback for the matching parse() argument. An explicit argument always wins. tabellio reads them from os.environ — it does not load a .env file; your application manages its own environment.

The key is never logged and never stored.

Command line

export TABELLIO_PROVIDER=gemini
export TABELLIO_KEY=...              # in your shell
python -m tabellio examples/sample_act.jpg [--hint baptism] [--lang fr] [--context "..."] \
    [--output simple|transcription] [--format gedcom] [-v]

Prints the result as JSON — Act, ActSummary or Transcription per --output — on stdout, warnings on stderr; --format gedcom prints a GEDCOM 7 document instead. The key is only read from the environment, never a CLI argument.

Providers

Set TABELLIO_MODEL (or pass model=) to switch model — quality vs speed is your call. Keyword options pass straight through: timeout= (seconds, default 120), base_url=, host=, max_tokens=.

NIM caveat. On the hosted free endpoint (integrate.api.nvidia.com) the 90B vision model routinely times out; the default is therefore meta/llama-3.2-11b-vision-instruct, which answers in seconds but is a small model — it hallucinates names and dates on old cursive and returns poor results here. Use --output simple with it, and treat gemini as the real path. The endpoint also rejects inline images larger than ~180 KB: tabellio re-encodes an oversized image in memory (lower JPEG quality, then downscale, long edge kept ≥ 1200 px) before sending — the source file is never touched — and logs a warning; pass shrink=False to get an error instead. NVCF asset upload is not implemented.

Local / general-purpose VLMs hallucinate plausible names and dates on old cursive. The ollama provider exists for convenience and testing, not for production genealogy.

Scope

In: extraction of a single record to schema, ambiguity flags, confidence surfacing, provider-agnostic adapter, JSON and GEDCOM 7 output.

Out: embedded/trained model, home-grown HTR, storage, accounts, billing, GEDCOM import, building or merging a family tree, image segmentation / deskew, translation of any transcribed text.

Development

make install     # uv sync --all-extras
make lint        # ruff check + format --check
make test        # pytest

To run python -m tabellio against a real provider while developing, export TABELLIO_KEY (and optionally TABELLIO_PROVIDER / TABELLIO_MODEL) in your shell, or use your own direnv / dotenv — the repo ships none.

The examples/ directory holds a sample act and its extracted output — fictional or public-domain only, see examples/README.md.

License

Apache-2.0. See LICENSE and NOTICE.