paratext-cli 0.1.0__py3-none-any.whl

paratext/AGENTS.md

@@ -0,0 +1,254 @@
+# paratext — agent guide
+
+Modular metadata-extraction pipeline for digitised library/archive collections,
+driven by a multimodal model. This orients AI coding agents; `README.md` is the
+human guide.
+
+paratext is **built to be adapted**. Most users need a project of their own, and
+some need a new input format or a new export format. Those are extension points,
+not forks — the recipes below are the supported way in.
+
+## Rules
+
+- **Read a file in full before editing it.** These modules are short by design;
+  there is no excuse for patching blind.
+- **Run `uv run pytest -q` and `uv run ruff check` after every change.** Both must
+  be clean. Line length 100, rules E/F/I/W.
+- **Never edit `prompt.md` and `schema.py` casually.** A prompt edit rolls a new
+  review round and invalidates comparisons; a schema change needs a
+  `schema_version` bump. Both cost a model run to re-evaluate.
+- **Keep a project's fields in step across all three places** (schema, prompt,
+  view) and call `audit_project()` from its tests.
+- **Add a notice, not a silent fallback.** If you write a degraded path, append to
+  `Source.notices` so `extract` reports it. Preprocessing that silently does
+  nothing looks like a clean run and quietly costs accuracy — it is the worst
+  failure mode in this codebase.
+- **Don't hardcode institution-specific values.** Detector weights, verso
+  thresholds and grid regions are all config.
+- **Ask before removing functionality** that looks deliberate.
+- Say "multimodal model" or "the model" — never "VLM".
+
+## Setup, build, run
+
+```bash
+uv sync --extra dev          # add --extra detector for the torch card detector
+uv run paratext …            # run the CLI against local source
+uv run pytest -q
+uv run ruff check
+```
+
+Python 3.11–3.13 via `uv`. The model endpoint is any OpenAI-compatible server;
+set `base-url` in `paratext.toml` or `PARATEXT_BASE_URL`.
+
+**The installed-vs-source trap:** a non-editable install means source edits do
+nothing until reinstall. `paratext inspect` reports the *installed* project — when
+behaviour doesn't match the source you're reading, check there first.
+
+## Architecture
+
+One CLI runs the loop: `extract` (model → JSONL) → `package` (JSONL → review
+dataset) → `review` (web UI) → `export` (publish). `run` does extract+package.
+
+| Module | Responsibility |
+| --- | --- |
+| `cli.py` | Argparse wiring, round resolution, command bodies |
+| `config.py` | `paratext.toml` + `PARATEXT_*` resolution |
+| `projects/` | The `Project` plug-in contract, `View`, `audit_project` |
+| `sources.py` | Input adapters (`image_source`, `pdf_source`) |
+| `extract.py` / `runner.py` | Sample loop; model call, retries, image encoding |
+| `packaging.py` | JSONL → `samples.json` + `images/` + `view.json` |
+| `store.py` | SQLite annotations + gold labels — **not** web code |
+| `datasets.py` | Discovering packaged rounds, resolving `view.json` |
+| `records.py` | Format-neutral gold selection, shared by all exports |
+| `catalogue.py` / `hf_export.py` | MARC/DC and Hugging Face exports |
+| `review/server.py` | HTTP layer only; `review/static/` is the vanilla-JS frontend |
+| `cards.py` | Optional scanned-card tools (verso, crop, show-through) |
+| `carbon.py` | Grid-intensity providers for `--green` |
+
+`store.py` and `datasets.py` are deliberately outside `review/` so exporters can
+read a reviewed round without starting a web server.
+
+Output is JSONL with a `_provenance` header (git commit, prompt hash, model,
+schema version, timestamp), and is resumable: re-running with the same `--output`
+skips ids already present.
+
+---
+
+# Extending paratext
+
+## Recipe: a new project
+
+The common case — a user's own collection. `paratext new <name>` scaffolds it;
+these are the files it makes and what to edit.
+
+```
+my_cards/
+    prompt.md     # behaviour lives here
+    schema.py     # Pydantic model = the fields, also sent as response_format
+    __init__.py   # wires them into PROJECT
+```
+
+```python
+from paratext.projects import Project, load_prompt
+from paratext.sources import image_source
+
+from .schema import Record
+
+PROJECT = Project(
+    name="my-cards",
+    schema_version="v1",
+    prompt=load_prompt(__file__),
+    schema=Record,
+    source=image_source(),
+)
+```
+
+Register it in the user's `pyproject.toml`:
+
+```toml
+[project.entry-points."paratext.projects"]
+my-cards = "my_cards:PROJECT"
+```
+
+Then `uv sync` — entry points are re-scanned per request, so the review server
+picks up a new project without a restart, but **framework code changes still need
+one**.
+
+Optional, in order of how often you'll need them:
+
+- `view=View(...)` — curate the review display. Defaults to every schema field.
+- `curate(rec) -> Curation` — `keep` / `drop` / `quarantine` per record.
+- `build_record(rec, images)` — extra keys in `samples.json`.
+- `ground_truth(rec)` — attach existing catalogue data for side-by-side review.
+- `disable_thinking` — some models need `enable_thinking=False`.
+
+**Field discipline:** the field names appear in schema, prompt and view with no
+compile-time link. `audit_project(PROJECT)` checks the view's fields exist in the
+schema and that every model-output field is named in the prompt. Call it from a
+test — `paratext new` generates one.
+
+Put behaviour in `prompt.md` (prose, examples, versioned by the prompt hash).
+Keep `Field(description=...)` short and structural: it is *also* sent to the
+model and shouldn't restate the prompt in a second voice.
+
+## Recipe: a new source adapter
+
+For an input shape neither `image_source` nor `pdf_source` covers — a IIIF
+manifest, a METS/ALTO tree, a CSV of URLs, a database query.
+
+A `Source` is two functions that must agree on the metadata they pass between
+them, plus optional notices and a descriptive config:
+
+```python
+from paratext.sources import Source
+from paratext.projects import Sample
+from paratext.packaging import save_image
+
+def iiif_source(*, max_items: int | None = None) -> Source:
+    notices: list[str] = []
+
+    def _iter(source: Path, limit: int | None) -> Iterator[Sample]:
+        # yield one Sample per unit of work
+        yield Sample(
+            id="unique-stable-id",           # becomes the JSONL/review key
+            images=[pil_image],              # what the model sees
+            metadata={"iiif_url": url},      # anything materialise() will need
+        )
+
+    def _materialise(rec: dict, out: Path, max_size: int) -> list[str]:
+        # write review images under `out`, return paths relative to it
+        rel = f"images/{rec['id']}/image.jpg"
+        save_image(local_path, out / rel, max_size)
+        return [rel]
+
+    return Source(
+        iter_samples=_iter,
+        materialise=_materialise,
+        notices=notices,
+        config={"kind": "iiif", "max_items": max_items},   # shown by `paratext inspect`
+    )
+```
+
+Rules specific to sources:
+
+- **`id` must be stable across runs** — resume and round-updating both key on it.
+- **`metadata` is the contract between the two halves.** `_materialise` runs at
+  packaging time, long after iteration, and gets only the JSONL record.
+- **Degrade loudly.** Falling back? `notices.append(...)`, and say what to do
+  about it.
+- **`config` is descriptive only** — nothing reads it back to make decisions; it
+  exists so `paratext inspect` can show what preprocessing was applied.
+- If your images need no special handling, reuse `packaging.default_materialise`
+  rather than writing the same six lines again.
+
+A source can also pre-classify a sample to skip the model entirely: set
+`metadata["preclassified"] = {...}` and `extract` writes it straight through
+(this is how the verso filter avoids paying for blank card backs).
+
+## Recipe: a new export format
+
+For a shape the built-in `hf` / `marc` / `dc` don't cover — EAD, MODS, a local ILS
+format, a CSV for a spreadsheet workflow.
+
+**Never re-derive which records are gold.** `records.select_records()` is the
+single source of truth and every format shares it:
+
+```python
+from paratext.records import select_records
+
+sel = select_records(dataset_dir, project, db_path=db_path)
+for rec in sel.records:
+    rec.label            # {field: value} — the gold label
+    rec.status           # verified | corrected | rejected
+    rec.verdict          # the model's original verdict, kept for corrected rows
+    rec.document_id      # stable id, falls back to rec.sid
+    rec.images           # resolved paths (may be empty)
+sel.schema_fields        # field order from the Pydantic schema
+sel.provenance           # model, prompt_hash, schema_version
+```
+
+Then follow `catalogue.py`'s shape — a pure `build_*(records, mapping)` that
+returns bytes or a tree, and a `run()` that selects, builds and writes:
+
+1. Write `build_ead(records, mapping) -> ET.ElementTree` with no I/O, so it's
+   unit-testable without a dataset on disk.
+2. Add a `run(dataset_dir, project, fmt)` that calls `select_records`, builds,
+   and writes to `EXPORT_ROOT`.
+3. Plumb it into the CLI in `cli.py`: add the name to `--format`'s `choices`, to
+   `_FORMAT_MENU`, and to the `_cmd_export` dispatch.
+4. If fields need mapping to a target vocabulary, reuse `resolve_mapping()` —
+   config under `[project.<name>.export.<fmt>]` wins, then canonical inference by
+   field name, and anything left goes to the wizard. Unmapped fields are dropped
+   with a warning, **never** a hard error.
+
+To expose it in the review UI's export modal as well, add a tab in
+`review/static/app.js` (`openExportModal`) and an endpoint in `review/server.py`
+alongside `_api_export_catalogue`.
+
+---
+
+## Conventions
+
+- **Rounds.** `run` writes `review/<project>-r<N>`, keyed on the **prompt hash**:
+  editing `prompt.md` rolls a new round (the UI diffs the two latest); re-running
+  the same prompt updates the current round in place, keeping its annotations.
+  Rounds are linear — reverting a prompt starts a new round, it doesn't return to
+  the old one. `--round N` forces; `--fresh` rebuilds and discards annotations.
+- **Reading feedback back into the prompt.** `annotations.db` holds verdicts and
+  notes (`annotations`) and human-corrected answers (`gold_labels`); read it via
+  `store.py`. With the server running, `GET /api/stats` gives accuracy and
+  eval-gold size. Tighten the fields reviewers corrected most.
+  The `annotations.corrections` column is unrelated — handwritten corrections on
+  the *card*, not reviewer edits.
+- **Export gold** is `good_enough` rows (`_label_status: verified`) plus corrected
+  `gold_labels` rows (`corrected`). `_verdict` preserves the model's original
+  verdict, so accuracy still measures the model, not the human who fixed it.
+- **`paratext.cards`** (verso filter, RetinaNet crop, show-through suppression) is
+  **off by default** and calibrated on one library's scans. Never assume it
+  transfers to another collection unchanged.
+- **`paratext carbon` / `run --green`** is opt-in carbon-aware scheduling; the grid
+  region is declared in `[carbon]`, never auto-detected. Stdlib `urllib` only.
+- **Config keys are kebab-case** (`base-url`), matching the CLI flag that sets
+  them. Snake_case still parses so old configs work, but generate kebab-case.
+- **Bump `schema_version`** when a schema changes.
+- **Commits:** short messages, no Co-Authored-By trailer.

paratext/__init__.py

@@ -0,0 +1,3 @@
+"""paratext — metadata extraction from digitised collections with a multimodal model."""
+
+__version__ = "0.1.0"