Codify

Codify turns a statute book into law a machine can read: each act structured as Akoma Ntoso 3.0, the open standard for legal documents, addressable by FRBR URI and citable to the provision, with quoted amending text lifted out of the prose into markup a machine can address.

This repository provides all the components of the core pipeline: acquisition, transcription, anchor-driven structuring, retrieval and comparison. Documents converted from scans include coverage metrics and validation reports, while pre-existing Akoma Ntoso documents skip the scan passes and are normalised (identifiers, unique eIds) and validated.

The code is Apache 2.0 and fully standalone. While TBI offers commercial services built on top of it, this core pipeline requires no external proprietary services and can be run independently.

This package contains the domain logic for manipulating legal text, exposed as reusable functions and classes. Application-level concerns like HTTP routing and workflow orchestration are downstream.

Quick start

You need:

• Python 3.12 or newer and uv
• poppler for scanned PDFs (brew install poppler / apt install poppler-utils); text input needs nothing
• access to a chat model behind an OpenAI-compatible endpoint. The repo makes it easy to add a Gemini API key, but a LiteLLM gateway will also easily work
• Docker, only for the database tests

To use the library from your own project, uv add codify-core (or pip install codify-core); the extras are serve, mcp and migrations. To work in this repository:

uv sync
cp .env.example .env   # add your key
uv run codify ingest-one data/fixtures/synthetic/xa/legislation-act-1992.pdf --jurisdiction xa --out bundle/

The sample command processes a five-page synthetic act from xa, a fictional test jurisdiction. It takes under a minute, costs a few cents in model API calls, and outputs locally without touching any database.

The generated bundle/ directory contains:

Structuring follows an anchor-driven model: a deterministic skeleton is parsed using the jurisdiction's rules, and the LLM then fills the text within each anchored block.

Inspection order. Inspect anchors.jsonl before final.akn.xml. The model only populates text inside anchored basic units (e.g., sections, articles). If a pass identifies higher-level containers but no basic units, it logs body_fill_skipped and exports only the skeleton. If no anchors are found, it logs scaffold_no_anchors and preserves the source text unparsed. (Note: the bundled xa fixture places section numbers in the margin, which the parser does not currently extract; it anchors six parts but no sections.)

Dry runs. To inspect detected anchors across .txt and text-layer .pdf files without making model calls or requiring a database, run the anchor scan directly:

uv run codify scan-corpus <dir> --jurisdiction xa

The summary is JSON on stdout; diagnostics such as an empty sweep go to stderr, and --per-document writes one row per scanned document to a path of your choosing.

Log messages.

• layout_pass_failed ... OcrNotConfigured: the optional secondary OCR engine (Azure AI Foundry) is not configured. The pipeline reads the scan once via the model's vision endpoint rather than twice, and the run otherwise proceeds.
• page_diverted_to_ocr: the page's extracted text layer was rejected (the event's reason names the test it failed: too short, garbled, letter-spaced, presentation forms, divergent from the scan, or a visible annotation) and the page was rasterised for OCR.

Store, search, compare

The bundle is the end of ingest-one. To search it, or set it beside another version, load it into Postgres. The compose file builds one with pgvector and pg_textsearch:

docker compose up -d --wait postgres          # first build takes a minute or two
uv run alembic -c alembic.ini upgrade head
uv run codify load bundle/ --embed
uv run codify search "right of access" --jurisdiction xa

load reads final.akn.xml and manifest.json from a bundle and writes the law, its version and its provisions; the jurisdiction and the title come from the manifest, so a bundle needs no flags. A bare .akn.xml has no manifest and needs --jurisdiction and --title. --embed also embeds the provisions, which search needs. Loading the same file twice returns the version it already holds. search runs the hybrid retrieval (lexical and vector, fused) and prints one JSON match per line, best first, with the provision's eId, score and text; it exits 1 when nothing matched.

Embeddings go to an OpenAI-compatible embeddings endpoint. EMBEDDING_BASE_URL, EMBEDDING_API_KEY and EMBEDDING_MODEL name it; unset, the LITELLM_* chat settings are used, so one gateway can serve both.

uv run codify compare a.akn.xml b.akn.xml --out report.json
uv run codify compare <version-id> <version-id>

compare aligns the second document's provisions against the first, article by article, through the chat model, and writes a report with a summary (aligned, partial, gap) and every alignment. Either side is an AKN file or a stored version id. It costs a model call per provision.

POSTGRES_URL names the database. Unset, it is the compose one while ENVIRONMENT is unset, localhost or development; under any other value an unset URL is an error.

Serve

The same reads and the ingest, behind HTTP, for a script or a UI that is not on the box:

uv sync --extra serve
uv run codify serve                            # http://127.0.0.1:8000, docs at /docs
curl -F file=@act.pdf -F jurisdiction=xa localhost:8000/runs/ingest
curl -N localhost:8000/runs/<run-id>/stream    # server-sent events until the run ends

Reads: /jurisdictions, /jurisdictions/{code}, /laws, /laws/{id}, /versions/{id} (with the AKN), /search?q=&jurisdiction=. Ingest: POST /runs/ingest (a file) or POST /runs/ingest-url (a URL) return a run at once; /runs/{id} is its state, /runs/{id}/stream replays every event so far and then follows it, and /runs/{id}/cancel and /runs/{id}/retry do what they say. A succeeded ingest is stored, so the reads see it. ingest-url takes an EU publications URL (eur-lex or publications.europa.eu) under jurisdiction eu, the only lane that fetches; a PDF goes through ingest. Runs live in the server's memory: a restart forgets them, each run says so (lost_on_restart), and the 200 most recent finished runs stay readable. There is no authentication; bind it to localhost or put it behind something that has.

Contract

contract/openapi.json is the server's OpenAPI schema, generated from the routes' response models (codify/serve/schemas.py and the library models they carry, such as JurisdictionConfig); a test fails when it drifts. After a route change:

uv run codify serve --openapi > contract/openapi.json

A client generates its types from that file (the UI does, with openapi-typescript), so the schema is the one place the two agree.

MCP server

The same reads, and compare, as tools for an agent over the Model Context Protocol:

uv sync --extra mcp
uv run codify mcp                     # stdio, for a client that launches the server
uv run codify mcp --http --port 8001  # streamable HTTP at http://127.0.0.1:8001/mcp

A client that speaks stdio launches the command itself; this is the shape most take:

{
  "mcpServers": {
    "codify": {
      "command": "uv",
      "args": ["run", "--directory", "/path/to/codify", "codify", "mcp"],
      "env": { "POSTGRES_URL": "postgresql://codify:codify@localhost:5432/codify" }
    }
  }
}

The tools, all reads:

• search_provisions(query, jurisdiction, language?, k?): hybrid search over one jurisdiction's provisions; matches best first, each with its provision id, eId, score and text. Needs the embeddings endpoint, as search does; unset, the tool says so.
• list_laws(jurisdiction?, doctype?, year?, q?, limit?, offset?): a page of stored laws with their ids.
• get_law(law_id): one law's fields, its jurisdiction and every stored version.
• get_version(version_id, include_xml?): one version's metadata; with include_xml, the Akoma Ntoso XML, cut at a million characters and flagged when cut.
• list_jurisdictions(): every configured jurisdiction's code, name and languages.
• get_jurisdiction(code): a jurisdiction's names, tradition, calendar, languages and document classes.
• compare_versions(reference_version_id, domestic_version_id): the compare report, each provision of the reference assessed against the domestic version. Needs the chat and embeddings endpoints and costs a model call per reference provision that carries text; refused above 200 of them.

A failure reads back as the tool's error with a plain message. --http binds to localhost unless --host says otherwise; there is no authentication here either.

Configuration

The CLI loads environment variables from a .env file in the working directory or parent directories; explicitly exported shell variables take precedence. Note that pytest and alembic do not read .env files automatically and require variables to be exported in your environment. See .env.example for all available options.

Jurisdictions

Configurations are included in distributed package wheels if their config.json sets either synthetic or public_reference to true. This filtering is enforced by codify/open_wheel.py at build time to prevent the bundled set from drifting out of sync.

• Synthetic jurisdictions: contain mock legislation formatted to real-world legislative structures, allowing test suites to assert against known-good parses without distributing copyrighted corpora.
• Public reference configurations: cover jurisdictions that publish their legal texts openly.

A source checkout reads the repository's data/ directory, wherever the command is run from; an installed wheel reads the data bundled inside the package. To supply a custom dataset, set CODIFY_DATA_ROOT to an absolute path containing jurisdictions/ and frameworks/ directories before starting Python. Relative paths are rejected, and custom data roots completely replace bundled data.

To add a jurisdiction, see docs/jurisdictions/adding-a-jurisdiction.md.

Layout

• codify/akn/: AKN 3.0 element models, parsing, rendering, eId generation, reference resolution, and schema validation
• codify/pipeline/: End-to-end ingestion from raw bytes to AKN. Contains input format parsers (formats/) and structural enrichment passes (enrich/).
• codify/acquisition/: Source adapters, scrape manifests, and rate limiting for jurisdiction data sources.
• codify/embed/: Provider-agnostic text embedding client over OpenAI-compatible endpoints.
• codify/retrieve/: Hybrid retrieval over statutory provisions using dense embeddings and BM25 fused via Reciprocal Rank Fusion (RRF).
• codify/compare/: Statutory compliance comparator, including alignment logic, prompt templates, and schema-validated model outputs.
• codify/storage/: Typed PostgreSQL data access layer.
• codify/lenses/: Analysis plugins and the extensible lens registry.
• codify/repair/: Automated AKN repair agent performing finding-grounded, transactional XML edits.
• codify/translate/: Structure-preserving legal translation with batching, clause parity checks, and quality scoring.
• codify/core/: Shared LLM client, OpenTelemetry/Langfuse tracing, internationalization, and log redaction.
• codify/frbr.py: FRBR URI generation and parsing.
• codify/jurisdictions.py: Jurisdiction configuration loader and schema validator.

Tests

For a detailed breakdown of test scopes, see the test guide.

Unit tests do not require an API key or a database:

uv sync --group dev --extra migrations
uv run pytest tests -m "not integration and not live_llm" -q

This runs the unit suite in about a minute and matches the standard CI check. Tests for unbundled jurisdictions are skipped automatically.

Note: Integration tests require the same Postgres as Store, search, compare, with the pgvector and pg_textsearch extensions:

# Start test database and apply schema migrations
docker compose up -d --wait postgres
uv run alembic -c alembic.ini upgrade head

# Run integration suite
REQUIRE_DB=1 uv run pytest tests -m "integration and not live_llm" -q

• Database safety: Integration tests write and drop data. Never set POSTGRES_URL to a production or shared database. If port 5432 is already bound locally, set CODIFY_PG_PORT and update POSTGRES_URL accordingly.
• Live LLM tests: Tests marked live_llm issue requests to the configured OpenAI-compatible gateway and incur API charges.
• CI checks: In addition to unit tests, CI enforces formatting, type checking, and wheel builds via Ruff, strict mypy, and uv build with the Hatchling backend (see .github/workflows/ci.yml).

Standards

Codify targets the Akoma Ntoso 3.0 specification. Generated documents pass a structural validator whose findings ride the bundle and the run; the OASIS schema is checked on one acquisition route only, not on every output. FRBR URIs identify works, expressions and manifestations, and eIds address the individual provision within them.

Structural validity alone does not guarantee semantic fidelity to the source text or seamless compatibility with external tooling. Internal structural conventions, such as how annex content is inlined, can diverge from specific downstream profiles like AKN4EU or platforms such as Indigo. You should validate intended interchange workflows using representative documents directly within the consuming system. For details on compatibility boundaries and known profile differences, see the interoperability scope documentation.

More

• docs/architecture.md
• docs/ocr-cascade.md
• docs/decisions/: architecture decision records
• CONTRIBUTING.md
• CODE_OF_CONDUCT.md
• SECURITY.md
• AGENTS.md: conventions for coding agents working in this repository

Codify is built by CentreAI at the Tony Blair Institute for Global Change. For more information on the vision and access to a hosted version, visit codify.centreai.global.