| --- |
| license: mit |
| language: |
| - ja |
| tags: |
| - legal |
| - japanese-law |
| - retrieval |
| - embeddings |
| --- |
| |
| # JuriCode-JP retrieval index and citation registry |
|
|
| Structured Japanese statutory law for retrieval, with a citation registry that lets a |
| consumer verify that a quoted passage is byte-identical to the source text. |
|
|
| ## Contents |
|
|
| | Path | What it is | |
| |---|---| |
| | `index/v0.2-aug-v9-gemini.npy` | Dense vectors, `float32`, shape `(143749, 3072)` | |
| | `index/v0.2-aug-v9-gemini.meta.jsonl` | One row per vector, in the same order: `chunk_id`, `article_id`, `law_id`, and section metadata | |
| | `registry/registry-2026-07/documents.jsonl` | One row per document, with its content hash | |
| | `registry/registry-2026-07/chunks.jsonl` | `chunk_id` -> document id | |
| | `registry/registry-2026-07/_manifest.json` | Provenance: `pipeline_version`, paired index, sha256 of inputs and outputs | |
|
|
| Embedding model: `gemini-embedding-001` (3072 dimensions, `RETRIEVAL_DOCUMENT` task type). |
| Registry `pipeline_version`: `w6-registry/2026-07`. |
|
|
| The `.npy` row at position *i* corresponds to the `.meta.jsonl` row at position *i*. |
|
|
| ## What `hash_basis` means (read this before trusting a hash) |
| |
| Every document in the registry carries a `hash_basis` field. **The three values do not |
| carry the same strength of evidence.** Do not treat them as interchangeable. |
|
|
| | `hash_basis` | Meaning | |
| |---|---| |
| | `egov-xml-canonical` | The text was checked against the e-Gov source XML. The hash is the one recorded in the per-law source manifest. | |
| | `egov-xml-derived` | Supplementary provisions (附則). Derived from the same source XML, but there is no per-article manifest anchor for them yet. | |
| | `juricode-stored-text` | The hash is over the text as stored here. **It is self-reported.** No external source was re-fetched to confirm it. | |
|
|
| This distribution is **not** "all verified". It is verified where it says it is verified. |
|
|
| ## No pickle |
|
|
| Only `.npy` and `.jsonl` are distributed. The `.vec.pkl` file used internally is **not** |
| published: it holds the same vectors as the `.npy`, and unpickling executes arbitrary |
| code in the reader's process. A project whose value proposition is verifiability should |
| not ship a format that asks the consumer to trust it. |
|
|
| ## Licensing |
|
|
| The structuring layer (schema, ids, chunking, this index) is MIT. |
| The underlying texts are not ours to relicense: statutes come from e-Gov, circulars and |
| tax guidance from the National Tax Agency, and rulings from the National Tax Tribunal; |
| each is used under its own terms. See the source repository for the per-source table. |
|
|
| ## Comparing registries |
|
|
| Do not diff two registries with different `pipeline_version` values. A change in the |
| pipeline changes how text is chunked and hashed, so a row-by-row diff across versions |
| measures the pipeline, not the law. |
|
|
| ## Integrity |
|
|
| The sha256 of every file here is recorded in `CHANGES.md` in the source repository. That |
| file is the anchor: it lets a third party confirm that what they downloaded is what was |
| published, and that it was not replaced afterwards. |
|
|
| Source: https://github.com/JuriCode-JP/JuriCode-JP |
|
|
