README.md

---
license: etalab-2.0
language:
  - fr
tags:
  - legal
  - french-law
  - embeddings
  - legifrance
  - mistral
size_categories:
  - 10K<n<100K
---

# Open Codes

Open dataset of French legal code articles with embeddings.

Chunked articles from French legal codes sourced from
[Legifrance](https://www.legifrance.gouv.fr/) via the PISTE API,
with 1024-dimensional embeddings generated by Mistral AI.

Each row is a text chunk enriched with full article metadata from the parent
legal code article.

## Dataset Description

- **Source**: [PISTE Legifrance API](https://piste.gouv.fr/) (official French government legal database)
- **License**: [Licence Ouverte / Etalab 2.0](https://www.etalab.gouv.fr/licence-ouverte-open-licence/)
- **Embeddings**: Mistral AI `mistral-embed` (1024 dimensions)
- **Update frequency**: Daily (nightly sync at 02:00 UTC, dataset push at 06:00 UTC)
- **Codes**: Dynamically sourced from `LEX_codes_piste` (active codes only)
- **Quality**: Dedup + stale-chunk filtering applied before every push

### Legal codes included

The list of legal codes is dynamic and managed in the `LEX_codes_piste` table.
Adding a new code there (with `actif=true`) automatically includes it in the next sync and export.

## Schema

### Chunk fields
| Column | Type | Description |
|--------|------|-------------|
| `chunk_text` | string | Text content of the chunk |
| `embedding` | float32[1024] | Mistral AI embedding vector |
| `id_legifrance` | string | Legifrance article identifier |
| `code_name` | string | Human-readable code name (e.g. "Code civil") |
| `chunk_index` | int32 | Chunk position within the article (0-indexed) |
| `start_position` | int32 | Character offset in original article text |
| `end_position` | int32 | End character offset in original article text |
| `code` | string | Legal code identifier (e.g. LEGITEXT000006070721) |
| `num` | string | Article number (e.g. "L. 1234-5") |
| `etat` | string | Article status (VIGUEUR, ABROGE, etc.) |
| `fullSectionsTitre` | string | Full hierarchy path in the code |

### Article metadata fields (prefixed `article_`)

#### Identifiers
| Column | Type | Description |
|--------|------|-------------|
| `article_id_legifrance` | string | Legifrance article ID |
| `article_code` | string | Legal code ID |
| `article_num` | string | Article number |
| `article_cid` | string | Consolidated ID |
| `article_idEli` | string | ELI (European Legislation Identifier) |
| `article_idEliAlias` | string | ELI alias |
| `article_idTexte` | string | Text ID |
| `article_cidTexte` | string | Consolidated text ID |

#### Content
| Column | Type | Description |
|--------|------|-------------|
| `article_texte` | string | Full article plain text |
| `article_texteHtml` | string | Full article HTML |
| `article_nota` | string | Article notes (plain text) |
| `article_notaHtml` | string | Article notes (HTML) |
| `article_surtitre` | string | Article subtitle |
| `article_historique` | string | Article history |

#### Dates & Status

**Date format**: `article_dateDebut` and `article_dateFin` are stored as **Unix timestamps in milliseconds** (string type). This is the raw format returned by the Legifrance PISTE API.

To convert to a human-readable date in Python:

```python
from datetime import datetime, timezone

timestamp_ms = "1301529600000"
dt = datetime.fromtimestamp(int(timestamp_ms) / 1000, tz=timezone.utc)
print(dt)  # 2011-03-31 00:00:00+00:00
```

Special value: `32472144000000` (year 2999) means "no end date" — the article is in force indefinitely.

| Column | Type | Description |
|--------|------|-------------|
| `article_dateDebut` | string | Effective start date (Unix ms) |
| `article_dateFin` | string | Effective end date (Unix ms, `32472144000000` = indefinite) |
| `article_dateDebutExtension` | string | Extension start date (Unix ms) |
| `article_dateFinExtension` | string | Extension end date (Unix ms) |
| `article_etat` | string | Status: `VIGUEUR` (in force), `ABROGE` (repealed), etc. |
| `article_type_article` | string | Article type |
| `article_nature` | string | Legal nature |
| `article_origine` | string | Origin (e.g. `LEGI`) |
| `article_version_article` | string | Version identifier |
| `article_versionPrecedente` | string | Previous version ID |
| `article_multipleVersions` | bool | Has multiple versions |

#### Hierarchy
| Column | Type | Description |
|--------|------|-------------|
| `article_sectionParentId` | string | Parent section ID |
| `article_sectionParentCid` | string | Parent section consolidated ID |
| `article_sectionParentTitre` | string | Parent section title |
| `article_fullSectionsTitre` | string | Full hierarchy path |
| `article_ordre` | int32 | Sort order within the code |
| `article_partie` | string | Partie (e.g. "Partie legislative") |
| `article_livre` | string | Livre |
| `article_titre` | string | Titre |
| `article_chapitre` | string | Chapitre |
| `article_section` | string | Section |
| `article_sous_section` | string | Sous-section |
| `article_paragraphe` | string | Paragraphe |

#### Extras
| Column | Type | Description |
|--------|------|-------------|
| `article_infosComplementaires` | string | Additional info (plain text) |
| `article_infosComplementairesHtml` | string | Additional info (HTML) |
| `article_conditionDiffere` | string | Deferred condition |
| `article_infosRestructurationBranche` | string | Branch restructuring info |
| `article_infosRestructurationBrancheHtml` | string | Branch restructuring (HTML) |
| `article_renvoi` | string | Cross-references |
| `article_comporteLiensSP` | bool | Contains SP links |
| `article_idTechInjection` | string | Technical injection ID |
| `article_refInjection` | string | Injection reference |
| `article_numeroBo` | string | BO number |
| `article_inap` | string | INAP code |

## Usage

```python
from datasets import load_dataset
from datetime import datetime, timezone

ds = load_dataset("ArthurSrz/open_codes", split="train")

# Access a chunk with its embedding and article metadata
row = ds[0]
print(row["code_name"])                # e.g. "Code civil"
print(row["chunk_text"][:200])
print(len(row["embedding"]))           # 1024

# Convert dates from Unix ms to datetime
date_debut = datetime.fromtimestamp(int(row["article_dateDebut"]) / 1000, tz=timezone.utc)
print(date_debut)                      # e.g. 2011-03-31 00:00:00+00:00

# Filter by legal code
code_civil = ds.filter(lambda x: x["code_name"] == "Code civil")

# Filter active articles only
en_vigueur = ds.filter(lambda x: x["article_etat"] == "VIGUEUR")

# Use embeddings for semantic search
import numpy as np
query_emb = np.array(ds[0]["embedding"])
```

## Provenance

Built by the [marIAnne](https://github.com/ArthurSrz/open_codes) project.
Sync pipeline fetches articles nightly from PISTE Legifrance, chunks them,
and generates embeddings via Mistral AI.