ID_REG_RAG_2611

A ready-to-load retrieval corpus of Indonesian laws and regulations: 1,716,628 chunks from 45,179 documents, each with its provenance, a dense embedding, a BM25 sparse vector, and document-level relations (what revokes, amends or is based on what). Built November 2026 by Ravel, the offline corpus compiler for Vera, and published here as an open, versioned backup.

Everything lives under bundle/. It is a sealed bundle: manifest.json lists every file with its SHA-256 and row count.

Query-side recipe (read before embedding queries). These vectors were pooled at a closing token placed after the tokenizer's own <|endoftext|>: the sequence ends ... <|endoftext|><|im_end|> (ids 151643, 151645). A plain tokenizer call ends at <|endoftext|> and lands in a different space (cosine ~0.64 against the same text). Embed queries with the same ending, i.e. append token id 151645 to the tokenised query before last-token pooling. The manifest's reference.npy canary is the check: with the recipe the query-templated reference.txt reproduces it at cosine 1.000. Vera's embed server supports this as APPEND_TOKEN_ID=151645. Chunks, BM25 vocabulary, sparse vectors, relations, status and quality files do not depend on it.

Contents

Path	What
bundle/chunks/part-*.parquet	The chunks: id, doc_id, body, regulation type/number/year, chapter/article, identifier, source_title, source_url, locator, chunker lineage.
bundle/vectors/part-*.parquet	chunk_id, embedding (1024-d, fp16, L2-normalised).
bundle/sparse/part-*.parquet	chunk_id, sparse (pgvector sparsevec text, BM25 weights, 1-based).
bundle/corpus.bm25.json	The BM25 vocabulary and IDF (20,000 terms, k1 1.5, b 0.75), fitted once over the whole corpus.
bundle/relations.parquet	741,609 citation edges (revokes, replaces, amends, based_on) read from the verb beside each citation.
bundle/doc_status.parquet	Per-document status derived from those edges (in force, revoked, amended).
bundle/doc_quality.parquet	Per-document text-layer verdict and an indexable flag.
bundle/failed.jsonl	The 25 documents that produced nothing, and why.
bundle/schema.sql, indexes.sql	PostgreSQL + pgvector schema and indexes generated from the manifest.
bundle/reference.npy, reference.txt	A canary: a known sentence and the vector the corpus embedder gives it.
bundle/manifest.json	The recipe: model and revision, pooling, instruction strings, chunker version, checksums.

Embedding recipe

Qwen/Qwen3-Embedding-0.6B at revision 97b0c614be4d77ee51c0cef4e5f07c00f9eb65b3, 1024 dimensions, last-token pooling, left padding, normalised, fp16 storage. Documents are embedded without an instruction, behind a header of regulation type, number, year, subject, chapter and article. Queries use the Qwen3 template:

Instruct: Given a legal or regulatory question, retrieve passages that answer it
Query: {your question}

Embed queries with the same model, template and sequence ending (see the recipe note at the top), or the vectors are not comparable. Check your setup against reference.npy: embed reference.txt with the query template and expect cosine ≥ 0.999.

Loading

The bundle is built to load into PostgreSQL with pgvector, all or nothing, with Ravel (ravel load bundle --dsn ...); the loader runs a preflight on checksums, row counts and provenance first. The parquet files also read directly with any parquet reader.

Known gaps

• 25 documents produced no chunks (failed.jsonl).
• Around 1,000 documents have an unusable or empty text layer; they are flagged in doc_quality.parquet (indexable = false) and would need OCR.
• No cluster assignment: clustering is a deployment concern, done by the serving engine.
• The relations are read from citation wording, not verified by a lawyer. Treat status as a strong hint, not legal advice.

Source and licence

The texts are Indonesian laws and regulations, public official documents, collected as markdown from the ID_REG_Markdown corpus and re-chunked here. Each chunk carries a source_title and a source_url. Note that source_url is the path of the markdown file inside the ID_REG_Markdown collection (an hf://datasets/Azzindani/ID_REG_Markdown/... path, not an official government link); that collection is not public, so the URL identifies the file but does not open for outside readers. Derived fields (chunking, embeddings, sparse vectors, relations, status, quality) are produced by Ravel. This card's license: other reflects that the exact licence is not yet fixed; verify each text against its official publication before redistributing it commercially.