Prithv122/production-rag-index

production-rag — prebuilt heading index

Retrieval artefacts for the production-rag Space. This repository holds no original content: it is a chunked, indexed derivative of three public documentation repositories, pinned at exact commits.

It exists so the Space searches the same index the published evaluation numbers came from instead of a rebuild, and so a free CPU Space does not spend ten minutes re-encoding 24k chunks on every restart.

Contents

heading/chunks.jsonl      chunk text, offsets, headings and source URLs
heading/bm25/             sparse weight matrix + vocabulary (scipy)
heading/dense/            float32 vectors, BAAI/bge-small-en-v1.5, L2-normalised

heading is one of three chunking strategies the project measures; it won on nDCG@10. The chunker splits on markdown headings, keeps code fences whole, and records the character span each chunk occupies in its source document.

Sources and licences

Tool	Repository	Commit	Licence
DuckDB	duckdb/duckdb-web (docs/current only)	6f6cd1659f0e2ddd1965b1d3f1833e7fc512e7ac	MIT
dbt	dbt-labs/docs.getdbt.com (website/docs)	cd0e5b0e2b77302807f7dd126faac871e728edab	Apache-2.0
Dagster	dagster-io/dagster (docs/docs)	76eed340c6b84517d91a86163c461c022ebef8d8	Apache-2.0

Each chunk carries its tool, its source doc_id and a link back to the canonical published page. DuckDB's superseded version trees (docs/0.10 … docs/lts) are deliberately not included — they are near-duplicates of current and measurably wreck retrieval precision.

Rebuilding this yourself

git clone https://github.com/Prithv122/production-rag
cd production-rag
uv run production-rag ingest                        # clones at the pinned SHAs
uv run --extra embed production-rag index --strategy heading

The result is byte-identical given the same commits and model.