abhishektuteja/rag-sec-db

RAG-SEC preloaded database

A PostgreSQL dump of the chunks table for github.com/abhishektuteja01/RAG-SEC: retrieval-augmented QA over SEC 10-K filings. Restoring it skips building the corpus and embedding it (hours of GPU time).

• 99,654 chunks from 799 SEC 10-K filings, each ~900 tokens (bge-m3 tokenizer; variant A, the serving corpus)
• columns: id, filing_stem, chunk_index, heading, n_tokens, text, embedding vector(1024), variant, excluded_by_variant
• one 1024-d dense embedding per chunk from BAAI/bge-m3, under an HNSW cosine index
• a ParadeDB pg_search BM25 index on (id, text)
• rag_sec_chunks.dump: pg_dump -Fc -t chunks --no-owner --no-privileges (custom format, ~652 MB)
• rag_sec_chunks.dump.sha256: checksum

Index contents are not stored in the dump. Both indexes are rebuilt during restore, which took about 2 minutes on a laptop.

Requirements

• PostgreSQL 18
• pgvector 0.8.6 (vector extension)
• ParadeDB pg_search 0.25.1 (pg_search extension, loaded via shared_preload_libraries=pg_search)

The repo's deploy/Dockerfile.postgres builds this image (pgvector/pgvector:0.8.6-pg18-trixie plus pg_search). Give the container shm_size: 2gb: with a larger maintenance_work_mem, a parallel HNSW build needs more than Docker's default 64 MB /dev/shm.

Restore

Easiest: in the repo, run scripts/setup_db.sh (downloads, verifies and restores). By hand:

hf download abhishektuteja/rag-sec-db --repo-type dataset --local-dir data/db
(cd data/db && shasum -a 256 -c rag_sec_chunks.dump.sha256)

# into a running Postgres container (service "postgres", user/db from .env)
docker compose exec -T postgres psql -U "$POSTGRES_USER" -d "$POSTGRES_DB" -v ON_ERROR_STOP=1 \
  -c "CREATE EXTENSION IF NOT EXISTS vector; CREATE EXTENSION IF NOT EXISTS pg_search;"
docker compose cp data/db/rag_sec_chunks.dump postgres:/tmp/rag_sec_chunks.dump
docker compose exec -T postgres pg_restore -U "$POSTGRES_USER" -d "$POSTGRES_DB" \
  --no-owner --no-privileges --clean --if-exists --exit-on-error -j 4 /tmp/rag_sec_chunks.dump
docker compose exec -T postgres rm /tmp/rag_sec_chunks.dump
docker compose exec -T postgres psql -U "$POSTGRES_USER" -d "$POSTGRES_DB" -c "ANALYZE chunks;"

Check: SELECT count(*) FROM chunks; returns 99654, and both chunks_embedding_hnsw and chunks_bm25_idx show up in pg_indexes.

Source and license

• Filing text: SEC EDGAR 10-K filings, U.S. government public records.
• Which filings are included comes from T²-RAGBench (G4KMU/t2-ragbench, CC-BY-4.0, arXiv:2506.12071): every unique (CIK, report year) pair in its FinQA and ConvFinQA subsets. Credit to the T²-RAGBench authors.
• This dataset (chunking, embeddings, dump) is released under CC-BY-4.0.

No warranty. The chunk text is parsed filing text and may contain parsing errors. Do not use it as financial advice.