README.md

metadata

pretty_name: SAR retrieval indexes and queries
tags:
  - retrieval
  - faiss
  - rag
  - qwen3-embedding

SAR retrieval indexes and queries

Prebuilt Qwen3-Embedding-8B FAISS indexes and query files for four search-agent benchmarks, laid out exactly as the sar (search agent research) code expects under scratchspace/data/. Every index was built without a query/document instruction prefix; embeddings are last-token pooled and L2-normalized, and the metric is inner product.

Download

From the root of the sar repo:

hf download ozelMetalUp/sar-data --repo-type dataset --local-dir scratchspace/data
# or just one dataset:
hf download ozelMetalUp/sar-data --repo-type dataset --local-dir scratchspace/data --include "openresearcher/*"

Layout

browsecompplus/indexes/qwen3-8b-embedding/   flat (IndexFlatIP), 100,195 docs          ~5 GB
flashrag/indexes/qwen3-8b-embedding/         IVF32768,SQ8 (nprobe 64), 21,015,324 passages ~101 GB
flashrag/{nq,hotpotqa,2wikimultihopqa,musique,bamboogle,popqa}/{train,dev,test}.jsonl
openresearcher/indexes/qwen3-8b-embedding/   IVF32768,SQ8 (nprobe 64), 14,878,084 docs ~110 GB
openresearcher/queries/{train,dev,sft_train_queries}.jsonl
trqa/indexes/ecommerce/qwen3-8b-embedding/   flat (SQfp16), 3,282,927 docs              ~29 GB
trqa/indexes/wiki/qwen3-8b-embedding/        flat (SQfp16), 5,910,177 docs              ~52 GB
trqa/indexes/{ecommerce,wiki}/bm25/          Pyserini/Lucene BM25 over the same docs
trqa/queries/*.encrypted.jsonl               encrypted, see Sources
trqa/qrels/*.encrypted.jsonl                 encrypted, see Sources

Each index directory holds:

• index.faiss: the FAISS index; faiss id i is row i of docids.jsonl.
• docids.jsonl: one {"docid", "text"} row per indexed document (this is the corpus text, so no separate corpus download is needed to serve an index).
• config.json: the model and index settings the searcher reads (index_type, nprobe/efSearch, ...). corpus_path is informational only.

Query files are JSONL rows with query_id, question, and golden_answers.

Sources

• FlashRAG: QA splits from RUC-NLPIR/FlashRAG_datasets; the index covers the FlashRAG Wikipedia (2018, 100-word passage) retrieval corpus.

• OpenResearcher: queries from OpenResearcher/OpenResearcher-Dataset (seed_42). The index is built from the authors' released Qwen3-Embedding-8B document encodings (OpenResearcher/OpenResearcher-Indexes, qwen3-embedding-8b/corpus.pkl), requantized to IVF+SQ8.

• BrowseComp-Plus: index over the Tevatron/browsecomp-plus-corpus corpus. Queries and qrels are intentionally not included: the authors distribute them encrypted to keep them out of training crawls. Recreate them with the sar scripts:

  BCP=scratchspace/data/browsecompplus
  uv run scripts/download_browsecompplus_dataset.py --output-dir $BCP/queries \
      --hf-cache-dir scratchspace/hf_cache_$USER
  uv run scripts/decrypt_browsecompplus_dataset.py --input $BCP/queries/test.jsonl \
      --output $BCP/queries/test.decrypted.jsonl
  mv $BCP/queries/test.decrypted.jsonl $BCP/queries/test.jsonl
  uv run scripts/split_browsecompplus_qrels.py --queries-file $BCP/queries/test.jsonl \
      --qrels-file $BCP/qrels/test.tsv
  

• TRQA: indexes over the TRQA rewritten e-commerce corpus and the rewritten Wikipedia (2025-10-01) partial corpus. Queries and qrels are encrypted (every field except query_id and canary is XOR'd with SHA256 of the TRQA canary string, as BrowseComp-Plus does), since the authors distribute them encrypted too. Decrypt them in place after downloading:

  for d in queries qrels; do
    uv run scripts/crypt_trqa_dataset.py decrypt \
        --input-dir scratchspace/data/trqa/$d --output-dir scratchspace/data/trqa/$d
  done
  

  Decrypted query rows carry query_id, question, golden_answers plus TRQA's original fields (query, answer, gold_docs, negative_docs, evidence_docs).

Please follow the licenses of the original datasets.

Serving

uv run python -m sar.searcher.faiss_searcher \
    --index-dir scratchspace/data/openresearcher/indexes/qwen3-8b-embedding \
    --model Qwen/Qwen3-Embedding-8B

The flat BrowseComp-Plus index is moved to a GPU (as are the flat TRQA indexes, if one fits). The IVF indexes are searched on CPU and need roughly index.faiss + 1.9 x docids.jsonl of host RAM (about 180 GB for OpenResearcher and 115 GB for FlashRAG); the GPU only runs the query encoder.