🤖 AI Engineering · RAG Pipelines

Implement hybrid keyword + vector search

Combine BM25-style lexical search with dense embeddings and fuse results so both rare terms and paraphrases are found.

intermediate~35 minAI EngineersML EngineersLLM App Developers

Steps

1Stand up a lexical index (BM25) alongside your vector index on the same chunks
2Embed queries with the same model and distance metric used at ingestion
3Retrieve top-k from both indexes with k at least double your final context size
4Fuse rankings with Reciprocal Rank Fusion (RRF), k=60 as default
5Optionally rerank fused candidates with a cross-encoder on top-50
6A/B against pure-vector on your golden questions; keep the winner per query class

Common Pitfalls

▲Mixing similarity metrics between ingestion and query time
▲Fusing raw scores instead of ranks across incompatible scales

Success Signals

✓Hit-rate lift over best single retriever
✓P95 retrieval latency under 300ms

Commands

Install with skills CLI

$ npx skills add aniruddhaadak80/skills --skill rag-pipelines-hybrid-search

Install globally

$ npx skills add aniruddhaadak80/skills --skill rag-pipelines-hybrid-search -g

Tags

#rag#search#bm25#ai-engineering#rag-pipelines

Install this skill

$ npx skills add aniruddhaadak80/skills --skill rag-pipelines-hybrid-search -a opencode

SKILL.md

# Implement hybrid keyword + vector search

> Combine BM25-style lexical search with dense embeddings and fuse results so both rare terms and paraphrases are found.

## When to Use This Skill

Combine BM25-style lexical search with dense embeddings and fuse results so both rare terms and paraphrases are found.

## Steps

1. Stand up a lexical index (BM25) alongside your vector index on the same chunks
2. Embed queries with the same model and distance metric used at ingestion
3. Retrieve top-k from both indexes with k at least double your final context size
4. Fuse rankings with Reciprocal Rank Fusion (RRF), k=60 as default
5. Optionally rerank fused candidates with a cross-encoder on top-50
6. A/B against pure-vector on your golden questions; keep the winner per query class

## Common Pitfalls

- Mixing similarity metrics between ingestion and query time
- Fusing raw scores instead of ranks across incompatible scales

View source on GitHub ↗

Choose chunk sizes, overlaps, and structure-aware splits so retrieved context actually helps the model answer.

🤖 AI Engineering·~25m

Open →

Evaluate RAG answer quality automaticallyadvanced

Score groundedness, relevance, and completeness with judge prompts plus deterministic checks wired into CI.

🤖 AI Engineering·~45m