RAG & Semantic Search
Grounded AI over your knowledge base — retrieval, reranking, evals.
Production RAG that retrieves the right chunks, grounds answers with citations, and stays accurate as your corpus grows.
Problems I solve
- LLM answers not grounded in your docs
- Vector search returns irrelevant chunks
- No way to measure retrieval quality
- Corpus updates break answer quality
What you get
- Chunking strategy tuned to your content
- Hybrid search (BM25 + vector) with reranking
- Citation-required prompting
- Retrieval + answer evals on a golden set
Use cases
Docs-grounded support chat
Answers with citations from your help center, resolved tickets, and PDFs.
Internal knowledge assistant
Chat over Notion + Google Drive + Confluence with permissioned retrieval.
Legal / policy search
Clause-level retrieval with jurisdiction filters and reranking.
Sales enablement search
Retrieve the right case study, deck, or objection response instantly.
Examples I've shipped
Support RAG on pgvector
Outcome — Recall@5 above 92% across 20k chunks.
Permissioned Notion RAG
Outcome — Enterprise-safe search across 12k pages with per-user ACLs.
How I work
- 1
Corpus audit
Understand content shapes, update frequency, and access rules.
- 2
Chunk + index
Choose chunking + embedding model, build hybrid index.
- 3
Rerank + prompt
Add cross-encoder reranking and citation-required prompts.
- 4
Eval
Golden Q&A set, measure recall + answer faithfulness.
- 5
Ship + monitor
Deploy behind observability, add drift alerts.
Deliverables
- Production RAG service with API
- Retrieval + answer eval suite
- Ingestion pipeline for corpus updates
- Dashboards for recall + cost
Benefits
- Grounded, cited answers
- Search quality you can measure
- Safe to expose to real users
Frequently asked questions
Do I need a vector database?+
Usually pgvector on Postgres is enough up to millions of chunks. I only reach for dedicated vector DBs when latency or scale demands it.