Case study
When Semantic Search Misses Proper Nouns
Live semantic Q&A over portfolio content, Bedrock embeddings, Pinecone retrieval, Nova Pro generation, visitor uploads with citations.
AI WorkflowsPlatform Performance / Reliability
The challenge
The live portfolio RAG demo returned "no context" for Narrow Gate even though the phrase existed in an indexed upload. Dense-only Pinecone retrieval ranked portfolio architecture chunks above the story chunk, a classic hybrid-search gap for rare proper nouns.
Architecture approach
- ·Reproduced the failure in production: vague entity queries vs rich contextual queries
- ·Shipped interim fixes, reserved upload retrieval slots, corpus-busy query lock, session-scoped history
- ·Planned Pinecone dense + sparse hybrid index (BM25 / pinecone-sparse-english-v0 + Titan v2)
- ·Defined eval queries (Narrow Gate, Mara Voss clues, portfolio-only regression checks)
- ·Structured logging (query.retrieval) for uploadHits and rewrite diagnostics
Tech stack
BedrockPineconeLambdaDynamoDBNext.js
Results
- Narrow Gate query passes with upload-slot retrieval; full hybrid search documented as next step
- Case study published alongside Live Portfolio RAG Demo on /work
- Roadmap captured in portfolio-rag-api ROADMAP.md, not implemented yet by design