RAG Implementation Checklist — AI Builders Network
===================================================

Chunking:
[ ] 300-800 tokens per chunk with overlap
[ ] Metadata: source, title, chunk_index
[ ] Code blocks kept intact

Embeddings:
[ ] Same model for index and query
[ ] Vectors stored with metadata filters

Retrieval:
[ ] Top-k tuned (start with 5)
[ ] Similarity threshold set
[ ] Source citations in responses

Eval:
[ ] 20 test questions with expected sources
[ ] Retrieval accuracy scored before prompt tuning

Production:
[ ] Access control on document sets
[ ] Fallback when no relevant chunks found
[ ] Cost per query measured