This Site's AI Twin — RAG over my own content

A portfolio chatbot that answers from a general model will confidently invent a project I never built or a stack I've never used. For a site whose whole job is to represent my work accurately, a plausible hallucination is worse than no assistant at all.

Built a retrieval layer over the site's own content. An ingest script chunks every project, essay, service and FAQ entry, embeds each chunk with text-embedding-3-small, and stores the 1536-dimension vectors in Postgres via pgvector. At query time the question is embedded and matched with an HNSW cosine index through a `match_content_chunks` RPC that enforces a similarity floor, and only the retrieved chunks are put in front of the model.

• 01Answers are grounded in retrieved chunks, each carrying the page it came from, so the assistant links back to real content instead of describing work that doesn't exist
• 02Re-ingest is content-hash addressed: unchanged chunks are skipped and stale ones pruned, so editing one project re-embeds one project
• 03HNSW over IVFFlat for the index — better recall at this corpus size, and no training step to schedule
• 04Retrieval runs as a Postgres function with a similarity floor, so an off-topic question returns nothing rather than the least-bad match