4 min read
Why eval sets go stale
An eval set built in March measures something else by June. The three things that drift — corpus, query distribution, judge — and how to catch each one.
#rag
10 posts tagged rag.
7 min read
RAG evaluation: the harness I actually run
Most RAG evaluation advice stops at build a labelled set and use LLM-as-judge. This is the workflow: sampling real queries, and calibrating the judge.
5 min read
Re-ranking: the highest-leverage 50 lines
Reranking top-50 down to top-5 routinely beats raw retrieval by 10+ recall points. What a cross-encoder does differently, and the 50 lines that call one.
7 min read
Swap an embedding model without breaking it
Re-embedding the corpus is the easy part. What actually breaks when you change embedding models in production, and the evals I run on either side of the swap.
6 min read
Multi-tenant RAG: relevance is not access control
Vector search ranks by similarity, not permission. Why multi-tenant RAG leaks across tenants by default, and where the authorization check actually has to sit.
5 min read
The Postgres-only AI stack
pgvector for embeddings, LISTEN/NOTIFY for jobs, row-level security for tenant isolation. One database instead of four services — and the signals to split.
5 min read
Why hybrid search beats vector search
Dense embeddings and BM25 fail on different queries, and real corpora have both. Reciprocal rank fusion in pgvector, and when pure vector search still wins.
5 min read
Chunking strategies for RAG, and when each one bites you
Fixed-size, semantic, and structure-aware chunking all look fine in a demo. Here's how each one actually fails once real documents hit your production pipeline.
5 min read
Choosing a vector store in 2026
A practical comparison of pgvector, Pinecone, Qdrant, and Weaviate — real cost and latency numbers from production, and why pgvector is usually enough to start.
7 min read
A production RAG checklist (the boring half)
What separates a working RAG demo from a production RAG system isn't the retrieval — it's the evaluation, observability, and failure-mode handling around it.