Blog

• reranking that costs more than it saves

  a rerank pass can lift the right chunk in a demo and still add more latency and spend than the answers justify.

• hybrid search without turning it into a science fair

  keep keyword and vector search, merge them simply, and skip the fusion roadmap until the boring combination actually fails.

• metadata you skipped that would have saved retrieval

  skipping ingest metadata is how versioned docs, drafts, and marketing pages keep winning similarity contests they should never enter.

• when you should not build a knowledge base yet

  a knowledge base will not invent process, ownership, or truth. if those are missing, embeddings just make the mess searchable.

• chunking mistakes that look like model problems

  most "the model is dumb" tickets in rag systems are chunking bugs wearing a lab coat.

• The cheapest part of RAG is the embeddings

  Why most RAG systems return crap from a working retrieval pipeline, and the corpus-curation work that actually fixes it.

• When your retrieval is slow, the answer is rarely the embeddings

  A guest post from the SaaSPerform team on the five places production RAG latency actually lives, and why the embeddings are usually fine.

• An eval harness that survives past month two

  How to build the evaluation pipeline a RAG team will actually keep using, in five pieces, with the loop that prevents abandonment.