Building production RAG systems that don't hallucinate
Retrieval quality, evaluation harnesses, and human review loops — the difference between a demo and a system you can trust.
Why prototypes fail after launch
Most RAG prototypes look convincing in a slide deck and fall apart the week after launch. The failure mode is rarely the model — it is retrieval, evaluation, and the absence of a human-in-the-loop for regulated decisions.
Corpus audit and chunking strategy
We start with a corpus audit: what documents are authoritative, how they change, and which chunks must never be mixed. Chunking strategy follows the domain, not a default token window. Metadata (source, effective date, access tier) travels with every retrieval hit.
Evaluation before production
Evaluation is not optional. Before production we ship a golden set of questions with expected citations, plus adversarial prompts that try to pull PII or invent policy. Regression runs on every index rebuild.
Confidence gates in production
In production, answers without sufficient retrieval confidence either refuse or escalate. That single constraint — no answer without evidence — is what keeps hallucination rates acceptable for enterprise buyers.
If you are wiring RAG into underwriting, support, or internal knowledge, treat it as a system with SLOs, not a chatbot feature. The model is the easy part.
