Retrieval-augmented generation works when the retrieval half is treated as seriously as the generation half. In our deployments, most hallucination problems trace back to retrieval — not the model. Chunking strategy, re-ranking, and citation discipline matter more than which model you pick.
We start by mapping the actual query patterns a system will see, not the content structure we happen to have. That shapes chunk size and overlap. From there, a re-ranking pass over the top candidates catches the cases where pure vector similarity misleads — near-duplicate content, outdated versions, or superficially similar but substantively different passages.
Finally, every generated answer traces back to the passages that produced it. If a claim can't be traced to a retrieved passage, the pipeline treats it as unsupported and asks a clarifying question instead of guessing.