Retrieval systems
When a retrieval system gives a bad answer, the model is usually innocent. It answered faithfully from the wrong three paragraphs.
Everything downstream is decided by what gets retrieved, which means the unglamorous parts carry the quality. How documents are split matters enormously: chunks that cut through the middle of a table, or that separate a heading from the rule it introduces, produce confident answers to questions the source never addressed. Metadata matters as much — a policy that was superseded in March will be retrieved forever unless the system knows about dates and versions.
Pure vector search is rarely the right answer on its own. It is good at meaning and bad at exactness, which is a problem when people search for a part number, an error code, a person's name or an acronym — precisely the queries where being wrong is most obvious. Hybrid retrieval, combining keyword matching with semantic similarity and then reranking the shortlist, is what closes that gap, and the reranking step is usually the single highest-return improvement available.
Citations are not a nicety; they are the mechanism by which the system becomes checkable. An answer that shows its sources can be verified by the person reading it, which converts an unfalsifiable claim into something with an audit trail. It also changes behaviour on your side: once every answer must point at a passage, the failures you were tolerating in retrieval become visible immediately.
How we work
- Chunking is designed around the documents you actually have, and checked by reading the chunks rather than trusting the splitter.
- Hybrid retrieval with reranking, because pure vector search fails exactly on part numbers and names.
- Every answer cites its passages, so a wrong answer is diagnosable instead of merely disappointing.
- The index refreshes incrementally, so a document edited this morning is not answered from last month's copy.
What this includes
Pick what you need and send it over.
Questions
- How much material do we need?
- Less than people expect — a few hundred well-structured documents is plenty to be useful. Volume is rarely the constraint; contradictions are. If three documents state different versions of the same policy, retrieval will faithfully surface all three, and no amount of tuning resolves a disagreement your organisation has not resolved.