A folder of contracts is not something you can ask questions of. Getting from one to the other is data engineering, not a model choice.
Retrieval over documents demos beautifully and disappoints quietly. You point it at a shared drive, ask an easy question, get a good answer, and assume the rest will follow. Then somebody asks about a clause that only makes sense alongside an amendment two documents away, and the answer is confidently wrong.
Most of the work in a document system is not the model. It is knowing that a contract has parties, dates and versions, that an invoice belongs to an order, that the appendix overrides the summary. Chunking a PDF into paragraphs throws all of that away and hopes similarity search recovers it.
So we model documents the way the business talks about them, then tune retrieval to that shape. It is less exciting than swapping models, and it is the difference between a toy and something legal will let you use.
Every answer has to carry a source. Not a footnote for credibility, a link the reader can open, because eventually one of them will.
Each claim the system makes should point at the passage it came from. This does two things: it lets a human verify in seconds, and it makes invention visible instead of plausible. Where there is no source, there is no answer.
A system that cannot say it does not know is a system that guesses. We build refusal in deliberately: if confidence is low, or the retrieved passages do not actually address the question, it says so and routes to a person. Teams trust that behaviour far faster than they trust a system that is right most of the time.
The wins are rarely the ones people expect. Not chat with our knowledge base, but the specific repetitive lookups: what did we agree on renewal terms, which supplier covers this region, has this claim been made before. Narrow questions asked hundreds of times a month, each currently costing someone fifteen minutes and a guess.
