Document Chatbot with Live Lineage
A high-precision multilingual chatbot over a living document set — every answer traceable to its source passage, with the index updated in real time as documents change.
- Role
- Team lead & delivery
- Retrieval
- Grounded, with lineage
- Languages
- Multilingual
The problem
Most document chatbots quietly assume two things that were not true here: that the documents sit still, and that a plausible answer is good enough. This one had to work against a corpus being edited while people were querying it, and in a setting where a confidently wrong answer was worse than no answer at all. Multilingual made both harder — the same question asked in two languages has to reach the same source material and come back with the same answer, not two differently-shaped guesses.
The approach
I led the team and owned the delivery. The design pushed correctness into the structure rather than relying on prompt discipline:
- Lineage on every answer, surfaced in real time — a response carries the passages it was derived from, so a reader can check the claim instead of trusting it. This is the difference between an answer that is useful and one that is defensible.
- Real-time upsert, so creating or editing a document updates what retrieval sees immediately. No overnight reindex, no window where the assistant confidently cites a version that no longer exists.
- Precision over recall, tuned so the system returns less rather than returning something shaky, and declines outright when the source material does not cover the question.
- One corpus, many languages, so multilingual support is a retrieval and generation concern rather than a separate stack per language drifting out of sync.
The outcome
Users get answers they can verify at the point of reading, against documents that are current as of the moment they asked — and the failure mode, when it comes, is an honest “not covered” rather than a fluent invention.