The challenge
Most organisations already hold the answers their staff spend hours looking for — in contracts, policies, handbooks and past proposals. A general-purpose chatbot cannot read any of it, and an assistant that answers confidently from nothing is worse than no assistant at all.
Our approach
A concept study in making an assistant that is checkable. Every claim it makes carries the passage it came from, and when retrieval finds nothing relevant the assistant says so rather than filling the gap. Trust in this kind of system is built entirely from being able to verify it.
The constraints
The hard limit is that a wrong answer delivered confidently is worse than no answer. That constraint drove the whole design — permissions scoped per reader, citations on every claim, and an explicit no-answer state rather than a plausible guess.
What we built
- Document ingestion that preserves structure, so a clause keeps the section it belongs to
- Retrieval over the organisation’s own corpus, scoped by the reader’s permissions
- Answers with inline citations that open the source passage directly
- An explicit no-answer state when retrieval returns nothing relevant
- Streamed responses so long answers start reading immediately
E-commerce Platform
Commerce / Storefront