Sample study. This is illustrative content showing the shape of the work, not a real client outcome. It is excluded from search indexing.
AI knowledge base
Answers with citations, or no answer
A knowledge base the support team actually trusts
Retrieval tuned for recall rather than demo quality, so the assistant stops confidently not knowing things the documentation already says.
- Client
- A logistics software vendor
- Work
- AI knowledge base
- Sector
- Software
- Year
- 2026
Results
- Answers carrying citations
- Every one
- Behaviour with no match
- Refuses, escalates
- Retrieval tuned for
- Recall, not latency
The story
The challenge
An earlier assistant answered fluently and wrongly often enough that the support team stopped using it. Nobody could say which answers came from the documentation and which the model had invented.
The approach
Recall measured before anything shipped. Default vector search settings returned roughly 88% of relevant passages, which reads as an assistant that simply does not know things — the failure that destroys trust fastest.
What we built
Retrieval tuned for recall, every answer grounded in cited passages that link back to the source, and an explicit refusal when nothing relevant is retrieved rather than a plausible guess.
Services used
Got somethinglike this?
Describe it in plain language and get a priced scope back before you make an account.