An assistant that shows its sources
A retrieval-augmented assistant built so every answer carries a clickable citation to the document it came from — designed for a buyer who has no reason to take an AI's word for it.
This work is being carried out by Bipasha Das inside her current full-time employer, not under OPERIYA and not for an OPERIYA client. It is published here as implementation experience, described in her own words and generalised: the organisation is not named, and its commercial data is not disclosed.
- Answers that carry a source link
- 100%A property of the build rather than a measurement: the citation is part of the response, so there is no answer without one.Design decision, stated by Bipasha Das.
The problem
An assistant that answers questions about a business is only useful if the answers can be relied on, and a confident paragraph from a language model gives a reader nothing to check. For a purchase of any size the questions that matter — what is included, what it costs, what the terms are — are exactly the ones where a plausible-sounding wrong answer is worse than no answer at all. Meanwhile the same questions, asked over and over, occupy the people whose job is to qualify and talk to buyers.
What changed
The assistant was built as retrieval-augmented rather than conversational-only: it answers from the organisation's own material rather than from a model's general knowledge. The part that matters for trust is what it does with that material — every answer carries a clickable citation to the source it was drawn from, so a reader can open the document and check the claim instead of taking it on faith.
The second purpose is qualification. The assistant handles the repeated, answerable questions at the front of a conversation, which is the part of qualification that does not need a person, and leaves the judgement to the people who are good at it.
Why it was built this way
Citations were a requirement of the build, not a feature added to it. An assistant that cannot show where an answer came from puts the burden of verification on the reader while removing their means of doing it, and in a sales context that is a liability rather than an efficiency: the cost of one confidently wrong answer about price or terms is larger than the time saved by answering a hundred easy questions.
A question goes to retrieval, retrieval reads the organisation's own documents, and an answer is returned. A second line runs from the answer back to the documents: the citation, which is what lets a reader check the claim.
What it produced
The assistant answers from the organisation's own documents, with a link to the source on every answer, and takes the repeated front-of-conversation questions off the people doing qualification.
The figure at the top of this page describes how the assistant is built, not what it achieved. Nothing was measured here: the effect on buyer trust and on qualification was the reason for the build, not something captured as a before and an after. Stating an unmeasured intention as a result is the move this site refuses to make, so the intention is labelled as one.
- Every answer is traceable to the document it came from.
- The repeated opening questions are handled without a person.
- A reader can verify an answer rather than trust it.
What it says about operations generally
Trust in an AI system is an interface problem as much as a model problem. Retrieval decides whether the answer is grounded; the citation decides whether anyone can tell. Build both, or the accuracy is unverifiable and therefore, to a buyer, indistinguishable from confidence.
Is something like this running in your business?
The audit starts by finding out. One hour, no charge, and you leave knowing whether it is worth buying.