Signal.

The challenge

A fluent response is not enough to establish trust. A useful knowledge assistant needs to show what it found, explain when its sources are insufficient, and make consequential actions reviewable.

The direction

Make the workflow visible: query, retrieval, grounded excerpts, human review. This prototype runs against six synthetic documents in the browser, with deterministic token matching rather than a connected language model.

Under the surface

• The working prototype. Query tokens are normalised, common words are removed, and distinct term matches rank the sample documents. Up to two matching passages are returned with their document title and section.
• The agentic workflow. The interface models a planner-to-retriever-to-reviewer handoff. The review control changes only local session state; it cannot publish content or call external tools.
• A proposed production RAG architecture. A production adapter would ingest approved documents, enforce access during retrieval, combine keyword and vector search, rerank context, and ask a model to answer with citations. Those services are not connected to this demo.
• Evaluation before deployment. The proposed evaluation covers source relevance, unsupported answers, permission boundaries, and reviewer escalation. The current prototype is tested for supported queries, empty queries, and no-match responses.

In practice

The live experiment on this page demonstrates retrieval, source attribution, an explicit no-answer state, and a review interaction. It is a studio prototype, not a client deployment or a claim of measured business results.