Grounded assistant over support knowledge
Support agents search across a knowledge base, historical tickets and policy documents that live in different systems.
This is an illustrative reference design describing how Protogenies approaches this class of problem. It is not a named client engagement, and it contains no performance claims.
What has to be solved
Answers are inconsistent, and a plain chatbot cannot be trusted because it cannot show where an answer came from.
Technical constraints
- Answers must cite a source
- Agents must not retrieve content outside their permissions
- Cost per interaction has to stay bounded
Approach
A retrieval pipeline with hybrid search and reranking, generation constrained to retrieved context with required citations, permission filtering applied at query time, and an evaluation set built from real questions.
Implementation
- 01Incremental ingestion with per-source access metadata
- 02Chunking and embedding strategy validated against retrieval metrics
- 03Citation enforcement plus an explicit refusal path when evidence is thin
- 04Evaluation harness running in CI on prompt and model changes
- 05Tracing of prompts, context, tokens and latency per interaction
Technology stack
Expected qualitative outcomes
- Every answer links to the source an agent can open and verify
- Model and prompt changes are gated by an evaluation run
- Retrieval respects the permissions of the requesting user
- Cost and latency per interaction are observable and bounded
Things to weigh up
- Retrieval quality usually matters more than model choice
- Source content that is out of date will be faithfully repeated
Talk to the engineers who would do the work
Bring your current architecture, constraints and the problem you are trying to solve. We will tell you what we would change first, what it depends on, and where we would start.