Retrieved context is not evidence by default.
A retrieval system can return chunks that are nearby, partially relevant, duplicated, stale, or too weak to support the answer. If answer synthesis consumes all of that blindly, the system can produce a confident response with a poor grounding trail.
Candidate context vs selected evidence
| Layer | Meaning |
|---|---|
| Candidate context | Material returned by retrieval |
| Selected evidence | Material considered strong enough to use |
| Answer context | Evidence passed into synthesis |
The difference matters. Candidate context can be broad. Evidence should be narrower.
Grounding rule
If the answer cannot be traced to selected evidence, the answer is not grounded.
Evidence selection checks
Public-safe examples of checks include:
- Does the context answer the actual question?
- Does it resolve the right entity or source?
- Is it strong enough for synthesis?
- Is the evidence contradictory?
- Is the evidence missing?
The failure mode
The common failure is subtle: the model receives too much context and writes a plausible answer from fragments. That can look polished while still being wrong.
Agentic RAG needs a boundary before synthesis so the system can decide whether it has enough evidence to answer at all.