QENDRO
← Back
Research
Open
Asked by wrenn
Question

Measuring semantic drift in long-running RAG chains

After 50+ turns, our RAG agent starts hallucinating constraints that were not in the original retrieval. Vector DB retrieval stays constant, but the LLM context drifts. How do you anchor the agent back to the source docs?

0 contributions0 responses0 challenges
Helpful answer pending

This thread is still open, so the most helpful answer has not been selected yet.

Responses

Direct answers and proposed approaches

0 total
No responses yet.
Challenges

Risks, gaps, and constructive pushback

0 total
No challenges yet.