Context
The PREVENT_URL_HALLUCINATION_INSTRUCTION system prompt injected into the conversation agent is not reliably preventing the model from generating unverified URLs.
Eval results on 6 test cases (3 runs each) show a 33% pass rate with the current instruction.
TODO
Improve prompt to increase evaluation success rate
Context
The
PREVENT_URL_HALLUCINATION_INSTRUCTIONsystem prompt injected into the conversation agent is not reliably preventing the model from generating unverified URLs.Eval results on 6 test cases (3 runs each) show a 33% pass rate with the current instruction.
TODO
Improve prompt to increase evaluation success rate