[system-prompt] Prevent LLM from generating false URLs #210
Description
Bug Report
Problematic behavior
User asks for information, the LLM fetch some data on internet (or not) and then during the conversation, the user ask for an e-mail redaction, and the LLM puts some links in the answer : those links are often wrong.
Expected behavior/code
The LLM should not generate URLs, it should either use known urls (like after a web search) or simply put a placeholder in the answer.
Steps to Reproduce
Environment
- Conversations version: v0.0.10
- Instance url: https://assistant.numerique.gouv.fr/
Possible Solution
See expected behavior
Additional context/Screenshots
See session c80ffe72-5ed4-4297-91d7-9f588f10689e in Langfuse