feat: code-level filter for user-only knowledge extraction #9
Description
Context
From #6 — when using smaller/local models (e.g. Qwen), assistant messages can leak into extracted knowledge despite prompt-level instructions to only extract from user messages.
Problem
The extraction prompts instruct the LLM to only extract from >>> USER: lines and ignore ASSISTANT: lines. This works well with capable models but is unreliable with smaller or local models that don't follow instructions as strictly.
Proposal
Add a code-level filter in the extraction pipeline that strips or validates extractions against the actual user message content, so extraction correctness doesn't depend entirely on model instruction-following capability.