persona_preference_extraction

Post-response extraction pipeline for persona preference memory.

Runs as a fire-and-forget hook after each LLM response. Detects expressed opinions in the bot’s output and stores them as persona preferences via PersonaPreferenceManager.

Opinions are allowed to be wrong. No fact-checking. No alignment filtering.

async persona_preference_extraction.extract_persona_preferences(response_text, persona_id, manager, openrouter, redis, channel_id, platform='', triggering_user_id='', triggering_message_id='', context_messages=None)

Main extraction entry point. Designed for fire-and-forget.

1. Gate checks (length, regex, rate limit)

2. LLM extraction call (with optional surrounding context)

3. Dedup / reinforce / conflict-create

4. KG entity resolution (best-effort)

context_messages is an optional list of {role, content} dicts representing the conversation turns surrounding the triggering exchange (typically 2 before + the trigger + the bot reply). When provided, these are prepended to the extraction prompt so the LLM understands what the opinion was a reaction to.

Return type:

None

Parameters:

• response_text (str)

• persona_id (str)

• manager (PersonaPreferenceManager)

• openrouter (Any)

• redis (Any)

• channel_id (str)

• platform (str)

• triggering_user_id (str)

• triggering_message_id (str)

• context_messages (list[dict] | None)