Retrieve relevant conversation history snippets before prompt construction

[lbedner]

Summary

When generating a reply, run a filtered vector search on conversation-memory to retrieve relevant older context and include hits in the prompt.

Acceptance Criteria

• Chat flow performs vector search for current user message (and optionally last 1-2 turns)
• Search filters by conversation_id and user_id (no cross-convo leakage)
• Top-K and min_score configurable (default top_k=8, min_score=0.0)
• Retrieved snippets de-duplicated with recent messages window (don't re-insert same content)
• Prompt template includes a framed "Retrieved memory" section
• Unit/integration test verifies retrieval + inclusion in prompt (mock vector store)

Implementation Notes

Modify prompt-building code path in AIService.chat() (where recent_messages = conversation.messages[-10:]) to:

1. Call rag.search(query=user_message, filter={"conversation_id": id})
2. Dedupe hits against recent_messages (by message_id)
3. Insert retrieved hits in a "Retrieved memory" block before the user message

Keep the last-N message window unchanged (memory is augmentation, not replacement).

Files to Change

• app/services/ai/ai_service.py - Or wherever chat prompt is built
• app/services/ai/rag.py - Search helper for conversation memory
• tests/services/ai/test_chat_rag.py - Integration tests

Dependencies

• Requires Index conversation messages into conversation-memory collection #419 (Index conversation messages) to be completed first

Estimate

1-2 days dev + testing