Persona extraction — sample → LLM extraction → merge → store PersonaCard JSON

[Punkte]

What to build

An extraction pipeline that samples representative message groups from the conversation, sends them to Claude for structured analysis, merges the partial results into a single PersonaCard, and stores it in the personas table. This is the core intelligence of the product — validate quality before building UI around it.

Acceptance criteria

• Sampler picks messages from across the full timeline (5+ temporal segments), prioritising longer messages and complete exchanges
• Extraction prompt produces valid PersonaCard JSON matching the defined schema (writingStyle, communicationPatterns, personality, relationshipDynamics, signatureExpressions, language)
• Fusion prompt merges N partial cards into one coherent card, prioritising recent style
• PersonaCard stored in personas table linked to conversation
• Manual validation: run extraction on a real WhatsApp export and confirm the card is accurate and non-generic
• GET /conversations/:id/persona returns the stored card

Blocked by

• Chunking + embedding pipeline — temporal sliding window → pgvector #4