Collection Intelligence

User collections stored in Supabase (PostgreSQL) with Row-Level Security. Each user's vinyl, books, and songs are isolated via auth policies. Data is fetched at runtime and injected into the LLM context window.

Collection data is serialized into a structured system prompt. The LLM receives the full dataset as context — artist names, genres, ratings, statuses — enabling it to reason across the collection without a vector database for this scale.

Claude API (Anthropic) handles natural language understanding via a secure serverless proxy (Netlify Functions). The API key never reaches the browser — it lives as an environment variable on the server. The system prompt defines the assistant's persona, knowledge boundaries, and response style. Multi-turn conversation history is maintained client-side.

Intent recognition (recommendations, analysis, comparison, search). Entity extraction (artist names, genres, ratings). Sentiment-aware responses. Cross-collection reasoning (pairing books with albums by theme).

User query → context assembly → serverless function proxy → Claude API with system prompt + conversation history → response → UI render. The API key is stored as a Netlify environment variable — zero exposure to the client. Token-efficient design keeps costs low.

For larger collections: embeddings via OpenAI/Cohere → vector DB (Pinecone/pgvector) → RAG pipeline. LangChain for tool-augmented generation. Currently unnecessary at this collection size but architecturally planned.