Platform to build search over your data (e.g. emojis, memes). Allow to build optiomal search and validate quality - easily switch models/datasets/search/etc.
Building search
- Upload – JSON array of rows → one dataset and many documents (content + metadata).
- Search datasets – From one dataset create multiple search configs (description prompt, description model, embedding model/dimension). Each config gets its own search_dataset and search_documents (content copy + LLM description + embedding in 384/768/1536/3072 dimensions).
- Describe – LLM (e.g. OpenAI) fills descriptions per row using the config prompt.
- Vectorize – Embed content + description and store in the chosen embedding column.
- Search – Hybrid (vector + full-text) over a search_dataset; returns ranked docs.
Measuring quality
- Validation – Upload query + expected doc IDs; run search and get recall@k / MRR.
Stack: Next.js, Supabase (Postgres + pgvector), OpenAI (or configurable) for describe/embed. All Supabase access from the server (API routes) with the service role.