RAG Pipeline Starter Kit

Building a RAG system from scratch means stitching together embeddings, vector storage, chunking logic, retrieval, and prompt assembly — then wiring it to an LLM. This starter kit does all that work so you can focus on your domain. **What's included:** `ingest.ts` — Document ingestion pipeline that reads files, splits them into overlapping chunks (configurable size and overlap), calls `text-embedding-3-small`, and upserts vectors into a Supabase `pgvector` table. `retrieve.ts` — Query function that embeds the user question, runs a cosine similarity search against pgvector, and returns top-k chunks with metadata. Supports a configurable similarity threshold to filter low-relevance results. `rag-query/route.ts` — Next.js API route that orchestrates the full RAG flow: retrieve relevant chunks → assemble a system prompt with injected context → stream the response from OpenAI GPT-4o. `schema.sql` — Supabase migration that creates the `documents` table with a `pgvector` column, installs the `vector` extension, and sets up an HNSW index for fast ANN search. **Configuration options:** chunk size, overlap, embedding model, top-k retrieval count, similarity threshold, LLM model selection, and system prompt template. Deployable on Vercel in under 30 minutes. Environment variables documented in the included `.env.example`.