Your docs, chunked, embedded, and interrogated

What it does

This Next.js starter ingests .mdx files from your pages directory, splits them into chunks, and stores 1536-dimension embeddings in a Supabase Postgres database using pgvector. At runtime, an Edge Function turns a user’s query into an embedding, performs vector similarity search to find the most relevant chunks, and injects that context into an OpenAI GPT-3 text completion prompt, streaming the synthesized answer back to the client.

The interesting bit

The build-time generate-embeddings script checksums each .mdx file to avoid regenerating embeddings for unchanged docs. That small bookkeeping detail is where the boring infrastructure work actually pays off.

Key highlights

• Build-time embedding generation with incremental updates via file checksums.
• Vector similarity search powered by Postgres pgvector.
• Streams OpenAI completion responses to the client via an Edge Function.
• Database schema and local Supabase setup managed through migrations.
• Ships as a deployable Vercel template with Supabase integration.

Caveats

• Explicitly targets the older OpenAI Text Completion API and GPT-3, not the newer Chat Completions format.
• Requires documentation to be in .mdx format by default; plain .md files need renaming.
• It is template code, not a reusable library: expect to fork and modify the plumbing.

Verdict

Worth a look if you run Next.js and Supabase and want a concrete, deployable example of doc RAG without writing the wiring from scratch. Skip it if you need a framework-agnostic or drop-in search component.