How it works

Four stages. One repo.

Your MDX is the corpus. The build step is the indexer. A small API route is the retriever. A streamed call to your model is the synthesizer.

Stage 01

Chunk

Each .mdx file under content/docs/ becomes a page. Wrap retrievable units in <Chunk id>. Frontmatter carries the metadata. Versioned with git.

Stage 02

Index

npm run ingest walks the tree, embeds chunks with Voyage AI using your key, and writes them to a single sqlite-vec file at data/docs.db.

Stage 03

Retrieve

/api/docs/search embeds the query, runs a vec0 ANN scan with a lexical filter, and returns top-k chunks. Spotlight (⌘K) calls the same route.

Stage 04

Answer

The chat panel packs chunks into a system prompt, appends history, and streams the model's answer through your provider key. Citations link to the source chunk.

Hosted RAG stack

Managed vector database
Embedding pipeline (queue, worker, retry)
Backend & query broker
Retrieval observability platform
Monthly bill
On-call rotation

The doks way

sqlite-vec in a single file
npm run ingest at build time
One Next.js route handler
The Network tab + your own chunks
Whatever your static host already costs
Nothing to operate

Legal