DocQuery

The interesting part of a RAG product is not the chat interface — it is retrieval quality. A naive chunk-and-embed pipeline will confidently answer a question using text from the wrong page, and the user has no way to tell. The failure is silent, which makes it the worst kind.

Most of the engineering went into the boundary between the document and the model: how big a chunk should be, how much chunks should overlap so a sentence split across a boundary is not lost, how many neighbours to retrieve, and how much of that retrieved context to actually hand the model before it starts ignoring the middle of it.

Uploads land through UploadThing so large PDFs never pass through a serverless function body. The file is parsed, split into overlapping windows, embedded, and upserted into Pinecone under a namespace scoped to the file — so retrieval can never leak across documents or across users.

At query time the question is embedded with the same model that indexed the document, the nearest chunks are pulled back, and the question plus retrieved context plus the recent turns of the conversation are streamed to the model through the Vercel AI SDK. The response streams token by token into the client rather than making the user wait on a complete answer.

Billing is a Stripe subscription with a free tier enforced server-side at upload time. Auth is Kinde, so I did not hand-roll sessions.

The whole stack is typed end to end with no generated client and no hand-written request types. Prisma types the database, tRPC v11 infers the router's input and output types straight into the React client, and Zod validates at the boundary. Renaming a database column surfaces as a red squiggle in a component, at author time, not as a runtime error in production.

That property is the reason I would build it this way again. It removes an entire category of bug rather than catching it later.