Ship a RAG chatbot on your database — embeddings stay yours.

Retrieval-augmented generation as ordinary endpoints on a database you already run. Ingest your documents, store the embeddings next to them, and answer questions from your own data — with no third-party vector store in the middle.

RAG without a proprietary vector silo

The embeddings live in your database, beside the documents they came from. There's no separate vector service to sync, secure, pay for, or explain to your compliance team — and nothing about your corpus has to leave your infrastructure.

• Postgres + pgvector, or MongoDB Atlas Vector Search
• Documents and embeddings in one place, under your backup policy
• Self-host it and the corpus never leaves your network

Two paths, same shape

The RAG Chatbot pack runs on Postgres with pgvector. The MongoDB Atlas Vector Search RAG pack does the same on Atlas. Fork whichever matches the database you already have — the endpoints look the same either way.

• RAG Chatbot — Postgres + pgvector with OpenAI embeddings
• MongoDB Atlas Vector Search RAG — the Atlas path
• LLM Gateway — a companion pack for completion and moderation

Ingest, chat, history, delete

A chatbot is a handful of endpoints, so that's what you get: seed and ingest documents, chat against them, read history, and delete. They're readable config — extend them like any other Air Pipe interface.

Observable, like every other endpoint

Retrieval is where RAG quietly goes wrong. Every call ships OpenTelemetry traces and Prometheus metrics automatically, so you can see what was retrieved and how long it took — rather than guessing why an answer was poor.

Frequently asked questions

• Do my embeddings leave my infrastructure? Not if you self-host. Embeddings are stored in your own database next to the source documents, so they stay wherever that database runs. Generating embeddings calls whichever model provider you configure — that call leaves your network; where the vectors live afterwards does not.
• Should I use pgvector or MongoDB Atlas Vector Search? Use whichever database you already run. Both paths ship as marketplace packs with the same endpoint shape, so the decision is about your existing stack and operational comfort rather than about Air Pipe.
• Can I self-host a RAG chatbot? Yes. The same config runs on the managed cloud or on the single self-hosted binary, so you can keep the documents, the embeddings and the database entirely on your own infrastructure.
• Which model providers can I use? The packs use OpenAI embeddings by default. The model, the API key and the system prompt are configuration, so you can point them at another provider without changing the endpoint shape.