Skip to main content
Version: 3.6.1-rc.2

What are RAG endpoints?

What it is​

A RAG endpoint (Retrieval-Augmented Generation) is a retrieval system you deploy on the platform to ground model answers in your own documents. It stores your documents, splits and indexes them, and returns the passages most relevant to a query.

Each RAG endpoint is its own deployment with its own API, at https://api.<your-domain>/v1/api/{endpoint-id}. The portal shows this URL on the RAG endpoint's page. You add documents in two ways:

  • Upload files in the portal, or send them to the endpoint's API.
  • Connect a data source, a storage location where your data is stored. The endpoint checks it on a schedule, or when you start a sync, and picks up new, changed and deleted files.

A RAG endpoint returns context, not a finished answer. Answers come from a model that uses the endpoint as a tool.

Why it exists​

A model only knows what it was trained on. A RAG endpoint gives it more context and lets it answer from your own documents instead:

  • Answers from current, private data. The model searches your documents at question time.
  • Your documents stay in your tenant. Each endpoint has its own database, and only callers with access can search it.
  • One index, many consumers. Chat, your applications and coding agents all search the same endpoint.

How it fits in the platform​

files, data sources ──► RAG endpoint ◄── search ── Chat, your applications, MCP clients
│
└──► Model Gateway (embedding and other models)
  1. You fill the endpoint by uploading files or connecting a data source. To index them, the endpoint calls its embedding and other models through the Model Gateway.
  2. A caller searches it: with POST /context on the endpoint's API, as a knowledge base in Chat, as a tool through the Responses API, or from an MCP client such as a coding agent, through the gateway's MCP endpoint.
  3. When a model uses it as a tool, the gateway runs the search and passes the passages back to the model, which writes the answer. See Tools and preset agents.

More resources​