Set up your agent's retrieval in 5 minutes with Claix Retrieve

What agentic RAG is, and where SQL retrieval fits

RAG means a model uses information retrieved from an external source. It does not require documents, embeddings or a vector database. When an agent chooses to call a tool to fetch information, the retrieval is agentic. That tool can search documents, query an index, or run a controlled SQL query. “Agentic” describes how retrieval is triggered. “Vector” describes one way to find results. They are not synonyms.

Claix Retrieve is a tool for relational data. If the agent needs tomorrow's bookings or which orders are still pending, it can ask the live database instead of a copied index. Claix does not require copying your tables into a vector store. A precise name is controlled SQL retrieval for AI agents. It can sit inside an agentic RAG architecture without an embeddings pipeline.

When to use Claix Retrieve

Typed formats include strings, integers, booleans, dates, timestamps, JSON and some lists. In raw mode Claix returns the query rows. For semantic search over large text collections, a vector index can still be the right tool. For statuses, amounts, dates, counts and current SQL records, direct retrieval can skip an extra index.

What you need first

A Claix account, an API key, and a read-only database connection. Pick one concrete question, such as “How many orders are confirmed?”. Claix documents SQL databases including Postgres, MySQL, MariaDB and Postgres-compatible setups such as Supabase.

Step 1. Connect the database and review the Data Map

Claix drafts a Data Map of tables, descriptions and relationships for you to review. A column named status might mean payment, shipping or an internal check. The connection uses a read-only role. Retrieve is built to query data, not to run writes.

Step 2. Create a route with one job

A route is a retrieval capability: the tables and columns it may use, plus a short guide with your vocabulary and rules. A support route and an internal analytics route can share a database without exposing the same columns.

Step 3. Create the agent and choose the response

In reasoned mode every question must include the required format. If the data is missing or the type is not safe to produce, the documented result can be null.

Step 4. Connect Claix over MCP

The Claix MCP server is at https://claix.dev/mcp. The documented retrieval tool is claix.retrieve.ask. A Cursor-style config:

{
  "mcpServers": {
    "claix": {
      "url": "https://claix.dev/mcp",
      "headers": { "x-api-key": "YOUR_API_KEY" }
    }
  }
}

Keep the API key in the client's secret store. MCP is the connection. Access control still depends on the database role, the routes and the agent.

Step 5. Ask once and read the trace

curl -X POST "https://claix.dev/retrieve/order-support" \
  -H "x-api-key: YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"questions":[{"question":"How many orders are confirmed?","format":"int"}]}'

Replace order-support with an active agent slug that has a published route. The response includes the questions, the results and a log_id. You can send up to five questions per request. Also try questions the route should not be able to answer.

Why this setup is low-code

Production apps still need auth, secrets and error handling. What you stop building is the retrieval layer itself. Before go-live, check ambiguous questions, empty results and requests outside the route. If you use your own model key, test it in the Sandbox. Table access is not a substitute for per-user or per-tenant authorization.

Start with one question

Pick one useful question, connect the database with a read-only role and create a small route. When that query works and you can read its trace, widen the perimeter only if the use case needs it. Start at https://www.claix.dev/ and read the MCP docs at https://www.claix.dev/documentation/mcp.