Overview

A map of the SID-1 retrieval pipeline split into the three stages this Build section covers. A question feeds into SID-1; SID-1 exchanges repeated tool calls (BM25, ANN, and similar) with a search backend and receives content and metadata back; SID-1 returns ranked results. An outer dashed box labeled Agent loop encloses the entire process. Inside it, Make a request encloses the question and SID-1, and Run the tools on your backend encloses the search backend. Together they show how the Build pages map onto the parts of the SID-1 agent.

The quickstart covers a basic implementation of SID-1; now we'll go deeper on integrating SID-1 into your search stack.

Build overview

Make a request: call SID-1 with the system prompt and tool names.
Run the tools on your backend: connect search, text_search, and read to your index.
Loop and exit: append tool results, call SID-1 again, and stop when it reports IDs.
Best practices: add parallelism, timeouts, retries, turn caps, and evals.

Each step maps onto one part of the SID-1 pipeline you saw on the overview; the diagram above shows which part each page covers.

What you implement vs what SID-1 does

You implement the search tools and the loop that runs them against your backend. SID-1 decides what to search, how to phrase each query, and when to stop.

See what you build, and what SID-1 does for the split, and run the tools on your backend for the detail.

Build overview

What you implement vs what SID-1 does

Content