RAG + Tools

Build agents that use retrieval-augmented generation (RAG) or tool-calling patterns.

Approach

1. Define events — e.g. rag-request (query + context) and rag-response (answer chunks)
2. Agent logic — Fetch from your vector store, build context, call the LLM with the augmented prompt
3. Streaming — Emit response chunks as they arrive from the LLM

Tool-Calling Pattern

For agents that call tools (e.g. search, calculator):

1. Emit a tool-request event with the tool name and params
2. Another agent (or external system) handles the tool and emits tool-response
3. The original agent receives tool-response and continues

This can be implemented with multiple agents on different channels, or with a single agent that manages tool state internally.

For a single agent calling an LLM directly, define tools with Tool.of() and pass them to the provider with tools.toOpenAI(), tools.toAnthropic() or tools.toAiSdk({ jsonSchema }). Feed the results back with executeForModel(). See How to add tools to an agent.

Example Repos

• core-example — Basic streaming agent; extend with RAG by adding retrieval before the LLM call
• open-ai-speech-to-speech-example — Uses OpenAI Realtime API; shows integration patterns

Common Recipes

See Common Recipes for copyable snippets (retrieval, tool loops, etc.).