# Chat

Chat is an agentic RAG system that lets you ask questions grounded in your Atomic knowledge base.

## How It Works

The chat agent has tools to search your atoms during conversation. When you ask a question:

1. The agent decides whether to search your notes.
2. It formulates search queries and retrieves relevant chunks.
3. It synthesizes an answer grounded in retrieved content.
4. Responses stream back in real time over WebSocket events.

Chat can emit tool-start and tool-complete events, citations, and canvas actions. The REST call that sends a message returns the final assistant message, while the UI updates from streaming events as the model responds.

## Scoped Conversations

Conversations can be scoped to specific tags. When scoped, the agent only searches atoms under those tags, giving you focused answers about a particular topic.

## Conversations

Chat conversations are persisted. You can revisit previous conversations and continue where you left off. Each conversation tracks messages and scoped tags.

Conversations can also be renamed, archived, or deleted through the API/UI.

## API and Events

The primary endpoints are:

- `POST /api/conversations`
- `GET /api/conversations`
- `GET /api/conversations/{id}`
- `PUT /api/conversations/{id}`
- `DELETE /api/conversations/{id}`
- `PUT /api/conversations/{id}/scope`
- `POST /api/conversations/{id}/messages`

Streaming event names exposed to the frontend include `chat-stream-delta`, `chat-tool-start`, `chat-tool-complete`, `chat-complete`, `chat-canvas-action`, and `chat-error`.

## Provider Notes

Chat requires an LLM provider and model that can handle the conversation and tool-use workload. If chat fails but embeddings work, check the chat model setting separately from the embedding model.