Gateway Chat | PraisonAIUI Docs

Gateway ChatReal-time AI agent chat with WebSocket streaming, markdown rendering, tool call display, and session management.!Chat InterfaceQuick StartHow It WorksFull Streaming Data FlowA chat message travels through 5 layers before reaching the browser:Key Files| File | Layer | Responsibility ||------|-------|----------------|| aiui/plugins/views/chat.js | Frontend | WS connection, DOM rendering, session history || src/praisonaiui/features/chat.py | Backend | ChatManager, _run_and_broadcast, history API || src/praisonaiui/server.py | Backend | SSE streaming path (run_agent endpoint) || src/praisonaiui/providers/__init__.py | Provider | SDK → RunEvent bridge, _run_direct_mode || src/praisonaiui/datastore.py | Storage | MemoryDataStore / JSONFileDataStore || src/praisonaiui/provider.py | Protocol | RunEventType enum definition |ArchitectureAgent Tool ResolutionAgents created via YAML config, CRUD API, jobs, or channel bots automatically get tools resolved through the ToolResolver. Tool names in config are resolved to callable Python functions from 4 sources:1. Local tools.py file (backward compatibility)2. praisonaiagents.tools.TOOL_MAPPINGS (built-in tools)3. praisonai_tools package (community tools)4. Tool registry (programmatically registered tools)YAML ConfigurationCRUD APIAgents created via POST /api/agents also support tools:Where Tool Resolution Applies| Component | Tool Resolution ||-----------|----------------|| Gateway _create_agents_from_config() | ✅ ToolResolver || Integration create_gateway_from_yaml() | ✅ ToolResolver || Provider _get_or_create_agent() | ✅ Default tools || Channel bot _start_channel_bot() | ✅ Default tools || Jobs _execute_job() fallback | ✅ ToolResolver || CRUD agents _run() fallback | ✅ ToolResolver || CRUD agents _sync_to_gateway() | ✅ ToolResolver |FeaturesMarkdown RenderingAssistant messages are rendered with full markdown support:Bold, *italic*, ~~strikethrough~~Inline code and fenced code blocks with syntax highlightingOrdered and unordered listsLinks (auto-detected and rendered safely)Tool Call Display!Tool Call StreamingWhen agents use tools, each call is displayed as a collapsible card:Tool name and status indicator (running/completed/failed)Input arguments (JSON)Output results (rendered in code blocks)Tool Call LifecycleTool calls go through 4 stages: emit → enrich → dedup → persist.1. Emit — The SDK fires TOOL_CALL_START / DELTA_TOOL_CALL / TOOL_CALL_END events via StreamEventEmitter. The provider maps these to RunEvent types (TOOL_CALL_STARTED, TOOL_CALL_COMPLETED).2. Enrich — _run_and_broadcast() in chat.py enriches each tool call event with display-friendly fields via _enrich_tool_payload():| Field | Source | Example ||-------|--------|---------|| icon | Mapped from tool name | 🔍, 📝, 💾 || description | Generated from name + args | "🔍 Searching for 'Django latest version'" || step_number | Auto-incrementing counter per run | 1, 2, 3 || formatted_result | Truncated result string | "✓ Done" || tool_call_id | SDK-assigned or UUID fallback | "call_abc123" |3. Dedup — Both DELTA_TOOL_CALL (stream) and hook callbacks fire for the same tool call. The handler uses _seen_tool_started / _seen_tool_completed sets keyed by tool_call_id and name to suppress duplicates. A re-broadcast with has_complete_args=True is allowed to update the description with keyword-rich text.4. Persist — After the run completes, all enriched tool calls are merged by tool_call_id into collected_tool_calls and saved alongside the assistant message:Tool Call Persistence SchemaTool calls are persisted alongside assistant messages in the datastore. When a session is reloaded from history, tool call steps are reconstructed from the toolCalls array on each assistant message.Assistant message schema in the datastore:History Reload RenderingSession ManagementNew Session: Click "New Chat" to create a fresh sessionSession History: Switch between sessions in the sidebarMessage History: Full conversation history per session via REST APIMessage AbortClick the Stop button (or send chat_abort via WebSocket) to cancel an in-progress agent runThe partial response is preserved in historyFile AttachmentsUpload files to include with your chat messages — see Attachments for details.Session PersistenceChat history is persisted via the BaseDataStore interface (datastore.py). Two implementations are available:| DataStore | Persistence | Default ||-----------|-------------|--------|| MemoryDataStore | Volatile (in-memory only) | Yes (fallback) || JSONFileDataStore | Disk (~/.praisonaiui/sessions/) | Yes (when data_dir configured) |Each session is a JSON file containing:Messages are appended via add_message() and retrieved via get_messages(). Both methods are async to support future database backends.Response Content StrategyThe content field on assistant messages stores the SDK's final response from agent.chat() — the same text displayed by finalizeDelta() in the live view. This ensures the reloaded view matches what the user saw during streaming.> [!IMPORTANT]> During streaming, text tokens are accumulated in full_response. However, when RUN_COMPLETED arrives with event.content (the SDK's authoritative return value), it replaces the accumulated tokens for storage. This prevents intermediate narrative text from being duplicated on reload.Chat Message ModelWebSocket ProtocolClient → ServerServer → ClientAll events include session_id and run_id.REST API| Endpoint | Method | Description ||----------|--------|-------------|| /api/chat/send | POST | Send a message (non-streaming) || /api/chat/history/{session_id} | GET | Get message history for a session || /api/chat/abort | POST | Abort an active run || /api/chat/ws | WebSocket | Real-time chat streaming |Send MessageGet HistoryResponse includes toolCalls on assistant messages: