Bella Chat Service
The Bella Chat Service is the AI orchestration engine for Bella Keys. It routes user queries to the right tool — financial data via MCP, knowledge search via RAG, or a direct LLM response — and streams the result back token-by-token over SSE.
- Capabilities
- Architecture
- Workspace
Financial Data Access Natural language queries against accounts, spending entries, and budgeting periods. Routed to the EMS MCP Server via LLM tool calls.
Semantic Knowledge Search (RAG) Retrieves grounded answers from the personal wiki using Qdrant vector search. Responses include source hyperlinks.
Persistent Conversation Memory
Multi-turn sessions are checkpointed to PostgreSQL via AsyncPostgresSaver. Conversation history survives Electron window restarts.
SSE Streaming
Responses stream token-by-token as text/event-stream. Event types: thinking, tool_call, tool_result, response, error, done.
Configurable LLM Backend
Supports Ollama (local, default: qwen2.5vl:7b) and Google Gemini, switchable via SYNTHESIS_MODEL_PROVIDER.
Observability
LangChain traces instrumented with Arize Phoenix via openinference. Trace data stored in PostgreSQL.
OrchestratorAgent
Built with LangGraph's create_agent ReAct loop. Receives a user message, decides which tool to call (or answers directly), and iterates until a final response is ready.
SimpleChatAgent
A lightweight fallback agent for persona-based Q&A when no tool is needed. Single-node graph: generate_response only.
State Persistence
All conversation turns are serialized into bella_chat_checkpoints (PostgreSQL) by AsyncPostgresSaver. Tables are auto-created on first startup via checkpointer.setup().
