Agentic Financial Advisor Chatbot

A conversational financial advisor agent built with LangGraph and LangChain's create_agent API. The agent manages multi-turn conversations about personal finances — checking balances, reviewing transactions, calculating budgets, and executing transfers — while adapting its behavior dynamically based on the user's membership tier.

Architecture

The agent is built around a ReAct-style tool-calling loop powered by LangGraph's stateful graph execution. User context (ID, name, tier, preferred currency) is injected at runtime via a typed UserContext dataclass and made available to all tools through LangChain's ToolRuntime.

Key Features

• Tier-based model routing — basic users get GPT-4o-mini (512 tokens), premium users get GPT-4o (2048 tokens), and platinum users get an uncapped GPT-4o instance, selected dynamically via middleware
• Dynamic system prompts — the agent's persona and verbosity adapt to the membership tier using a @dynamic_prompt middleware
• Financial tools — account balance lookup, transaction history retrieval (per account), budget allocation calculator, and money transfers with validation
• Structured output — platinum-tier responses return a FinancialResponse Pydantic model with summary, details, action items, warnings, and a confidence level
• Short-term memory — multi-turn conversation history is persisted with LangGraph's InMemorySaver checkpointer, scoped by thread_id
• Streaming — supports token-level streaming via agent.stream() in both updates and values modes
• Middleware error handling — a @wrap_tool_call middleware catches ValueError and KeyError from tools and returns graceful error messages instead of crashing

Tech Stack

• LangGraph / LangChain — agent orchestration, tool node, middleware pipeline
• OpenAI GPT-4o / GPT-4o-mini — language model backends
• Pydantic — structured response schema
• Python — implementation language