System Architecture

The diagram below shows the full stack: the Next.js/React UI communicates via REST API with a FastAPI backend deployed on Vercel. The backend runs a LangGraph supervisor that routes to specialist agents, each with access to specific tools and knowledge sources.

Agent Flow

Every request follows this path:

1. Model Documentation Reader — Reads and caches the financial model's documentation tab on first interaction.
2. Supervisor / Planner — Analyzes the user's question using structured output to decide which specialist agent should handle it.
3. Specialist Agent — Executes the appropriate playbook. Tool-using agents (Recall, Goal Seek) enter a tool-call loop, calling Google Sheets tools and returning results until the task is complete.
4. Response — The final agent response is returned through the API to the UI.

Memory Architecture

The system uses five memory types for comprehensive context management:

• Short-Term Memory — Conversation history via LangGraph checkpointer (MemorySaver).
• Long-Term Memory — Cross-session facts stored in InMemoryStore, persisted across threads.
• Semantic Memory — Embedding-indexed knowledge with text-embedding-3-small for similarity search.
• Episodic Memory — Timestamped records of past agent interactions and outcomes.
• Procedural Memory — Agent playbooks and standard operating procedures for each task type.