Finance-LLM LangGraph Workflow

Project Overview

This project implements a stateful, modular financial analytics engine built using LangGraph for workflow orchestration and the Gemini API for intent detection, grounding, and final synthesis.

Its purpose is to take complex financial or crypto-related user queries and intelligently route them through multiple computational stages—data loading, statistical analysis, visualization generation—before producing a grounded, high-quality academic or investment-focused answer.

---

Objectives

1. Dynamic Routing

Execute different workflow paths depending on query intent:

• simple
• single_asset
• compare
• visual_request

2. Statistical Analysis

Use mock BTC/ETH data to compute:

• Volatility
• Average price
• Price ranges

3. Visualization Artifacts

Generate:

• Mock chart artifacts (Base64 strings)
• Text summaries to support LLM reasoning

4. Gemini-Optimized Synthesis

Feed structured artifacts into gemini-2.5-flash-preview-09-2025 for accurate final outputs.

---

Tasks

Task 1 – Intent Detection, Asset Mode, & Mock Data Loading (fetch_data.py)

Intent/Mode Detection

• Classify the query and set asset_mode to:
  • simple
  • single_asset
  • compare
  • visual_request

Asset Identification

• Detect crypto assets (BTC, ETH).

Mock Data Loading

• Load pre-generated price data into shared LangGraph state.

---

Task 2 – Statistical Transformation (transform.py)

Perform computations using the loaded data:

• Average price
• Min & max price
• Volatility (std / mean * 100)

Store results in the shared state under stats.

---

Task 3 – Visualization Artifact Generation (visualize.py)

Base64 Chart Generation

• Produce mock Base64 charts for single-asset or comparison queries.

Textual Description

• Create a detailed explanation of the visual to help the LLM during final reasoning.

Artifacts stored:

• chart_base64
• chart_description

---

Task 4 – Final LLM Synthesis (model.py)

Simple Queries

• Answer using Gemini + optional Google Search grounding.

Asset Analysis & Comparison Integrate:

• Statistical metrics
• Visual descriptions
• Base64 charts
• Academic/investor system prompt

Output written to:

• final_answer

---

** src/

├─ memory/ │ └─ state.py # Shared LangGraph state model (FinanceState) │ ├─ nodes/ │ ├─ fetch_data.py # Task 1: Intent Detection (Sets asset_mode) │ ├─ transform.py # Task 2: Statistical Calculation │ └─ visualize.py # Task 3: Visual Artifact Description & Chart Generation │ ├─ graph/ │ ├─ model.py # Task 4: Final LLM Synthesis │ └─ graph_definition.py # Routers + workflow builder (Defines the Graph) │ notebooks/ └─ finance_workflow.ipynb # Minimal usage interface

2. Intelligent Routing Breakdown

Routing uses two routers that depend on asset_mode determined in fetch_data.

Asset Mode	Example Query	Workflow Path
simple	“Explain crypto volatility”	fetch_data → model_api
single_asset	“Analyze BTC last week”	fetch_data → transform → model_api
compare	“Compare BTC and ETH volatility”	fetch_data → transform → visualize → model_api
visual_request	“Show me a visual of BTC risk”	fetch_data → visualize → model_api

---

1. LangGraph Flow Diagram

flowchart TD

    A[fetch_data] --> B{Router 1<br>Asset Mode?};

    B -- simple --> M[model_api];
    B -- single_asset --> T[transform];
    B -- visual_request --> V[visualize];
    B -- compare --> T;

    T --> C{Router 2<br>Compare Mode?};

    C -- compare --> V;
    C -- single_asset --> M;

    V --> M;

    M --> E((END));

Loading