You can edit functions in graph_agent.py, and in evaluator.py can change test_mode (LEGACY is langchain mode, GRAPH is langgraph)

🛠️ Prerequisites

Before you begin, ensure you have the following installed:

• Python 3.11 (Strict requirement)

• Google Cloud API Key or other LLM Key

⚙️ Environment Setup

1. Virtual Environment Setup

It is highly recommended to use a virtual environment to manage dependencies.

For macOS / Linux:

# Create virtual environment
python -m venv venv

# Activate environment
source venv/bin/activate

For Windows:

# Create virtual environment
python -m venv venv

# Activate environment
venv\Scripts\activate

2. Install Dependencies

pip install -r requirements.txt

3. Environment Variables (.env)

Rename the file .env_example to .env in the root directory and add your API_KEY

📂 File Descriptions

• data/: Folder containing the raw PDF financial reports
• langgraph_agent.py: [MAIN WORKSPACE] This is where you will write your code. It contains the logic for:

  • PDF Ingestion: initialize_vector_dbs()

  • Graph Nodes: retrieve_node, grade_documents_node, generate_node, rewrite_node.

  • Legacy Agent: run_legacy_agent (The baseline for comparison).

• evaluator.py: The benchmark testing script. It runs a suite of test cases (Apple Revenue, Tesla R&D, Comparison, Traps) and uses "LLM-as-a-Judge" to score your agent (Pass/Fail).
• config.py: Configuration file that handles API key loading and initializes the LLM and Embedding models.

📝 Student Tasks

You need to complete the TODO sections in langgraph_agent.py.

• Task 1 (Legacy): Implement the run_legacy_agent Prompt Template to establish a baseline (langchain).

• Task 2 (Router): Implement the retrieve_node logic to route queries to "apple", "tesla", or "both".

• Task 3 (Grader): Implement the grade_documents_node to filter out irrelevant documents.

• Task 4 (Generator): Implement the generate_node to answer questions in English with Citations.

• Task 5 (Rewriter): Implement the rewrite_node to refine search queries when retrieval fails.

🚀 Execution Order

• Step1: python build_rag.py: Before running any agents, you must ingest the PDFs and convert them into vector embeddings. This allows you to experiment with different chunking strategies without re-running the evaluation logic every time.
• Step2: python evaluator.py: Once the database is ready, run the evaluator to benchmark your agent.