This project demonstrates an automated equity research analysis workflow for Brazilian energy sector companies, specifically Taesa and Engie. It leverages LlamaIndex for building a retrieval augmented generation (RAG) pipeline, Google's Gemini Pro as the Large Language Model (LLM) for analysis and summarization, and LlamaExtract (via LlamaCloud) for structured data extraction from PDF financial reports.

The workflow ingests quarterly financial reports (PDFs), extracts key financial and operational data, creates granular vector indexes, and then uses a multi-step LlamaIndex Workflow to:

1. Generate individual financial summaries for Taesa and Engie.
2. Perform a comparative analysis.
3. Produce a final equity research memo in a structured Pydantic format.

• Automated Data Extraction: Uses LlamaExtract with Pydantic schemas to pull structured data from PDF reports.
• Granular Indexing: Creates fine-grained nodes in LlamaIndex VectorStoreIndex for precise data retrieval.
• Multi-Step Analysis Workflow: Employs LlamaIndex Workflow to orchestrate asynchronous analysis steps.
  • Individual company summaries.
  • Comparative analysis.
  • Final memo generation.
• Structured Output: Uses Pydantic models for all LLM outputs, ensuring consistent and parsable results.
• LLM Integration: Leverages Google Gemini Pro for text generation and structured data prediction.
• Customizable: Prompts, Pydantic schemas, and modeling assumptions can be adapted for different companies or sectors.

Equity-Research-llama/
├── .env                        # Environment variables (API keys) - !! GITIGNORE !!
├── .venv/                      # Python virtual environment - !! GITIGNORE !!
├── data/
│   ├── reference/
│   │   └── modeling_assumptions.txt # Key assumptions for financial modeling
│   ├── release-engie/
│   │   └── ENGIE_REPORT.pdf    # Engie's PDF financial report
│   └── release-taese/
│       └── TAESA_REPORT.pdf    # Taesa's PDF financial report
├── notebooks/                  # Jupyter notebooks for experimentation (if any)
├── output/
│   ├── extracted_documents/    # JSON outputs from LlamaExtract
│   ├── final_equity_memo_YYYYMMDD_HHMMSS.json # Final generated memo
│   └── indexes/                # Persisted LlamaIndex vector stores
│       ├── engie/
│       └── taesa/
├── reference/                  # Reference materials, original notebooks
├── scripts/
│   ├── data_extractor.py       # Script for LlamaExtract data extraction
│   ├── index_builder.py        # Script for creating granular LlamaIndex indexes
│   ├── equity_analyzer_agent.py # Main script for running the analysis workflow
│   └── output_schemas.py       # Pydantic models for structured LLM outputs
├── README.md                   # This file
├── requirements.txt            # Python dependencies
└── tasks.md                    # Project task tracking

1. Clone the repository (or initialize if you have the files):

   # If cloning an existing repo:
   # git clone https://github.com/arthur0211/llama-equity-analyst-brasil.git
   # cd llama-equity-analyst-brasil

2. Create and activate a Python virtual environment:

   python -m venv .venv
   # On Windows
   .venv\Scripts\activate
   # On macOS/Linux
   source .venv/bin/activate

3. Install dependencies:

   pip install -r requirements.txt

   (Note: A requirements.txt will be generated in a later step. For now, ensure you have installed the packages mentioned in tasks.md)

4. Set up API Keys: Create a .env file in the project root with your API keys:

   GEMINI_API_KEY="YOUR_GOOGLE_GEMINI_API_KEY"
   LLAMA_CLOUD_API_KEY="YOUR_LLAMA_CLOUD_API_KEY"

   Replace YOUR_GOOGLE_GEMINI_API_KEY and YOUR_LLAMA_CLOUD_API_KEY with your actual keys.

5. Place Data Files:

   • Put Taesa's PDF report in data/release-taese/ (e.g., TAESA-Release-1T25.pdf).
   • Put Engie's PDF report in data/release-engie/ (e.g., 250507-Release-de-Resultados-1T25.pdf).
   • Update data/reference/modeling_assumptions.txt if needed.

The workflow consists of three main script executions:

1. Extract Data (LlamaExtract): This script uses LlamaExtract to parse the PDF reports and save the structured data as JSON. You'll need to run this for each company, typically by modifying the script to point to the correct PDF and output path.

   python -m scripts.data_extractor

   (Review scripts/data_extractor.py to ensure it's configured for the desired company and PDF before each run if not yet parameterized).

2. Build Indexes: This script takes the extracted JSON data and builds granular LlamaIndex vector stores.

   python -m scripts.index_builder

3. Run Equity Analyzer Workflow: This script loads the built indexes and runs the multi-step analysis to generate the final equity memo.

   python -m scripts.equity_analyzer_agent

   The output memo will be saved in the output/ directory.

• Parameterize scripts (data_extractor.py, index_builder.py, equity_analyzer_agent.py) using argparse for easier execution with different files/settings.
• Refine prompts for LLM steps to improve the quality and detail of generated summaries and analyses.
• Expand Pydantic schemas to capture more detailed financial or operational metrics.
• Implement more sophisticated error handling and logging.
• Add unit and integration tests.
• Explore different LLM models or LlamaIndex components.

Contributions, issues, and feature requests are welcome.

(This is a placeholder for a more detailed contributing guide if the project becomes more collaborative).