LlamaIndex Research Analyst

Welcome to the AI Equity Research Analyst demo project! This application showcases the power of Large Language Models (LLMs) in generating equity research reports.

Overview

This Streamlit app demonstrates the use of LLMs, specifically using the Google PaLM API, to analyze and extract data from 10-K filings of renowned NYSE listed companies. The LlamaIndex orchestration framework, along with its SubQuestionQuery Engine, breaks down complex prompts into sub-questions, retrieves answers, and assembles the final equity research report.

Demo Features:

• Select from 7 renowned NYSE listed companies.
• App reads, analyzes the latest 10-K filings, and extracts data based on complex prompts.
• LlamaIndex's SubQuestionQuery Engine decomposes complex prompts into sub-questions, fetches answers, and assembles the final report.

Potential Enhancements:

• API-driven financial data access.
• Dedicated Python REPL for accurate calculations.

Try the demo here!

Tools and Technologies Used:

• LLM: Google PaLM API
• LLM Orchestration Framework: LlamaIndex
• Embedding Model: Hugging Face bge-small-en
• Vector Store: Llamaindex's in-memory SimpleVectorStore
• Query Engine: SubQuestionQueryEngine
• UI: Streamlit

Project Structure:

• app.py: The main application script for the Streamlit app.
• requirements.txt: Lists the Python packages required for this project.
• storage/: Directory contains vector store index files which are pre-generated for the 10k files to improve demo app performance
• tenk/: Directory holding data or resources, likely related to 10-K filings.

Setup and Usage:

1. Clone the Repository:

   git clone https://github.com/AI-ANK/ai-equity-research-analyst
   

2. Install Required Packages:

   pip install -r requirements.txt
   

3. Run the Streamlit App:

   streamlit run app.py
   

Building Vector Index from Scratch

If you wish to build the vector index from scratch instead of using the pre-generated vector index, you can follow the steps below:

1. Make sure you have loaded the required files in the load_data function.
2. Uncomment the code snippet in app.py related to building the vector index.
3. Run the app.py script.

storage_context = StorageContext.from_defaults()
index = VectorStoreIndex.from_documents(tenk_company, service_context=service_context, use_async=True, storage_context = storage_context)
index.set_index_id("index_"+company)
index.storage_context.persist(persist_dir="storage")

Note: Building the vector index from scratch might take a considerable amount of time depending on the size of your data.

Feedback and Contributions:

Feel free to raise issues or submit pull requests if you think something can be improved or added. Your feedback is highly appreciated!

---

Developed by Harshad Suryawanshi. If you find this project useful, consider giving it a ⭐ on GitHub!