End-to-End LLM Project: Equity Research Analysis

Introduction

• The project involves building a news research tool for equity research analysis using an LLM.
• Utilizes Lang chain, OpenAI, and Streamlit.
• Features a storytelling aspect involving a character named Peter Pandey, inspired by a hypothetical Rocky Bhai.

Industry Use Case

• Mutual Funds: Common people invest in mutual funds, which then invest in individual stocks.
• Equity Research Analysts: Responsible for researching stocks (e.g., Tata Motors, Reliance) and providing insights.
• Analysts read numerous articles which can be tedious; hence this tool aids in compiling and querying news data.

Project Overview

• The tool takes multiple news article URLs, answers user queries based on these articles, and can summarize articles.
• Technical Framework:
  • Lang Chain: Used for processing and understanding language data.
  • Streamlit: Used for creating a user-friendly interface.
  • OpenAI: Provides the LLM capabilities.

Challenges with Existing Solutions

• Manual Entry: Copy-pasting articles is not feasible for busy analysts.
• Aggregate Knowledge Base: Necessary to compile data from various sources efficiently.
• Token Limit: ChatGPT has word limitations that are impractical for large article texts.

Technical Architecture

Initial Setup

• Document Loader: Gathers news articles.
• Text Splitting: Breaks down articles into manageable chunks.
• Vector Database: Stores these chunks to facilitate efficient searching.

Advanced Setup

• Semantic Search: Uses embeddings and vector databases for context-aware searches.
• Phase Usage: Lightweight in-memory vector database (alternative to Pinecone, Milvus, Chroma).

Tools and Libraries

• Lang Chain: Provides document loaders, text splitters, and embedding classes.
• OpenAI API: Used for creating embeddings and LLM interactions.
• Sentence Transformer: Converts text into vectors using predefined models.
• Phase Library: Allows fast similarity search within vector data.

Building the Tool

Steps Involved

1. Load News Articles: Use unstructured URL loader to gather articles' content.
2. Text Splitting: Utilize recursive text splitter for chunking articles.
3. Embedding: Convert text chunks into embeddings using OpenAI.
4. Create Vector Index: Use phase library to store embeddings and perform search operations.

Execution

• Streamlit Interface: Interactive UI with sidebars for URL input and question box for queries.
• Processing URLs: On clicking process, data is loaded, split, embedded, and stored in the vector index.
• Answering Queries: Retrieve and display answers based on question input using embeddings and vector search.

Advanced Features

• Retrieval QA with Sources Chain:
  • Stuff Method: Combines all relevant chunks in one LLM prompt.
  • Map Reduce Method: Divides the task into multiple LLM calls per chunk, then aggregates results.

Deployment Considerations

• Proof of Concept: Initial development using Streamlit for quick deployment and testing.
• Long Term Architecture: Incorporate reliable data ingestion systems and comprehensive vector databases.

Conclusion

• The tool is designed for real-world application by equity research analysts.
• It enhances efficiency by automating retrieval and analysis of news data.
• Encouraged to share and utilize the provided code for enhancing personal or professional projects.