Analyze Financial Documents with LlamaIndex

This notebook shows how to perform financial analysis over 10-K documents - the lengthy, jargon-heavy annual reports companies file with the SEC - using LlamaIndex, a data framework for building RAG systems in just a few lines of code. After installing llama-index and configuring gpt-3.5-turbo-instruct as the LLM via a global ServiceContext, two 10-K PDFs (Uber and Lyft, 2021) are loaded and parsed into per-page Document objects (each over 100 pages, so loading takes a while), then indexed into an in-memory VectorStoreIndex (which calls the OpenAI API to compute embeddings over document chunks). Simple QA is run via a QueryEngine configured with similarity_top_k controlling how many retrieved chunks (Node objects) are used as context per answer. For more complex, cross-document analysis - comparing and contrasting Uber's and Lyft's financials - a SubQuestionQueryEngine breaks a compound compare-and-contrast query into simpler sub-questions, each routed to its respective company's index, then synthesizes the sub-answers into a combined response.

The motivating problem is that extracting information and synthesizing insight from long financial documents is a core part of a financial analyst's job, and a 10-K - the annual report the SEC requires public companies to file, giving a comprehensive summary of financial performance - typically runs hundreds of pages with domain-specific terminology that's hard for a layperson to digest quickly. LlamaIndex is introduced as a general data framework for LLM applications: a RAG system can be stood up in a few lines of code, and for more advanced use it also offers a broader toolkit for data ingestion and indexing, retrieval and re-ranking modules, and composable components for building custom query engines beyond what this specific notebook demonstrates. The notebook itself follows a fixed outline - Introduction, Setup, Data Loading & Indexing, Simple QA, and Advanced QA (Compare and Contrast) - walking from installing the library and configuring the LLM through to the two query patterns.

Use the simple QueryEngine + VectorStoreIndex pattern for straightforward Q&A against a single long financial document. Use the SubQuestionQueryEngine specifically when a query requires synthesizing information across multiple documents (e.g. comparing two companies' financial metrics), since it decomposes the compound question into single-document sub-questions automatically rather than requiring you to manually split the query yourself.

Input: 10-K PDF filings (or similar long financial documents). Output: direct answers to single-document questions via simple QA, or synthesized compare-and-contrast answers across multiple companies' filings via the sub-question query engine.

Requires the llama-index library and an OpenAI API key/model (gpt-3.5-turbo-instruct in this example) for both embedding computation and query answering.

Financial analysts and developers who need to quickly extract information and synthesize insights from long, jargon-dense financial filings - including across multiple companies - without building a custom RAG pipeline from scratch.