Semantic Search Over Nike 10-K (Part 1)

This project demonstrates how to build a basic semantic search engine over a real-world 10-K corporate filing. It serves as the foundation for more advanced retrieval-augmented generation (RAG) systems, showcasing how to move from unstructured PDF text to meaningful, queryable insights.

Developed for the MSDS 442 course at Northwestern, this first iteration emphasizes core skills in document ingestion, embedding, and vector similarity querying using LangChain and GPT-4o-mini.

⸻

Features

• PDF Ingestion: Uses PyPDFLoader to extract and structure raw text from Nike’s 10-K filing.
• Chunking: Splits content into manageable pieces using RecursiveCharacterTextSplitter.
• Embedding & Storage: Transforms text into embeddings with OpenAI and stores them in a Chroma vectorstore.
• Semantic Search: Supports question-based search with ranked similarity scores.
• Retriever Construction: Compares default and custom retriever approaches for document relevance.
• Batch Querying: Demonstrates LangChain’s Runnable pipelines for efficient multi-query handling.

⸻

Key Insight

AI applications that rely on information retrieval should treat LLMs not as oracles, but as reasoning layers on top of curated memory. This project shows how to build that memory effectively.

View the source code on GitHub