Skip to content
All projects
Case studyCS/17
StatusPersonal project
TypeAI
Year2025

SEC Filings RAG

Ask questions about FAANG 10-K filings with retrieval-augmented generation

Loads the 2023 10-K PDFs of the FAANG companies, chunks and embeds them into a FAISS index with LangChain, and answers financial questions from the filings themselves in a Streamlit app.

GitHub
  • 5 annual reports
  • FAISS vector index
  • Grounded answers
  • Dockerised

§1

The problem

General LLMs don't know the latest filings and tend to make up numbers. Answers about a company's finances should come from its actual report.

§2

My role

Built the ingestion pipeline, the vector index, the retrieval chain and the Streamlit UI.

§3

Architecture

Fig. 1 · SEC Filings RAG architecture
context10-K PDFSFAANG 2023SPLITTERrecursiveFAISSembeddingsSTREAMLITquestionRETRIEVERLLMgrounded answer
externalservicedataclient

§4

Key decisions

  1. D/01

    Retrieve, then generate

    Relevant chunks from the filings are put in the prompt, so answers are grounded in the source documents.

  2. D/02

    Index built once

    PDFs are split with a recursive text splitter and embedded into FAISS at build time; queries only search.

§5

Challenges

  • Chunk size vs. answer quality: financial tables break easily when split mid-row.

§6

Stack

  • Python
  • LangChain
  • FAISS
  • OpenAI
  • Streamlit
  • Docker

Next case study

FinTafsir

Upload a bank statement PDF, get structured JSON back