PDF querying Faiss

Overview

This project ingests PDF files from a local docs/ directory, splits them into overlapping text chunks, and stores their vector embeddings in a FAISS index saved to disk. At query time, the user's question is embedded and matched against the index to retrieve the most relevant chunks. Those chunks are passed as context to an OpenAI GPT-4o model, which generates an answer constrained to the retrieved material.

Key capabilities:

• Automated PDF ingestion — recursively loads every *.pdf under docs/.
• Local vector store — FAISS index is built once and persisted (faiss_index/) for fast reuse.
• Semantic retrieval — top-k similarity search using all-mpnet-base-v2 embeddings.
• Grounded generation — GPT-4o answers strictly from the retrieved chunks.

Architecture

            +-------------------+
            |   PDFs in docs/   |
            +-------------------+
                     |
                     v
        +-----------------------------+
        |  DirectoryLoader (PyPDF)    |   load PDF pages
        +-----------------------------+
                     |
                     v
        +-----------------------------+
        | RecursiveCharacterText      |   chunk_size=1200
        | Splitter -> text chunks     |   overlap=200
        +-----------------------------+
                     |
                     v
        +-----------------------------+
        | HuggingFace Embeddings      |   all-mpnet-base-v2
        | (sentence-transformers)     |
        +-----------------------------+
                     |
                     v
        +-----------------------------+
        |   FAISS IndexFlatL2         |   save_local("faiss_index")
        |   vector store on disk      |
        +-----------------------------+
                     |
        =============|=============  (build phase above / query phase below)
                     |
   user question     |
        |            v
        |   +-----------------------------+
        +-->|  similarity_search(k=2)     |   retrieve top chunks
            +-----------------------------+
                     |
                     v
        +-----------------------------+
        |  Prompt = question + chunks |
        +-----------------------------+
                     |
                     v
        +-----------------------------+
        |   ChatOpenAI (GPT-4o)       |   temperature=0
        +-----------------------------+
                     |
                     v
            +-------------------+
            |  Grounded answer  |
            +-------------------+

Tech Stack