📄 Ask Your Docs — PDF Q&A Tool

Upload a PDF and ask plain-language questions against it. Answers are grounded in the document with source citations. This project was created to practice building RAG pipelines and evaluations.

---

Prerequisites

• Python 3.11–3.13
• uv
• A Pinecone account (free tier works)
• One of: an OpenAI API key (cloud) or Ollama running locally

---

Setup

1. Clone the repo and install dependencies

git clone <repo-url>
cd ask-your-docs
uv sync

2. Create a .env file in the project root

# Pinecone (required)
PINECONE_API_KEY=your_pinecone_api_key
PINECONE_INDEX=your_index_name
PINECONE_MODEL=multilingual-e5-large

# LLM — pick one of the two options below

# Option A: OpenAI (cloud)
OPENAI_API_KEY=your_openai_api_key
OPENAI_MODEL=gpt-4o-mini

# Option B: Ollama (local, used only if OPENAI_API_KEY is not set)
OLLAMA_MODEL=llama3

3. If using Ollama, pull your model

ollama pull llama3

4. Add your PDF

Place the PDF you want to query inside the textbooks/ folder.

---

Running the App

First run — ingest the PDF and start Q&A:

uv run ask-your-docs --ingest

Subsequent runs — PDF already indexed, skip ingestion:

uv run ask-your-docs

You'll see an interactive prompt:

RAG pipeline ready. Type 'exit' to quit.

Ask a question: What is the purpose of the TLB?
> The TLB (Translation Lookaside Buffer) is a cache for virtual-to-physical
  address translations...

Type exit to quit.

---

Running the Evaluation

Runs 150 questions from the golden dataset through the full pipeline and reports MRR, nDCG, and keyword coverage:

uv run run-evaluation

Example output:

========================================
Overall Results
========================================
  MRR:               0.6123
  nDCG:              0.5891
  Keyword Coverage:  0.7340
========================================

Results are saved to evaluation/results/.