DocQuery: Retrieval-Augmented Generation Pipeline

This project implements a Retrieval-Augmented Generation (RAG) pipeline for querying unstructured PDF documents.
It combines embeddings, vector search, and a large language model to return context-aware answers in real time. Note: Right now data is limited, in future data gets updated or you can

📊 Workflow

🚀 Features

• Document Ingestion (core/data_loader.py): Load and chunk PDF documents.
• Embeddings (core/embedding_manager.py): Generate 384-dim sentence embeddings with all-MiniLM-L6-v2.
• Vector Store (core/vector_store.py): Store and search embeddings using ChromaDB (HNSW indexing).
• Retriever (core/retriever.py): Fetch relevant context for queries.
• Pipeline (pipelines/rag_pipeline.py): Combine retriever + LLM (Google’s gemma2-9b-it) for RAG responses.
• Streamlit UI (main.py): Simple and interactive interface for querying documents.
• Configurable (config.py): Centralized settings for model, database, and pipeline options.
• Experiments (notebooks/rag_pipeline.ipynb).

⚙️ Setup

This project uses uv for Python package management.
Make sure you have uv installed first:

pip install uv

Clone the repo and install dependencies:

git clone https://github.com/<your-username>/<repo-name>.git
cd <repo-name>
uv sync

▶️ Usage

Build the databse (this is a onetime setup):

• Upload PDFs to the data/pdf_files path
• Then run this command

python main.py --build

API Setup:

• Get your API key to the gemma2-9b-it model from here groq-api-keys.
• Create a .env file in your project root path and assign your API key to GROQ_API_KEY.

Start the Streamlit app in local:

streamlit run app.py

or click on this link to view the app DocQuery

Type your query, and get context-aware answers.

📂 Project Structure

.
├── core/                    # Core components
│   ├── data_loader.py       # PDF loading + chunking
│   ├── embedding_manager.py # Embedding generation
│   ├── retriever.py         # Context retrieval
│   └── vector_store.py      # ChromaDB integration
│
├── data/                    # Input and storage
│   ├── pdf_files/           # Source documents
│   └── vector_store/        # Persisted ChromaDB index
│
├── notebooks/
│   └── rag_pipeline.ipynb   # Experiments & benchmarks
│
├── pipelines/
│   └── rag_pipeline.py      # Full RAG pipeline logic
│
├── config.py                # Global configs
├── main.py                  # Streamlit entry point
├── pyproject.toml           # uv dependencies
├── requirements.txt         # pip fallback
├── uv.lock                  # uv lock file
├── .gitignore
└── README.md

Future work

• Benchmark the retrieval strategies

Reference

• https://www.youtube.com/watch?v=fZM3oX4xEyg&list=PLZoTAELRMXVM8Pf4U67L4UuDRgV4TNX9D
• https://www.singlestore.com/blog/a-guide-to-retrieval-augmented-generation-rag/
• https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2
• https://python.langchain.com/docs/introduction/
• https://console.groq.com/docs/