Chatbot Project (RAG over PDFs)

A simple Retrieval-Augmented Generation (RAG) chatbot that lets you ask questions about a PDF. It:

• Loads a PDF, splits it into chunks
• Creates embeddings with OpenAI
• Stores / retrieves chunks from a local ChromaDB vector store
• Uses an LLM (OpenAI chat model by default) to answer using retrieved context
• Exposes a FastAPI endpoint with per-session conversation history

Project structure

• src/api/server.py — FastAPI server (GET /question-answering)
• src/document_loader/loader.py — PDF ingestion script (load → split → embed → persist to Chroma)
• src/question_answering/qa_system.py — retrieval + prompt building + LLM invocation
• src/vector_store/embeddings.py — helper for adding/searching/listing embeddings in ChromaDB
• src/retrieval/retriever.py — lower-level retrieval helper (debug/test)
• src/text_processing/splitter.py — text chunking helper

Requirements

• Python 3.10+ recommended
• An OpenAI API key

Install dependencies:

python -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate
pip install -r requirements.txt

Configuration

Create a .env file in the repo root:

OPENAI_API_KEY=your_key_here
# Optional
GPT_MODEL=gpt-4

1) Ingest a PDF (build the vector store)

By default, the ingestion script is hard-coded to use:

• data/WhitepaperServiceMeshAndAPIManagement.pdf

Run ingestion (from the repo root):

python -m src.document_loader.loader

This will create / update the persisted vector store under vector_store_db/.

Note: The current src/document_loader/loader.py uses Chroma.from_documents(...) which creates a Chroma collection. The rest of the code expects a collection named document_embeddings. If you see retrieval issues, ensure the same collection name is being used consistently.

2) Run the API

Start the FastAPI server:

It runs on port 8000 by default.

Query the chatbot

Example request:

curl -G "http://localhost:8000/question-answering" \
  --data-urlencode "query=What is a service mesh?" \
  --data-urlencode "session_id=demo"

Response includes:

• answer — the model's answer
• history — the session conversation history stored in-memory

Notes / limitations

• Conversation history is stored in memory only (lost on restart).
• vector_store_db/ is currently committed to the repo; you may want to add it to .gitignore for most workflows.
• A local LLM option via LlamaCpp exists in src/question_answering/qa_system.py but is commented out.

License

Add a license file if you plan to distribute this project.