1. Project Overview
This project is a production-ready Conversational RAG (Retrieval-Augmented Generation) chatbot built with Streamlit. Users upload one or more PDF files, and the app allows them to ask questions about the content in a multi-turn conversation. The app maintains full chat history across turns so follow-up questions are understood in context.
1.1 Key Features
- Upload multiple PDFs and chat with their combined content
- Session-based memory — follow-up questions understand previous context automatically
- Groq LLaMA 3.3 70B as the LLM — fast and 100% free
- HuggingFace MiniLM embeddings — no OpenAI key needed
- ChromaDB persistent vector store — stored locally in ./chroma_db
- API key loaded from .env — no input box in the UI
- Clean Streamlit UI with expandable chat history viewer
1.2 Tech Stack
| Layer | Technology | Purpose |
|---|---|---|
| UI | Streamlit | Web interface for file upload and chat |
| LLM | Groq API — llama-3.3-70b-versatile | Answering questions from retrieved context |
| Embeddings | HuggingFace all-MiniLM-L6-v2 | Converting text chunks to vectors |
| Vector Store | ChromaDB (PersistentClient) | Storing and retrieving document embeddings |
| Memory | RunnableWithMessageHistory | Session-based multi-turn conversation memory |
| Document Loader | PyPDFLoader | Loading and parsing uploaded PDF files |
| Text Splitter | RecursiveCharacterTextSplitter | Chunking documents (5000 chars, 500 overlap) |
| Chain | create_history_aware_retriever | Reformulates questions using chat history |
| Chain | create_retrieval_chain | Combines retriever + LLM into full RAG pipeline |
| Config | python-dotenv | Loads GROQ_API_KEY and HF_TOKEN from .env |
2. Architecture
2.1 How RAG + Memory Works
When a user asks a question, the system first reformulates it as a standalone question using chat history (so follow-ups like ‘tell me more about that’ work). Then it retrieves relevant chunks from ChromaDB, passes them as context to Groq LLaMA, and returns a concise answer.
PDF Upload
↓ PyPDFLoader → RecursiveCharacterTextSplitter
↓ HuggingFace Embeddings → ChromaDB (local)
User Question + Chat History
↓ create_history_aware_retriever (Groq) → Standalone Question
↓ ChromaDB Retriever → Relevant Chunks
↓ create_stuff_documents_chain (Groq LLaMA 3.3 70B)
↓ Answer → Streamlit UI + ChatMessageHistory updated
2.2 LangChain Components
| Component | What It Does |
|---|---|
| create_history_aware_retriever | Takes the LLM + retriever + contextualize prompt. Rewrites user’s question into a standalone question if chat history exists. |
| create_stuff_documents_chain | Takes the LLM + QA prompt. Stuffs retrieved docs into the context window and generates an answer. |
| create_retrieval_chain | Combines the history-aware retriever and the document chain into a single invokable pipeline. |
| RunnableWithMessageHistory | Wraps the RAG chain. Automatically loads and saves ChatMessageHistory per session_id. |
| ChatMessageHistory | In-memory store of HumanMessage and AIMessage objects per session. Stored in st.session_state. |
3. Installation & Setup
3.1 Prerequisites
| Requirement | Details |
|---|---|
| Python | 3.10 or higher (tested on 3.11) |
| pip | Latest version recommended |
| Git | To clone the repository |
| Groq API Key | Free at console.groq.com — no credit card needed |
| HuggingFace Token | Free at huggingface.co/settings/tokens |
3.2 Installation Steps
Step 1 — Clone the repo
git clone https://github.com/shivamrawat2002/conversational-rag-pdf-chatbot/tree/main
cd Conversational-QnA-Chatbot
Step 2 — Create and activate virtual environment
python -m venv venv
venv\Scripts\activate # Windows
source venv/bin/activate # Mac/Linux
Step 3 — Install dependencies
pip install streamlit langchain-classic langchain-core langchain-community
pip install langchain-chroma langchain-groq langchain-huggingface
pip install langchain-text-splitters chromadb pypdf python-dotenv
Step 4 — Create .env file in project root
GROQ_API_KEY=gsk_your_groq_key_here
HF_TOKEN=hf_your_huggingface_token_here
Step 5 — Run the app
streamlit run app.py
Open http://localhost:8501 in your browser. Upload a PDF and start chatting!
4. Usage Guide
4.1 How to Use the App
| Step | Action |
|---|---|
| 1 | Open the app at http://localhost:8501 |
| 2 | Optionally change the Session ID field (default: ‘default_session’) |
| 3 | Click ‘Choose PDF file(s)’ and upload one or more PDFs |
| 4 | Wait for documents to be processed and embedded into ChromaDB |
| 5 | Type your question in the ‘Your question:’ input box and press Enter |
| 6 | Read the Assistant’s answer below |
| 7 | Click ‘Chat History’ expander to see the full conversation so far |
| 8 | Ask follow-up questions — the app remembers context automatically |
4.2 Session ID Explained
The Session ID field lets you run multiple independent conversations. Each session has its own isolated chat history. If you want a fresh conversation, change the Session ID. Useful for comparing answers across different research sessions.
4.3 Tips for Best Results
- Use specific questions rather than vague ones — e.g. ‘What are the key findings in section 3?’ works better than ‘Tell me about the document’
- Upload multiple related PDFs together — they are all embedded into the same ChromaDB collection
- If answers seem off, try re-uploading the PDF — this re-chunks and re-embeds the content
- Keep Breadth and Depth in mind — larger PDFs with more chunks give richer, more accurate answers
- The chat history expander shows both Human and AI messages for full transparency
5. Configuration & Customization
5.1 Changing the LLM Model
Edit app.py line 33 to use a different Groq model:
# Current (recommended):
llm = ChatGroq(groq_api_key=api_key, model_name=’llama-3.3-70b-versatile’)
# Other free Groq models:
# model_name=’llama-3.1-8b-instant’ (faster, smaller)
# model_name=’mixtral-8x7b-32768′ (older, still free)
5.2 Changing Chunk Size
Edit app.py to adjust how PDFs are split:
text_splitter = RecursiveCharacterTextSplitter(
chunk_size=5000, # characters per chunk (increase for more context)
chunk_overlap=500 # overlap between chunks (increase to reduce cut-offs)
)
5.3 Changing the Embedding Model
Edit app.py to use a different HuggingFace embedding model:
# Current (fast, lightweight):
embeddings = HuggingFaceEmbeddings(model_name=’all-MiniLM-L6-v2′)
# Higher quality alternative:
# embeddings = HuggingFaceEmbeddings(model_name=’all-mpnet-base-v2′)
6. Project File Structure
Conversational-QnA-Chatbot/
├── app.py       # Main Streamlit app — all logic lives here
├── .env        # API keys (GROQ_API_KEY, HF_TOKEN) — never commit!
├── requirements.txt  # Python dependencies
├── chroma_db/     # Auto-created — persistent ChromaDB vector store
├── temp.pdf      # Temporary file for uploaded PDF processing
└── venv/       # Python virtual environment
Demo Video
