PDF Chatbot 🤖

An intelligent PDF document Q&A application using RAG (Retrieval-Augmented Generation) with LangChain and Groq API.

Features

✨ Smart Document Analysis

• Upload and analyze PDF files (CVs, reports, documents, etc.)
• Ask natural language questions about your documents
• Get accurate, context-aware responses

🔍 Advanced Retrieval

• Semantic search using embeddings
• FAISS vector store for efficient similarity search
• Multi-document chunk retrieval for comprehensive context

🚀 Fast Performance

• Groq API for lightning-fast LLM inference
• HuggingFace embeddings for semantic understanding
• Real-time streaming responses

💬 Conversation History

• Maintains chat history within a session
• Automatic reset when uploading a new document
• File-aware conversation management

Architecture

PDF Upload
    ↓
[Text Extraction] → PyPDF2
    ↓
[Chunking] → RecursiveCharacterTextSplitter (1000 chars, 200 overlap)
    ↓
[Embeddings] → HuggingFace "all-MiniLM-L6-v2"
    ↓
[Vector Store] → FAISS (4 top results per query)
    ↓
[RAG Chain] → Retriever | Prompt Template | Groq LLM
    ↓
Answer

Installation

Prerequisites

• Python 3.10+
• Groq API key (get free at https://groq.com)

Setup

# 1. Create virtual environment
python -m venv ml_env
source ml_env/bin/activate  # On Windows: ml_env\Scripts\activate

# 2. Install dependencies
pip install -r requirements.txt

# 3. Create .env file
echo "api_key=your_groq_api_key_here" > .env

Usage

# Activate environment
source ml_env/bin/activate

# Run the application
streamlit run app.py

The app will open at http://localhost:8501

How to Use

1. Enter API Key (optional if in .env)
2. Upload PDF via the sidebar
3. Ask Questions about the document content
4. Get Instant Answers based on document context

Configuration

Environment Variables (.env)

api_key=your_groq_api_key_here

Model Parameters

Parameter	Value	Description
Chunk Size	1000	Characters per text chunk
Chunk Overlap	200	Overlap between chunks for context continuity
Retrieval K	4	Number of chunks retrieved per query
Embedding Model	all-MiniLM-L6-v2	384-dim, lightweight embeddings
LLM	llama-3.3-70b	Groq's fastest available model

Project Structure

pdf_chatbot/
├── app.py              # Main Streamlit application
├── requirements.txt    # Python dependencies
├── .env               # API keys (git-ignored)
├── ml_env/           # Virtual environment
└── README.md         # This file

Dependencies

streamlit              # Web UI framework
langchain             # LLM orchestration
langchain-core        # Core abstractions
langchain-groq        # Groq API integration
langchain-text-splitters  # Document splitting
langchain-community   # Community integrations
sentence-transformers # Semantic embeddings
faiss-cpu            # Vector similarity search
PyPDF2               # PDF text extraction
python-dotenv        # Environment variable loading

Key Components

extract_text(pdf_file)

Extracts text from PDF using PyPDF2.

build_chain(text, api_key)

Creates RAG chain:

• Splits document into chunks
• Generates embeddings
• Builds FAISS vector store
• Creates LCEL chain combining retrieval + LLM

Session State Management

• messages: Chat history
• chain: RAG pipeline instance
• retriever: FAISS retriever
• current_file: Tracks loaded file to avoid reprocessing

Performance Tips

⚡ Optimization

• Uses CPU FAISS (install faiss-gpu for GPU acceleration)
• HuggingFace embeddings are cached locally
• Groq API provides sub-second inference

🎯 Best Practices

• Keep chunk size between 500-2000 for balance
• Use k=3-5 for retrieval to balance relevance/scope
• Small PDFs (<500KB) work best for instant processing

Troubleshooting

"No module named 'langchain.chains'"

Your installed LangChain version is incompatible. Update:

pip install --upgrade langchain langchain-core langchain-groq

"Input to ChatPromptTemplate is missing variables"

The prompt template expects specific variables. Ensure build_chain() passes correct keys.

Slow responses

• Reduce chunk_size to speed up embedding
• Reduce k (retrieval count) from 4 to 2
• Use faiss-gpu for GPU acceleration

Empty PDF extraction

Some PDFs are image-only. They need OCR (not supported in this version).

Advanced Customization

Change LLM Model

In build_chain():

llm = ChatGroq(groq_api_key=api_key, model_name="mixtral-8x7b-32768")

Adjust Chunk Strategy

splitter = RecursiveCharacterTextSplitter(
    chunk_size=2000,  # Larger chunks
    chunk_overlap=500  # More overlap
)

Custom System Prompt

Modify the system_prompt variable in build_chain() to change LLM behavior.

Limitations

⚠️ Known Limitations

• Cannot process image-based PDFs (OCR not implemented)
• Limited to 500MB+ file sizes depending on memory
• No support for multi-language documents (English optimized)
• Chat history resets on application restart

Future Enhancements

📚 Planned Features

• Multi-document Q&A
• Persistent chat history (database)
• OCR for image-based PDFs
• Citation tracking (show source pages)
• Custom prompt templates UI
• Batch document processing
• Export chat to PDF

License

MIT License - Feel free to modify and share!

Support

• Groq API: https://groq.com/documentation
• LangChain: https://python.langchain.com/
• Streamlit: https://docs.streamlit.io/

---

Created with ❤️ using LangChain, Streamlit, and Groq API