Doc Chat Studio

A modern, intelligent document chat application built with Streamlit, LangChain, and FAISS for conversational Q&A over your documents using RAG (Retrieval-Augmented Generation).

🎯 Overview

Doc Chat Studio allows users to upload multiple documents (PDFs, DOCX, TXT, MD), automatically indexes them using FAISS vector embeddings, and enables natural language conversations with multi-step reasoning powered by OpenAI's GPT models.

✨ Key Features

• Multi-Document Upload: Support for PDF, DOCX, TXT, and Markdown files
• Intelligent Chunking: Uses RecursiveCharacterTextSplitter for semantic text splitting
• LangChain Retriever: FAISS-based VectorStoreRetriever for efficient similarity search
• Multi-Step Reasoning: Sequential chain (Summarize → Analyze → Answer) using LangChain
• Conversation Memory: Context-aware responses using ConversationBufferMemory
• Source Citations: Automatic citation of source documents in answers
• Modern UI: Centered, responsive design with custom CSS styling
• In-Memory Storage: FAISS vectorstore managed by LangChain

🏗️ Architecture

Tech Stack

• UI Framework: Streamlit
• LLM: OpenAI GPT-3.5-turbo
• Embeddings: HuggingFaceEmbeddings (all-MiniLM-L6-v2) via LangChain
• Vector Database: LangChain FAISS VectorStore (in-memory)
• Retriever: VectorStoreRetriever with similarity search
• Orchestration: LangChain
• Document Processing: PyPDF, python-docx

RAG Pipeline

1. Document Upload → Extract text from files
2. Text Chunking → Split into semantic chunks (1000 chars, 200 overlap)
3. Embedding → Convert chunks to 384-dim vectors using HuggingFaceEmbeddings
4. Indexing → Store in LangChain FAISS VectorStore
5. Query → Embed user question
6. Retrieval → VectorStoreRetriever finds top-4 similar chunks using .invoke()
7. Multi-Step Reasoning:
   • Step 1: Summarize retrieved context
   • Step 2: Analyze relevance to question
   • Step 3: Generate final answer with citations
8. Memory → Save Q&A pairs for context-aware follow-ups

📋 Core Functions

Memory & LLM Management

Function	Description
_get_conversation_memory()	Initialize/return ConversationBufferMemory for chat history
_get_langchain_llm()	Create/cache ChatOpenAI instance with API key

Prompt Engineering

Function	Description
build_prompt()	Build OpenAI-compatible messages from template and context
_create_sequential_chain()	Create 3-step LangChain SequentialChain (summarize → analyze → answer)

Document Processing

Function	Description
_read_file()	Extract text from PDF/DOCX/TXT/MD files
_chunk_text()	Split text using RecursiveCharacterTextSplitter (1000/200)
_add_docs()	Process files: extract → chunk → create LangChain Documents → add to vectorstore

Vector Search

Function	Description
_get_embeddings()	Lazy-load LangChain HuggingFaceEmbeddings (all-MiniLM-L6-v2)
_get_embedder()	Lazy-load SentenceTransformer model (backup for compatibility)
_get_retriever()	Create VectorStoreRetriever from FAISS vectorstore with top-k
_search_index()	Retrieve top-k documents using retriever.invoke() and format them

Answer Generation

Function	Description
_call_llm()	Generate answer using multi-step reasoning chain with memory
_generate_answer()	Main RAG pipeline: retrieve contexts → call LLM

UI Components

Function	Description
_inject_styles()	Apply custom CSS for modern, centered design
_init_state()	Initialize Streamlit session state variables
_render_header()	Display app title
_render_doc_pills()	Show uploaded document badges with size
_render_left_panel()	Document upload UI and file list
_render_chat_area()	Chat history display and input box
main()	App entry point: initialize state → render tabs

🚀 Setup & Installation

Prerequisites

• Python 3.8+
• OpenAI API key

Installation

# Clone repository
cd capstone/aiassistant

# Install dependencies
pip install -r requirements.txt

# Set OpenAI API key
# Windows PowerShell:
$env:OPENAI_API_KEY="your-key-here"

# Or create .env file:
echo "OPENAI_API_KEY=your-key-here" > .env

Run the App

streamlit run app.py

The app will open in your browser at http://localhost:8501

📦 Dependencies

streamlit>=1.37.0
faiss-cpu>=1.8.0
sentence-transformers>=3.0.0
pypdf>=4.2.0
python-docx>=1.1.0
openai>=1.42.0
langchain>=0.2.0
langchain-openai>=0.2.1
langchain-community>=0.2.0
langchain-classic
python-dotenv
numpy>=1.24.0

💡 Usage

Upload Documents

1. Navigate to the Documents tab
2. Click "Browse files" or drag-and-drop
3. Select PDF, DOCX, TXT, or MD files
4. Click "Add to workspace"
5. Wait for indexing to complete

Ask Questions

1. Navigate to the Chat tab
2. Type your question in the input box
3. Press Enter
4. The app will:
   • Retrieve relevant document chunks
   • Summarize context (Step 1/3)
   • Analyze relevance (Step 2/3)
   • Generate answer with citations (Step 3/3)

Follow-up Questions

The app maintains conversation history, so you can ask follow-up questions that reference previous answers:

You: What is the leave policy?
Bot: Employees get 15 days annual leave... [Source: policy.pdf]

You: Can I carry forward unused leave?
Bot: Based on our previous discussion about the leave policy...

🔧 Configuration

Chunk Size

Modify chunking parameters in _chunk_text():

chunk_size=1000,  # Characters per chunk
overlap=200       # Overlapping characters

Top-K Retrieval

Change number of retrieved chunks in _generate_answer():

contexts = _search_index(question, top_k=4)  # Retrieve 4 chunks

LLM Model

Update model in _get_langchain_llm():

model_name="gpt-3.5-turbo",  # or "gpt-4", "gpt-4-turbo"
temperature=0.3,              # Lower = more deterministic

🎨 UI Features

• Centered Layout: Max-width 880px for optimal readability
• Light Theme: Clean white background with subtle borders
• Card Design: Elevated surfaces with shadow effects
• Document Pills: Visual badges showing uploaded files
• Chat Bubbles: Distinct styling for user/assistant messages
• Progress Indicators: Real-time feedback during multi-step reasoning
• Timestamps: UTC timestamps for each message

⚠️ Limitations

Current Limitations

• In-Memory Only: FAISS index is not persisted to disk (resets on app restart)
• No Persistence: Uploaded documents and chat history lost on page refresh
• Session-Based: Each browser session maintains separate state
• No Multi-User Support: Not designed for concurrent users

Future Enhancements

• Persistent FAISS storage (save/load from disk)
• Document metadata persistence (SQLite/JSON)
• Chat history export/import
• Multi-user support with authentication
• Analytics dashboard (document stats, query patterns)
• Advanced filters (date range, document type)
• Visualization of embeddings (t-SNE/UMAP)

🐛 Troubleshooting

"No module named 'langchain.chains'"

pip install langchain>=0.2.0 langchain-classic --upgrade

"One input key expected got [...]"

Ensure ConversationBufferMemory has input_key="question":

memory = ConversationBufferMemory(
    memory_key="chat_history",
    input_key="question",  # Required for multiple inputs
    output_key="final_answer"
)

Context Not Being Used

Check that chat history is being formatted correctly:

# Format messages as string before passing to chain
chat_history_str = "\n".join([f"{msg.type}: {msg.content}"
                               for msg in memory.chat_memory.messages])

FAISS Not Available

The app automatically falls back to NumPy cosine similarity:

# NumPy fallback for similarity search
if FAISS_AVAILABLE:
    # Use FAISS
else:
    # Use NumPy cosine similarity

📄 License

MIT License

🤝 Contributing

Contributions welcome! Please open an issue or submit a pull request.

📧 Contact

For questions or feedback, please open an issue in the repository.

---

Built with ❤️ using Streamlit, LangChain, and OpenAI