Multi-Agent Document Platform

A Streamlit application that allows users to chat with pre-built document agents or create their own custom agents by simply uploading a PDF file. Each agent is equipped with a trained intent classifier in real-time to handle queries efficiently using a Retrieval-Augmented Generation (RAG) pipeline.

---

Key Features

• Multi-Agent Architecture: Start with pre-built agents (e.g., "gem5 Expert") and dynamically create new, custom agents for any PDF document.

• Real-time Agent Creation: Uploading a PDF automatically triggers a full machine learning pipeline that chunks the document, creates a vector database, synthetically generates a training dataset, and trains a unique intent classifier for that agent.

• Dynamic Intent Classification: Each custom agent uses its own LogisticRegression classifier to distinguish between casual conversation and document-specific questions (doc_qna), ensuring resources are used efficiently.

• RAG Pipeline: Leverages SentenceTransformer embeddings and a persistent ChromaDB vector store to retrieve the most relevant context from a document before passing it to the language model.

• Source Attribution: Responses for custom agents include the source page numbers, enhancing trust and verifiability.

---

Application Workflow

1. Agent Creation

When a user uploads a new PDF, a 5-step process (visible via status toasts in the UI) creates a fully functional agent:

1. Chunking: The PDF is parsed, and its text is split into manageable chunks. A new ChromaDB collection is created to store the embeddings for these chunks.
2. Data Generation: Keywords are extracted from the text chunks using YAKE. These keywords are used to synthetically generate a list of relevant, document-specific questions, which will form the doc_qna portion of our training set.
3. Dataset Assembly: The generated doc_qna questions are combined with a predefined list of casual questions from casual_queries.csv to form a training dataset.
4. Embedding Generation: All training queries are converted into numerical vector embeddings using the SentenceTransformer model.
5. Classifier Training: A LogisticRegression model is trained on the embeddings. The final trained classifier and its corresponding LabelEncoder are saved as .joblib files, ready for use.

2. Chatting with an Agent

1. Intent Classification: The user's query is first converted into an embedding and passed to the agent's unique, trained LogisticRegression model. The model classifies the intent as either doc_qna or casual.
2. Conditional Routing:
   • If casual, the RAG pipeline is skipped, and a polite, pre-defined response is given.
   • If doc_qna, the RAG pipeline is triggered.
3. Retrieval (RAG): The query embedding is used to search the agent's ChromaDB collection, retrieving the top 5 most relevant text chunks from the source document.
4. Generation (RAG): The retrieved chunks are formatted into a detailed prompt along with the original query. This is sent to the gemini-1.5-flash LLM, which generates an answer based strictly on the provided context.
5. Display: The final response, including source citations, is displayed in the chat interface.

---

Tech Stack

• Frontend: Streamlit
• Intent Classifier: Scikit-learn (LogisticRegression)
• Embedding Model: sentence-transformers/all-MiniLM-L6-v2
• LLM: Google Gemini 1.5 Flash (gemini-1.5-flash)
• Vector Database: ChromaDB
• Keyword Extraction: YAKE

---

Setup and Local Installation

To run this application on your local machine, follow these steps.

1. Clone the Repository

   git clone https://github.com/Anand-786/multi-doc-agent.git
   cd multi-doc-agent

2. Create and Activate a Virtual Environment

   python -m venv venv
   source venv/bin/activate  # On Windows, use `venv\Scripts\activate`

3. Install Dependencies

   pip install -r requirements.txt

4. Set Up Environment Variables Create a file named .env in the project's root directory and add your Google API key:

   GOOGLE_API_KEY="your_api_key_here"
   

5. Create the Casual Queries File Create a file named casual_queries.csv in the root directory with the entries like this:

   query,intent
   "Hello there!",casual
   "Hey, how are you?",casual
   "What's up?",casual
   "Thank you!",casual
   "Thanks a lot",casual
   "bye",casual

6. Run the Application

   streamlit run src/app_multi_agent.py