Multi-Agent-Research-Assistant

🤖 Multi-Agent Research Assistant

An enterprise-grade, production-ready Multi-Agent Research Assistant built with LangGraph, LangChain, Ollama (llama3.1), FAISS, SQLite, Streamlit, and Tavily Web Search API.

It combines document-based RAG, real-time web search, intelligent query routing, answer verification, citations, and conversational memory to deliver context-aware research with PDF, DOCX, and Markdown export.

Unlike simple RAG chatbots, this application employs a Planner Agent to dynamically analyze query intent, uploaded document availability, retrieval confidence scores, and freshness requirements to route queries along optimal agent execution paths.

---

🌟 Key Features

• 8 Autonomous Specialized Agents:
  1. 🧠 Planner Agent: Analyzes user intent, document presence, and freshness to dynamically route execution.
  2. 📚 Retriever Agent: Parses PDF, DOCX, and TXT files, splits text into chunks, generates embeddings with HuggingFace (sentence-transformers/all-MiniLM-L6-v2), and indexes into FAISS vector store with normalized confidence scores.
  3. 🌐 Web Search Agent: Executes real-time web searches using Tavily API when external or live data is required.
  4. 💡 Reasoning Agent: Synthesizes facts across uploaded documents, web search snippets, and LLM reasoning into grounded answers.
  5. 🔍 Verifier Agent: Validates retrieval confidence, checks grounding, detects document vs. web search conflicts, and flags outdated documents.
  6. ✍️ Summarizer Agent: Formats responses into selectable styles (Brief, Detailed, Bullet Points, ELI5, Professional, Research).
  7. 📌 Citation Agent: Automatically formats and attaches exact inline and footer citations for uploaded files (filename & page number) and web URLs.
  8. 🧠 Memory Agent: Manages SQLite conversation history and reformulates follow-up queries for context awareness.
• Dynamic Execution Routing:
  • llm_only: General knowledge, programming, or coding questions.
  • rag_only: Questions strictly pertaining to uploaded documents.
  • web_search: Latest/current news, real-time events, external facts.
  • hybrid_rag_llm: Related to uploaded documents, but augmented with LLM reasoning when document facts are partial.
  • hybrid_rag_web: Comparing uploaded document facts with latest live web research.
• Enterprise Streamlit Interface:
  • Dark mode glassmorphic UI.
  • Dynamic route badges & retrieval confidence percentage indicators.
  • Expandable agent execution steps.
  • Multiple document uploader & vector DB status dashboard.
  • Export chat history as PDF, DOCX, or Markdown.

---

🏗️ Architecture & Workflow

graph TD
    User([User Input]) --> App[Streamlit UI]
    App --> Memory[Memory Agent: Follow-up Reformulation]
    Memory --> Planner[Planner Agent: Routing Logic]

    Planner -- LLM Direct --> Reasoning[Reasoning Agent]
    Planner -- Document Related --> Retriever[Retriever Agent: FAISS Index]
    Planner -- Fresh Info --> WebSearch[Web Search Agent: Tavily]
    Planner -- Hybrid RAG+Web --> Retriever & WebSearch

    Retriever --> Verifier[Verifier Agent: Grounding & Conflict Check]
    WebSearch --> Verifier

    Verifier --> Reasoning
    Reasoning --> Summarizer[Summarizer Agent: Response Style]
    Summarizer --> Citation[Citation Agent: Sources & References]
    Citation --> App
    Citation --> DB[(SQLite Database)]

---

📁 Directory Structure

multi-agent-research-assistant/
│
├── app.py                      # Main Streamlit UI dashboard
├── graph.py                    # LangGraph StateGraph orchestration
├── config.py                   # Centralized configuration & environment loader
├── database.py                 # SQLite database for persistent history & metadata
├── prompts.py                  # System & user prompts for all 8 agents
├── utils.py                    # Logging setup and PDF/DOCX/MD export utilities
│
├── agents/                     # Modular Agent Implementations
│   ├── __init__.py
│   ├── planner.py              # Query intent & routing agent
│   ├── retriever.py            # PDF/DOCX/TXT loader & FAISS vector search agent
│   ├── web_search.py           # Tavily web search agent
│   ├── reasoning.py            # Multi-source answer synthesis agent
│   ├── verifier.py             # Groundedness & conflict verification agent
│   ├── summarizer.py           # Multi-style formatting agent
│   ├── citation.py             # Citation extraction agent
│   └── memory.py               # History & follow-up reformulation agent
│
├── uploads/                    # Directory for uploaded documents
├── vector_store/               # FAISS vector store index directory
├── database/                   # SQLite database location
├── logs/                       # Application logs directory
├── sample_docs/                # Test documents for quick RAG testing
├── requirements.txt            # Python dependencies
├── README.md                   # Documentation
└── .env.example                # Sample environment variables

---

⚡ Quick Start & Installation

1. Prerequisites

• Python 3.10+
• (Optional) Ollama with llama3.1 pulled:

  ollama pull llama3.1
  

2. Create & Activate Virtual Environment

Windows (PowerShell):

python -m venv .venv
.\.venv\Scripts\Activate.ps1

Windows (CMD):

python -m venv .venv
.\.venv\Scripts\Activate.ps1

Linux / macOS:

python -m venv .venv
source .venv/bin/activate

3. Install Dependencies

pip install -r requirements.txt

Note: Always run from inside the activated .venv environment. If you see ModuleNotFoundError, ensure .venv is activated or use the full path to .venv\Scripts\python.exe -m streamlit run app.py.

4. Configure Environment Variables

Copy .env.example to .env:

cp .env.example .env   # Linux/macOS
copy .env.example .env  # Windows

Update .env with your Tavily API Key (or OpenAI key if using fallback):

OLLAMA_BASE_URL=http://localhost:11434
OLLAMA_MODEL=llama3.1
TAVILY_API_KEY=your_tavily_api_key_here

5. Run the Streamlit Application

After activating .venv:

Or directly with venv Python (no activation needed):

# Windows
.\.venv\Scripts\python.exe -m streamlit run app.py

# Linux/macOS
.venv/bin/python -m streamlit run app.py

---

🧪 Testing the Assistant

1. General Knowledge / Coding (LLM Only):
   • Ask: "What is LangChain?" or "Write a Python function to quicksort an array."
   • Route Badge: ROUTE: LLM_ONLY
2. RAG Document Search:
   • Upload sample_docs/sample_research.txt in the sidebar and click Process & Index Documents.
   • Ask: "What is Project Titan and what is its latency benchmark?"
   • Route Badge: ROUTE: RAG_ONLY or ROUTE: HYBRID_RAG_LLM with document citations.
3. Web Search:
   • Ask: "What are the latest updates on GPT-5?"
   • Route Badge: ROUTE: WEB_SEARCH with web source links.
4. Follow-up Conversation:
   • Ask: "Explain LangGraph."
   • Follow up: "Compare it with CrewAI."
   • Memory Agent automatically reformulates query context.
5. Conversation Export:
   • Click PDF, DOCX, or MD buttons in the sidebar to download conversation logs.

---

📝 License

MIT License. Built for enterprise Agentic AI development.