Multi-Agent Research Assistant

An AI-powered multi-agent system that automatically analyzes academic research papers using Retrieval-Augmented Generation (RAG), hybrid search, and specialized AI agents.

The system extracts information from research papers, retrieves relevant context, generates structured summaries, critiques research quality, and produces comprehensive reports.

---

🚀 Features

• 📄 Upload research papers (PDF)
• 🔍 Hybrid Retrieval (Dense Embeddings + BM25)
• 🧠 Multi-Agent Architecture
• 📑 Automatic metadata extraction
• 📖 OCR support for scanned PDFs
• ✂️ Section-aware document chunking
• 🚫 Reference page filtering
• 📂 Parent-child chunking strategy
• 🤖 Gemini with automatic Groq fallback
• 📝 Structured research report generation
• ⚡ FastAPI REST APIs
• 🗄️ ChromaDB Vector Database

---

System Architecture

                                User Uploads PDF
                                        │
                                        ▼
                              PDF Processing Service
                                        │
          ┌─────────────────────────────┼─────────────────────────────┐
          │                             │                             │
          ▼                             ▼                             ▼
     PyMuPDF Loader               OCR Fallback             Metadata Extraction
                                                      (Title, Authors, Abstract)
          │
          ▼
   Reference Page Filtering
          │
          ▼
     Section Detection
          │
          ▼
 Parent → Child Chunking
          │
          ▼
 Metadata Enrichment
          │
          ▼
      ChromaDB Vector Store
          │
      ┌───┴─────────────┐
      ▼                 ▼
 Dense Retrieval     BM25 Retrieval
      │                 │
      └────── Hybrid Retrieval ───────┘
                      │
                      ▼
            Parent Context Expansion
                      │
                      ▼
                Multi-Agent Pipeline
                      │
      ┌─────────────────────────────────────────────┐
      │                                             │
      ▼                                             ▼
 Planner Agent                               Search Agent
      │                                             │
      ▼                                             ▼
 Research Agent ----------------------------► Review.py
      │                                             │
      ▼                                             ▼
 Summarizer Agent --------------------------► Critic Agent
      │
      ▼
 Writer Agent
      │
      ▼
 Final Research Report

---

Project Structure

multi_agent_research_assistant/
│
├── agents/                # AI agents
│   ├── planner_agent.py
│   ├── search_agent.py
│   ├── research_agent.py
│   ├── summarizer_agent.py
│   ├── writer_agent.py
│   ├── critic_agent.py
│   └── review_agent.py
│
├── api/                   # FastAPI endpoints
│
├── core/                  # Configuration & utilities
│
├── database/              # Database operations
│
├── orchestration/         # Multi-agent workflow
│
├── prompts/               # Agent prompts
│
├── reports/               # Generated reports
│
├── schemas/               # Pydantic schemas
│
├── services/
│   ├── rag_service.py
│   ├── pdf_processor.py
│   └── embedding_service.py
│
├── tests/
│
├── tools/
│
├── uploads/
│
├── vectorstore/
│
├── logs/
│
├── .env
├── Dockerfile
├── requirements.txt
└── main.py

---

Workflow

1. Upload PDF

The user uploads a research paper through the FastAPI endpoint.

↓

2. PDF Processing

The system

• extracts text using PyMuPDF
• performs OCR if necessary
• extracts metadata
• removes reference pages
• identifies document sections

↓

3. Chunking

The processed document is converted into

• parent documents
• child chunks

Each chunk is enriched with metadata.

↓

4. Indexing

The chunks are stored inside ChromaDB while BM25 indexes are built for lexical retrieval.

↓

5. Query Processing

For every user query the system

• checks metadata cache
• performs hybrid retrieval
• expands parent context
• prepares context for the LLM

↓

6. Multi-Agent Analysis

The retrieved context flows through specialized agents.

Planner

↓

Search

↓

Research

↓

Summarizer

↓

Critic

↓

Writer

↓

Final Report

---

AI Agents

Planner Agent

Plans the research workflow and determines the execution order.

---

Search Agent

Retrieves relevant document chunks using hybrid retrieval.

---

Research Agent

Analyzes retrieved content and extracts key information.

---

Summarizer Agent

Produces structured summaries and highlights important findings.

---

Critic Agent

Evaluates limitations, weaknesses, and missing information.

---

Writer Agent

Combines outputs from all agents into a structured research report.

---

Review Agent

Performs quality validation before generating the final response.

---

Tech Stack

Backend

• FastAPI
• Python

LLMs

• Google Gemini
• Groq (Fallback)

Retrieval

• LangChain
• ChromaDB
• BM25
• Sentence Transformers

PDF Processing

• PyMuPDF
• OCR

AI

• Retrieval-Augmented Generation (RAG)
• Hybrid Search
• Multi-Agent Systems

---

Current Features

✅ PDF Upload

✅ OCR Support

✅ Metadata Extraction

✅ Hybrid Retrieval

✅ ChromaDB Integration

✅ BM25 Search

✅ Parent-Child Chunking

✅ Reference Filtering

✅ Multi-Agent Workflow

✅ Structured Report Generation

✅ Gemini → Groq Fallback

---

Future Improvements

• Query routing
• Multi-document analysis
• Citation-aware retrieval
• Agent memory
• Streaming responses
• Web research integration
• Research paper comparison
• Knowledge graph generation

---

Author

Prerna Mishra