ResearchGen AI

A multi-agent research assistant that lets you upload PDF papers, ask complex questions, and receive structured, evidence-grounded answers powered by RAG, LangGraph, and Google Gemini.

---

Features

• PDF upload & indexing — ChromaDB vector store with table-aware chunking
• Smart routing — SIMPLE questions get fast answers; ADVANCED questions use the full agent pipeline
• Multi-agent workflow — Researcher → Critic → Writer with retry loops
• Multi-query retrieval — Decomposes questions into theory / architecture / experiments searches
• Grounded critics — Requirement-aware evaluation with checklist (experimental, architectural, theoretical evidence)
• Second-pass QA — Final answer is reviewed before delivery; one revision if needed
• Chat UI — Streamlit interface with conversation history

---

Architecture

Upload PDFs → Ingest (chunk + embed) → ChromaDB

User question
    → Classify (SIMPLE / ADVANCED)
    → [ADVANCED]
        → Detect requirements
        → Multi-query retrieve
        → Researcher (evidence notes)
        → Critic (notes QA, retry up to 2×)
        → Writer (structured answer)
        → Final critic (answer QA, revise up to 1×)
    → [SIMPLE] → Direct short answer

Tech stack

Component	Technology
UI	Streamlit
Orchestration	LangGraph
LLM	Google Gemini 2.5 Flash
Embeddings	HuggingFace all-MiniLM-L6-v2
Vector DB	ChromaDB
PDF parsing	PyPDF

---

Project structure

Research-agent/
├── app.py                 # Streamlit entry point
├── agents/
│   ├── researcher.py      # Extracts evidence-backed notes
│   ├── critic.py          # Research + final answer evaluation
│   └── writer.py          # Produces user-facing answer
├── graph/
│   └── workflow.py        # LangGraph pipeline
├── rag/
│   ├── ingest.py          # PDF indexing (table-aware chunks)
│   ├── retriever.py       # MMR retrieval
│   ├── query_decompose.py # Sub-query generation
│   └── multi_retrieve.py  # Multi-angle retrieval
├── utils/
│   ├── llm.py             # Gemini client
│   ├── schemas.py         # Structured outputs (Pydantic)
│   └── requirements.py    # Question requirement detection
├── requirements.txt
├── .env.example
└── .gitignore

---

Prerequisites

• Python 3.10 or newer
• Google Gemini API key

---

Setup

1. Clone the repository

git clone https://github.com/SEFIN-official/researchgen-ai.git
cd researchgen-ai

2. Create and activate a virtual environment

Windows (PowerShell):

python -m venv venv
venv\Scripts\activate

macOS / Linux:

python -m venv venv
source venv/bin/activate

3. Install dependencies

pip install -r requirements.txt

4. Configure environment variables

Copy the example env file and add your API key:

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

Edit .env:

GOOGLE_API_KEY=your_api_key_here

---

Run the app

streamlit run app.py

Open the URL shown in the terminal (usually http://localhost:8501).

---

Usage

1. Upload one or more PDF research papers.
2. Click Upload & Index and wait for the success message.
3. Ask a question in the chat box.

Example flow

Step	Action
1	Upload a paper (e.g. Transformer / ML paper)
2	Index documents
3	Ask a complex question

Example question:

Compare the model's theoretical advantages, architectural components, and experimental results (including tables and metrics) using evidence from the uploaded paper.

4. Review the structured multi-agent answer with theory, architecture, experiments, and synthesis.

Question types

Type	Example	Pipeline
SIMPLE	"What is gradient descent?"	Fast direct answer
ADVANCED	"Compare Adam vs SGD using my paper"	Full RAG + multi-agent

---

Configuration (optional)

Setting	Location	Default
Retrieval chunk count	rag/retriever.py	DEFAULT_K = 8
Max research retries	graph/workflow.py	MAX_RETRIES = 2
Max final answer revisions	graph/workflow.py	MAX_FINAL_RETRIES = 1
LLM model	utils/llm.py	gemini-2.5-flash

---

Re-indexing after code updates

If you change ingestion logic, re-upload and re-index your PDFs so ChromaDB uses the latest chunking strategy.

---

Troubleshooting

Issue	Fix
API key error	Set GOOGLE_API_KEY in .env and restart Streamlit
Generic answers	Confirm PDFs were indexed; ask ADVANCED-style questions
Quota exceeded	Wait and retry; check Gemini API limits
Missing table data	Re-index PDFs; tables depend on PDF text extraction quality

---

Limitations

• Answers are only as good as retrieved PDF chunks.
• PDF tables may not parse perfectly as structured data.
• ADVANCED questions use more API calls (higher latency and cost).

---

License

MIT (or add your preferred license).

---

Acknowledgments

Built with LangChain, LangGraph, Streamlit, ChromaDB, and Google Gemini.