📚 Paper‑RAG Assistant

A research‑oriented Retrieval‑Augmented Generation (RAG) system that lets you search arXiv, ingest local PDFs, and ask conversational questions over your personalised paper collection. It pairs LangGraph state machines with Google Gemini models, Chroma DB vector storage and a lightweight MySQL catalogue, giving you a batteries‑included pipeline from raw PDF to high‑level insight.

---

✨ Key Features

Capability	Description
Multi‑source ingestion	‑ Search and persist top‑k arXiv papers.‑ Drop any PDF (single file or folder) and auto‑extract title + abstract via heuristics.
Structured metadata store	All papers are catalogued into a MySQL table and simultaneously embedded into a Chroma collection (papers).
LangGraph orchestration	An explicit state machine routes user queries to one of five intents: search_arxiv, upload_pdf, local_query, compare, other.
Google Gemini integration	• models/gemini‑2.5‑flash‑preview‑04‑17 for chat tasks.• models/embedding‑001 for text embeddings.
CLI	Use the built‑in terminal loop.
Deterministic persistence	Vector store is stored on disk (./chroma_db) so nothing is lost between sessions.

---

---

🛠️ Prerequisites

Install Python dependencies:

pip install \
  langchain langgraph langchain-community chromadb arxiv \
  mysql-connector-python python-dotenv google-generativeai pdfminer.six \
  tiktoken pypdf streamlit

langchain==0.1.17
langgraph==0.0.26
chromadb==0.5.0
arxiv==2.1.0
mysql-connector-python==8.4.0
python-dotenv==1.0.1
google-generativeai==0.5.1
pdfminer.six==20221105
pypdf==4.2.0
tiktoken==0.6.0

---

🔐 Environment Variables

Create a .env in the project root (or export vars in your shell):

# Google Gemini
GEMINI_API_KEY="<your‑key>"
GEMINI_MODEL="models/gemini-2.5-flash-preview-04-17"  # optional override

# MySQL connection
MYSQL_HOST=localhost
MYSQL_PORT=3306
MYSQL_USER=llmuser
MYSQL_PASSWORD=LLMproject_
MYSQL_DB=paper_db

# Chroma persistence directory
CHROMA_PERSIST_DIR=./chroma_db

Tip: The script will auto‑create the paper_db database if it does not exist (given sufficient privileges).

---

🗄️ MySQL Schema

The table is created automatically on first run, shown here for reference:

CREATE TABLE papers (
    id INT AUTO_INCREMENT PRIMARY KEY,
    title TEXT,
    abstract TEXT,
);

---

🚀 Quick Start (CLI)

python main.py

Then interact:

Welcome to the paper‑RAG assistant. Type 'exit' to quit.
>>> upload pdf
📄 Please enter a full PDF file path, or a directory containing PDFs:
>> PDF file or folder path: ./pdf
📑 Processing: ./pdf/SiLK- Simple Learned Keypoints.pdf...
✅ Stored PDF as paper ID 1: SiLK: Simple Learned Keypoints Pierre Gleize
...

>>> Compare dedode and dedodev2
💬 **Comparison Report: DeDoDe vs. DeDoDe v2**
**Problem Statement**
*   **DeDoDe:** Addresses the challenge of keypoint detection for 3D reconstruction, specifically the ....
...

The CLI keeps only the last two turns in memory, making it suitable for long sessions without hitting token limits.

---

🖼️ Architecture Overview

┌───────────────────────────────┐
│           User CLI            │
└──────────────┬────────────────┘
               │  user_input
        ┌──────▼──────┐ intent
        │  Router     │─────────┐
        └──────┬──────┘         │
  search_arxiv │     upload_pdf │ compare/local/other
        ┌──────▼──────┐         │
        │  arXiv API  │         │
        └──────┬──────┘         │
               │ meta & text    │
        ┌──────▼──────┐         │
        │ MySQL +     │         │
        │  Chroma     │◄────────┘
        └──────┬──────┘
               │ RAG context
        ┌──────▼──────┐
        │  Gemini LLM │
        └─────────────┘

• StateGraph – provides explicit edges between nodes for traceability.
• Google Gemini – used both for intent routing and final answer generation.
• Chroma DB – fast cosine similarity search on embeddings.
• MySQL – relational metadata store (source, ids, timestamps).

---

🏃‍♂️ Common Workflows

Task	Command	Notes
Search latest papers on Arxiv	>>> Arxiv search ...	Top‑3 arXiv hits are persisted automatically
Ingest a folder of local PDFs	Choose upload_pdf intent or type upload	Titles & abstracts are auto‑extracted; fallback to file‑name if OCR fails
Ask a question about stored papers	>>> What are the main limitations of current 3D diffusion models?	If nothing found, system will state so
Compare approaches	>>> compare How do SCORE‑based vs DDPM methods differ?	Produces a 5‑bullet structured report

---

🙏 Acknowledgements

• LangChain + LangGraph
• ChromaDB
• Google Gemini & GenerativeAI APIs
• The amazing arXiv open access repository