AI Research Paper Analyzer — Backend

A learning project to explore RAG (Retrieval-Augmented Generation), vector databases, and the building blocks of single-agent AI systems — using research papers as the data source.

---

Architecture

Client
  │
  └─ sends PDF
        │
        ▼
  FastAPI endpoint
        │
        ▼
  chunked into text segments
        │
        ▼
  vector embedded (sentence-transformers)
        │
        ▼
  stored in ChromaDB

  Query
  │
  └─ vector embedded
        │
        ▼
  ChromaDB similarity search
        │
        ▼
  top chunks → RAG prompt → Gemini
        │
        ▼
  answer returned to client

---

Stack

Layer	Technology
API	FastAPI
ORM / Migrations	SQLAlchemy + Alembic
Database	SQLite
PDF Parsing	pypdf
Embeddings	sentence-transformers (all-MiniLM-L6-v2)
Vector Store	ChromaDB
LLM	Google Gemini 2.5 Flash

---

Data Flow

Upload

1. PDF uploaded → text extracted page-by-page
2. Text split into 500-character chunks
3. Paper metadata and chunks saved to SQLite
4. Each chunk embedded → vectors stored in ChromaDB

Search

1. Query embedded into the same vector space
2. ChromaDB returns top 5 most similar chunks

Ask (RAG)

1. Question embedded → top 5 chunks retrieved
2. Chunks injected into a grounded prompt
3. Gemini answers using only the provided context

---

Roadmap

• PDF ingestion and text extraction
• Text chunking
• Semantic embeddings (local, sentence-transformers)
• Vector storage and similarity search (ChromaDB)
• RAG pipeline — retrieval-augmented generation via Google Gemini 2.5 Flash
• Multi-agent router — classify question type and route to the appropriate agent
• Summarization agent — full paper summarization
• Swap local embeddings for hosted provider (Google Gemini embedding API)
• Multi-paper search and cross-paper reasoning