AcademicRAG

A local-first RAG system for academic literature review. Built with ChromaDB, Streamlit, and multiple LLM backends. Designed for researchers, graduate students, and anyone working with a corpus of academic papers.

Originally built for a Master's thesis at Bocconi University on automation and political ideology, then generalized as a reusable tool.

Features

Core RAG

• Semantic search over your paper corpus using all-MiniLM-L6-v2 embeddings
• Knowledge graph integration — retrieved context is enriched with a concept graph linking ideas across papers
• Source citations — every answer shows the exact paper excerpts used, with author, year, and page number

Three LLM Backends

Mode	Cost	Privacy	Setup
Ollama (local)	Free	Full privacy — nothing leaves your machine	ollama serve
Claude Code (MAX)	Included in MAX subscription	Anthropic handles data	Install Claude Code
Claude API	Pay per token	Anthropic handles data	Add ANTHROPIC_API_KEY to .env

Academic Tools

Six specialized tools beyond basic Q&A:

Tool	What it does
Citation Generator	Extract citations from sources, export to LaTeX/BibTeX or APA
Literature Review Drafter	Auto-generate literature review sections with proper citations
Paper Comparator	Create comparison tables across methodology, findings, theory, etc.
Claim Extractor	Extract key claims and find contradictions between papers
Research Gap Finder	Identify unanswered questions and underexplored areas
Defense Prep	Generate potential defense questions and prepare counterarguments

Deep Paper Extraction (LLM-Powered)

Automated structured extraction pipeline that processes each paper and outputs:

• Constructs — theoretical concepts with definitions and verbatim quotes
• Variables — measured quantities with operationalization details
• Methods — empirical approaches with specifics (F-stats, controls, sample sizes)
• Claims — key findings with evidence and strength ratings (strong/moderate/weak)

Extractions feed into a concept knowledge graph that links ideas across papers with typed edges (causes, correlates, moderates, supports, contradicts).

Auto-Cataloging (No LLM Required)

The generate_catalog.py script builds structured metadata for your entire corpus using pure Python heuristics:

• Abstract extraction from PDF text
• Key findings via signal-phrase detection
• Methodology tagging (shift-share, TWFE, IV, RDD, survey, panel data, ...)
• BibTeX key matching (fuzzy author + year + title overlap)
• Category assignment (labor economics, political behavior, psychology, ...)
• Concept graph linking

Quick Start

1. Clone and install

git clone https://github.com/AlessandroCaforio/AcademicRAG.git
cd AcademicRAG

# Create virtual environment
python3 -m venv venv
source venv/bin/activate

# Install dependencies
pip install -r requirements.txt

2. Add your papers

# Add papers one at a time
python add_paper.py papers/MyPaper.pdf "Author Name" "Paper Title" "2024"

# Or drop PDFs into papers/ and add them in bulk
for pdf in papers/*.pdf; do
    python add_paper.py "$pdf" "Unknown" "$(basename "$pdf" .pdf)" "2024"
done

3. Run

# Quick start (handles venv + Ollama check)
./run.sh

# Or manually
streamlit run app/main.py

Open http://localhost:8501 in your browser.

4. (Optional) Extract structured knowledge

# Extract a single paper
python extract_paper.py "Paper Title" --backend claude

# Extract ALL unprocessed papers (batch mode)
python run_extraction.py --backend claude

# Build/update the concept graph from extractions
python build_graph.py extractions/

# See what would be processed (no LLM calls)
python run_extraction.py --dry-run

5. (Optional) Generate catalog

python generate_catalog.py
# Outputs: paper_catalog.json + paper_catalog.md

Configuration

Copy .env.example to .env:

cp .env.example .env

Variable	Default	Description
OLLAMA_MODEL	llama3.1:8b	Local LLM model
ANTHROPIC_API_KEY	—	For Claude API mode (optional)
CLAUDE_CODE_PATH	claude	Path to Claude Code CLI
SYSTEM_PROMPT	(built-in)	Custom system prompt for your research domain
APP_TITLE	AcademicRAG	UI title
TOP_K_RESULTS	5	Number of sources per query

Customizing the System Prompt

Set SYSTEM_PROMPT in .env to tailor responses to your research domain:

SYSTEM_PROMPT="You are a research assistant for a PhD thesis on climate economics. Base answers on retrieved paper excerpts and cite specific authors."

Architecture

User Question
     │
     ▼
┌─────────────┐     ┌──────────────────┐
│  Embedding   │────▶│  ChromaDB Vector  │
│ (MiniLM-L6) │     │     Store         │
└─────────────┘     └──────────────────┘
                            │
                     Top-K chunks
                            │
                            ▼
                    ┌───────────────┐
                    │ Context Builder │◀── Concept Graph (optional)
                    └───────────────┘
                            │
                    Context + Question
                            │
                            ▼
                 ┌─────────────────────┐
                 │   LLM Generation    │
                 │  (Ollama / Claude)   │
                 └─────────────────────┘
                            │
                            ▼
                   Answer + Sources

Key Design Decisions

• ChromaDB for the vector store — persistent, local, no server needed
• all-MiniLM-L6-v2 for embeddings — fast, good quality, runs on CPU
• Sentence-boundary chunking with overlap — preserves context across chunk boundaries
• Concept graph enrichment — optional knowledge graph adds structured relationships to raw retrieval
• Three LLM backends — privacy-first (Ollama) with cloud fallbacks

Project Structure

AcademicRAG/
├── app/
│   ├── config.py              # All settings (env-configurable)
│   ├── main.py                # Streamlit interface
│   ├── rag_engine.py          # Core RAG: indexing, retrieval, generation
│   └── academic_features.py   # 6 academic tools
├── papers/                    # Your PDFs + paper_chunks.json
├── extractions/               # Structured paper extractions (auto-generated)
├── data/vectors/              # ChromaDB storage (auto-generated)
├── add_paper.py               # PDF → chunks ingestion
├── extract_paper.py           # LLM-powered structured extraction (single paper)
├── build_graph.py             # Concept graph builder (from extractions)
├── run_extraction.py          # Batch orchestrator (all papers)
├── generate_catalog.py        # Auto-catalog (no LLM needed)
├── run.sh                     # Quick start script
├── requirements.txt
├── .env.example
└── docs/
    └── ARCHITECTURE.md        # Detailed architecture & concept schema

Example Queries

• "What are the main findings on automation and wages?"
• "Compare the methodologies used across papers"
• "What does Author (2024) argue about X?"
• "Find contradictions between Paper A and Paper B"
• "What research gaps exist in this literature?"
• "Generate defense questions for my thesis"

Technical Details

Component	Technology
Embeddings	all-MiniLM-L6-v2 (384-dim, ~80MB)
Vector Store	ChromaDB (persistent, local)
Chunking	~1500 chars, 200 char overlap, sentence-boundary aware
Retrieval	Top-K semantic search (cosine similarity)
UI	Streamlit
LLMs	Ollama / Claude API / Claude Code

License

MIT

Acknowledgments

Built during the Master's thesis "Automation and Political Ideology: Task Displacement and the Conservative Shift Among Exposed Workers" at Bocconi University, supervised by Prof. Massimo Anelli.

The concept graph schema and agentic extraction pipeline were inspired by the iterative literature analysis approach described in the architecture docs.