PaperPilot

PaperPilot extracts text from PDF research papers, chunks them, and uses the Gemini API to produce a structured review (summary, claims, method, results, reproducibility, risk breakdown, experiments, citations). It supports a replication plan, RAG-lite Q&A over chunks, and Markdown export.

What it does

• PDF upload → extract text (pdfplumber, fallback pypdf)
• Chunking → ~800–1200 token chunks with overlap; cap at 40 chunks (with batch summarization when longer)
• Review → Gemini returns structured JSON; risk score is computed in code from a deterministic rubric (dataset, splits, hyperparameters, etc.)
• Citations → Every claim/result/checklist/risk item includes chunk_id + quote, or “Not found in provided text” when absent
• Replication plan → Generate prerequisites, data/env, training/eval, ablations, compute, risks
• Q&A → TF-IDF retrieval over chunks → Gemini answer with citations; “Not found in provided text” if unsupported
• Export → Download Review and/or Review + Replication Plan as a single Markdown file

Setup

1. Virtual environment (recommended)

cd project
python -m venv .venv
source .venv/bin/activate   # Linux/macOS
# Windows:  .venv\Scripts\activate

2. Install dependencies

pip install -r requirements.txt

3. API key

Secrets are read only from the environment (e.g. via .env and python-dotenv).

Option A — .env

cp .env.example .env
# Edit .env and set:
# GEMINI_API_KEY=your_actual_key_here

Option B — export

export GEMINI_API_KEY=your_actual_key_here

Get a key at Google AI Studio. GOOGLE_API_KEY is also accepted.

Run

From the project directory:

streamlit run paperpilot.py

Open the URL shown (typically http://localhost:8501).

90-second demo

1. Run streamlit run paperpilot.py and open the app.
2. Upload a PDF research paper (e.g. a short ML paper).
3. Wait for extraction and chunking (and summarization if the paper is long).
4. Click Analyze with Gemini and wait for the review.
5. Check the risk score badge and Risk tab (breakdown table).
6. Click Generate Replication Plan, then Download as Markdown.
7. In Q&A, ask e.g. “What dataset did the authors use?” and check the answer and Evidence expander.

Known limitations

• Scanned PDFs: If extraction yields < 1500 characters, the app shows an error and suggests trying another PDF or using OCR.
• Long papers: Chunks are capped at 40; excess is batch-summarized before review. Very long papers may lose detail.
• Gemini: Rate limits or timeouts trigger one automatic retry, then a user-friendly error.

Requirements

See requirements.txt:

• streamlit — UI
• pdfplumber — PDF extraction (primary)
• pypdf — fallback extraction
• google-generativeai — Gemini API
• python-dotenv — load .env
• pydantic — schemas
• scikit-learn — TF-IDF retrieval

License

Use and modify as you like.