Idea2Paper — Upload → arXiv → SPECTER → PEGASUS (non-LLM scientific drafts)

A lightweight pipeline that turns your own notes/experiment logs into a short, research‑style draft by:

• extracting keywords from your upload,
• pulling related arXiv papers,
• ranking them with SPECTER (semantic similarity),
• and summarizing the background with PEGASUS‑arXiv (abstracts or full‑PDFs).

No paid LLM keys required; models run locally via Hugging Face + PyTorch. Works on CPU or CUDA if available.

---

✨ Pipeline

Pipeline (abstract)

flowchart TD
  A[Upload notes / draft] --> B[Keyword Extraction]
  B --> C[arXiv Retrieval]
  C --> D[SPECTER Ranking]
  D --> E[Summarize Background]
  E --> F[Compose Research-style Draft]
  F --> G[Markdown Output]

  %% optional helper (comment out if not wanted)
  A -. optional .-> X[Light edits / clarifications]
  X -. feeds .-> B

Loading

This diagram intentionally keeps the flow abstract and independent of any specific PDF/full‑text step.

---

🚀 Quickstart (Windows)

git clone https://github.com/<your-username>/Idea2Paper.git
cd Idea2Paper

# Create & activate venv
python -m venv .venv
. .venv/Scripts/activate

# Install deps
pip install --upgrade pip
pip install -r requirements.txt

# Run the Streamlit app
streamlit run src/app.py

GPU (optional)

If you have an NVIDIA GPU, CUDA Toolkit/Driver installed:

• The app auto‑detects CUDA. You can force it via env var:

  set IDEA2PAPER_DEVICE=cuda

• Models will use fp16 on GPU to save VRAM and speed up generation.

Torch vulnerability notice: Some Transformers checkpoints require torch>=2.6 (or safetensors) because of a security fix (CVE‑2025‑32434). If you hit a “please upgrade torch” error, use:

pip install --upgrade "torch>=2.6" --index-url https://download.pytorch.org/whl/cu121  # choose cu version to match your CUDA

---

🗂️ Repository Layout

Idea2Paper/
├─ flowchart/                  # Mermaid diagram(s) for the README
│  └─ pipeline.md
├─ data_samples/               # (Empty in repo; add your own test files) 
├─ drafts/                     # Generated drafts (git-ignored)
├─ notebook/
│  └─ demo.ipynb               # Optional smoke tests / quick experiments
├─ src/
│  ├─ app.py                   # Streamlit UI
│  ├─ config.py                # Central configuration
│  ├─ ingest.py                # Upload parsing (pdf/docx/txt/md/csv)
│  ├─ keywords.py              # Keyphrase extraction
│  ├─ retrieval.py             # arXiv query builder + fetch
│  ├─ ranker.py                # SPECTER-based re-ranking
│  ├─ summarizer.py            # PEGASUS summarization (abstracts / full-PDF)
│  ├─ generator.py             # Compose Markdown draft
│  ├─ utils.py                 # Helpers
│  └─ fulltext.py              # (Optional) PDF extraction helper
├─ .gitignore
├─ requirements.txt
└─ README.md

---

⚙️ Config (important knobs)

All tunable via src/config.py or environment variables:

• DEVICE — "cuda" or "cpu" (auto-detected; override with IDEA2PAPER_DEVICE)
• TOP_K — how many top ranked papers to use in the draft
• MAX_ARXIV_RESULTS — how many raw arXiv results to fetch before ranking
• USE_FULL_PDFS — true/false (if true, attempt to download & summarize full PDFs)
• PDF_TIMEOUT, PDF_MAX_BYTES, PDF_CHAR_CAP — safety rails for PDF downloads
• SUM_MODEL — default summarizer (google/pegasus-arxiv)
• EMBED_MODEL — SPECTER encoder (sentence-transformers/allenai-specter)

---

✅ What works well

• No API keys needed; local semantic ranking + summarization
• RAG based on your upload + arXiv related work
• Simple UI that makes it easy to iterate and save drafts

😕 What disappointed me

• The generated “paper” is concise and not as rich as LLM‑assisted drafts.
• Quality depends heavily on upload clarity and arXiv retrieval.
• Full‑PDF extraction can be slow/unreliable depending on source PDFs.

🧭 Roadmap / ideas to improve

• Swap to a local instruction‑tuned LLM (Ollama) for drafting & section structure
• Citation scraping and reference formatting (BibTeX export)
• Domain‑specific rankers (SciBERT, multi‑query retrieval, category filters)
• Smarter hierarchical long‑context summarization across multiple PDFs with chunk memory
• Better PDF parsing (layout‑aware extraction / figure caption mining)

---

🧰 Troubleshooting

• CUDA found but slow? Some models still do heavy tokenization on CPU; that’s expected. You can still speed up generation by using fp16 on GPU.
• Transformers says “upgrade torch”? Install torch>=2.6 or switch to safetensors‑only models.
• arXiv returns zero results? Tweak keywords; try fewer/more general terms.

---

🤝 Credit

No license file right now. If you use this repo in your work or demos, a short credit like

“Built on Idea2Paper (https://github.com/Sivarohitk/Idea2Paper)”
is highly appreciated.