papermind-ai 0.1.0

Read arXiv papers for real: structured analysis, grounded RAG Q&A, and reproduction tutoring from the command line.

PaperMind

English · 简体中文

Read any paper — from skim to reproduction — with every key claim traced back to the source. Paste an arXiv link, DOI, paper title, or any paper page / PDF URL (or upload a PDF): structured analysis, grounded & citation‑verified Q&A, a whole‑method framework diagram (downloadable SVG), and a runnable reproduction guide.

_{↑ A paper-style teaching diagram PaperMind generates for a technical point (real output) · full sample report · gallery →}

Why it isn't just another chat‑with‑PDF

• 🔬 We trained our own retrieval reranker — +14pt Recall@5 over a strong dense baseline (QASPER, held‑out test). Most "chat with a paper" tools wrap an API; PaperMind's evidence retrieval is a cross‑encoder we fine‑tuned and measured (see Benchmarks).
• 🔎 Key claims are labeled and citation‑verified. Answers are split into fact / inference (with confidence) / out‑of‑scope, and each cited quote is checked against the actual paper text — if it can't be found, it's flagged ⚠️ rather than silently trusted.
• 📐 Whole‑method framework diagram. Beyond per‑point figures, PaperMind reconstructs the paper's end‑to‑end method as a single Figure‑1‑style architecture diagram (steps the paper only implies are marked inferred) — view it on the web /framework tab and download it as SVG.
• 🛠️ Reproduction grounded in the real code repo. PaperMind locates the paper's code repository and builds setup.sh from the repo's actual dependency file and README run‑commands — not a hallucinated guess.
• 🆓 Runs fully local & free. papermind demo works offline; --local routes everything through Ollama; with no key it auto‑falls back to local.

	PaperMind	Typical chat‑with‑PDF / SaaS readers
Per‑claim citation with verification (flags unverifiable ⚠️)	✅	—
Reproduction → runnable setup.sh from the real repo	✅	—
Self‑trained RAG reranker, measured on QASPER	✅	—
Runs fully local / offline, zero cost	✅	✗
Open‑source & self‑hostable	✅	✗

_{“—” = not a documented capability we're aware of; open an issue to correct us. Open‑source/local rows are clear‑cut for closed SaaS tools.}

Three ways to use it

	How	Best for
🌐 Online	open papermind.try2026.cn — zero install	trying it instantly
🖥️ GUI	pip install "papermind-ai[web]" → papermind ui	full features, in your browser
⌨️ CLI / API	pip install papermind-ai → papermind analyze https://arxiv.org/abs/2307.08691	scripting, batch, integration

🆓 No spend: papermind demo shows the output offline (no key); add --local to any command to run fully on local Ollama at zero cost.

What you get

╭──────────────── Attention Is All You Need · 2017 · arXiv:1706.03762 ───────────────╮
🎯 Contributions   A fully attention-based model (Transformer): no recurrence/convolution…
🔬 Technical       1. Scaled Dot-Product Attention [high]  softmax(QKᵀ/√d_k)·V; scaling avoids
                      saturated gradients   💡 analogy + 📊 paper-style teaching diagram (SVG)
💬 Grounded Q&A    "Why divide by √d_k?" → [FACT] large d_k inflates dot-product variance…
                      📌 source Section 3.2.1 (p.4) ✓ verified
🛠️ Reproduction    code repo (★24k) github.com/… → setup.sh (real deps + README run commands)

Full rendered reports (open on GitHub — paper‑style diagrams, jump‑to‑source citations, reproduction tables): Transformer · ViT · PPO · the full gallery →

Gallery — 7 papers across 5 areas

PaperMind isn't NLP‑only. Real reports generated by the pipeline, grouped by area:

Area	Papers
NLP / Attention	Transformer · FlashAttention‑2 · Llama 2
Computer Vision	ViT
Generative / Diffusion	Latent Diffusion (Stable Diffusion)
Reinforcement Learning	PPO
Efficient fine‑tuning	LoRA

Benchmarks — the reranker is trained and measured, not a black box

We fine‑tuned a cross‑encoder reranker (bge‑reranker‑base) on QASPER. Over all‑paragraph candidates, versus a strong dense baseline:

	Recall@5	MRR	nDCG@10
Dense (bge‑small‑en‑v1.5)	0.519	0.463	0.469
+ our reranker	0.660	0.612	0.609

Consistent across dev (888 q) and a held‑out test (1309 q) — no overfitting. Reproduce it: docs/RESEARCH_PLAN.md (trainer/ to train · evaluation/ to evaluate).

Self‑host the online service

One Docker command. Public exposure is safe by default (demo mode is read‑only cache, never spends a key):

git clone https://github.com/Wenhao-Hua/papermind && cd papermind
docker build -t papermind . && docker run -p 8080:8080 papermind            # demo mode

docker run -p 8080:8080 -e OPENAI_API_KEY=sk-... -e PAPERMIND_TRUST_PROXY=1 papermind \
       papermind serve --host 0.0.0.0 --port 8080 --live                    # live (your key, billed)

--live is rate‑limited by default (8/IP/day, 300/day global; tune with --rate-per-ip / --rate-global, 0 = unlimited) so a public deployment can't drain your key. Free ops (search / cached / offline demo) aren't limited.

Behind Cloudflare/a reverse proxy, set PAPERMIND_TRUST_PROXY=1 — otherwise every visitor shares one per‑IP quota (the proxy's IP) and per‑IP limiting silently breaks. Only set it when you're actually behind a trusted proxy that sets cf-connecting-ip.

Faster figures: diagrams default to the main model (cleanest layout, slower). Point figures at a fast non‑thinking model to speed them up: papermind config set figure-model deepseek/deepseek-chat.

What it can do

analyze (4‑module report) · summary (TL;DR) · ask/chat (grounded Q&A) · tutor/debug (reproduction help) · compare (multi‑paper) · reproduce (export setup.sh / notebook) · search/batch/list · cite.

Models via litellm: OpenAI / Anthropic / DeepSeek / Gemini / Qwen (DashScope) / local Ollama. Results are cached locally — re‑runs are instant. More: papermind --help.

Contributing / License

PRs and issues welcome (CONTRIBUTING.md) · MIT