reviewer2

AI-powered academic paper reviewer for Node/TypeScript — the reviewer #2 every paper deserves. Based on OpenAIReview (Python), redesigned for web-app integration. On top of the original it adds:

• Built for web apps — headless and JSON-first: every review returns visualization-ready JSON you can render directly; no CLI, no result files to manage.
• Survives long reviews — a thorough review can run for an hour; the pipeline breaks into durable, retryable steps that drop into Inngest (or any background-job runtime), with live progress updates and cancellation.
• Reference accuracy check — opt-in verification of every bibliography entry against Crossref/OpenAlex/arXiv (free, keyless) to catch hallucinated or inaccurate citations, with links to the matched records.
• Any LLM provider — OpenAI, OpenRouter, Anthropic, or Gemini out of the box, including per-user API keys for multi-tenant apps.
• Easy prompt customization — swap the review criteria, the tone, or entire prompts with plain-string overrides, and keep named presets per journal or track.
• Cost tracking with live pricing — every review reports what it cost in dollars, and you can pull current per-model prices from LiteLLM or OpenRouter instead of relying on a static table.
• Papers from anywhere, running anywhere — beyond the original's local files and arXiv links: any file URL (including presigned S3/GCS links) and in-memory buffers, parsed in pure JS so it runs on serverless with no system dependencies.

See What reviewer2 adds on top of OpenAIReview for the full list.

Install

npm install reviewer2

Node ≥ 20, ESM and CJS builds included.

Quick start

export OPENAI_API_KEY=sk-...   # or OPENROUTER_/ANTHROPIC_/GEMINI_API_KEY

import { parseDocument, reviewPaper } from "reviewer2";

// 1. Get the paper text (PDF/DOCX/TeX/MD file, or an arXiv URL)
const doc = await parseDocument("paper.pdf"); // or parseDocumentBuffer(bytes, "pdf")

// 2. Review it — returns viz-compatible JSON
const { paper, result } = await reviewPaper(doc.text, {
  title: doc.title,
  ocr: doc.wasOcr,
  method: "progressive",              // default; also: zero_shot | local
  model: "gpt-5.2",                   // default depends on provider
  checkReferences: true,              // opt-in: verify citations against Crossref/OpenAlex/arXiv
  onProgress: (e) => console.log(e),  // { stage: "passage", current, total, ... }
});

console.log(JSON.stringify(paper, null, 2));

Keys can also be passed explicitly per call ({ provider, apiKey }) for multi-tenant apps — see Providers & API keys.

Documentation

| Guide | What's inside | |---|---| | How it works | Output JSON contract, pipeline flowchart, the three review methods compared, building a UI | | Reference accuracy check | Catching hallucinated/inaccurate citations: usage, per-entry results with source links, how matching works, tuning | | Customizing prompts | Block and template overrides, placeholders, presets per journal/track, validation | | Providers & API keys | Env vars vs explicit keys, provider auto-detection, per-provider notes | | Document parsing | PDF/DOCX/LaTeX/MD, arXiv and file URLs, buffers, OCR handling | | Cost tracking | Pricing tables, live pricing, per-model usage breakdown | | Long-running reviews (Inngest) | Durable step-by-step execution in a Next.js app | | vs. OpenAIReview | Everything reviewer2 adds on top of the original |

Development

npm install
npm test          # vitest (110 tests, no API calls or network)
npm run build     # tsup → dist/ (ESM + CJS + d.ts)
npm run typecheck

Agent skill

Prefer running reviews inside an AI coding agent instead of integrating the library? harrywang/reviewer2-skill packages the same review pipeline as an agent skill for Claude Code, Cursor, Codex, and others — install with npx skills add harrywang/reviewer2-skill, then run /reviewer2 paper.pdf. It produces the same viz-compatible JSON and bundles a local web viewer.

Credits

reviewer2 is based on OpenAIReview by ChicagoHAI (MIT licensed) — not a direct port, but a TypeScript reimplementation that follows its review pipeline design (progressive summary-based review, deep-check prompts, consolidation), prompt set, and fuzzy quote-to-paragraph anchoring, then extends it for web-app use (see what reviewer2 adds).

License

MIT