PaperLens

An evidence-grounded workspace for reading and comparing research papers.

Upload a PDF or import an arXiv paper, then read, translate, ask questions, inspect citations, and compare papers without losing the path back to the source.

简体中文 · Documentation · Contributing · Roadmap

Product tour

Read the paper first

The single-paper reader is the product center. It reconstructs the document as a readable paper, keeps the original PDF one click away, and leaves only the outline, paper, and one Ask the paper entry visible by default. Figures, tables, references, translation controls, and parse warnings appear only in their reading context.

Move from a library into focused work

Paper library	Translation controls

The library is for opening and selecting papers—not creating a separate project model. Select two or three papers to enter comparison mode. Terminology and fixed translations live under reader settings, where they support bilingual reading without competing with it in the main navigation.

Compare with evidence, not flattened summaries

Comparison starts from papers you are already reading. Each paper is extracted independently before PaperLens aligns methods, experiments, metrics, limitations, and custom dimensions. Incomparable conditions remain visible, and supported claims retain links back to their paper evidence.

Why PaperLens?

Academic PDFs are not plain text. Reading order, columns, equations, tables, citations, and page geometry all matter. Sending an entire PDF directly to a language model often destroys that structure and makes answers difficult to verify.

PaperLens first builds a deterministic document representation. Retrieval and language-model workflows operate on that representation, and evidence links connect generated claims back to exact blocks, character spans, pages, and PDF coordinates.

What it can do

• Structured import: upload PDF files or import arXiv papers. Modern arXiv pages use structured HTML first; PDF parsing remains the fallback.
• Canonical-first Parser v2: probe, plan and fuse structure from optional Docling plus local PyMuPDF/pdfplumber backends, score paragraph/order/table/ formula/reference quality, then selectively repair only weak pages. GROBID and PaddleOCR-VL are optional semantic and visual providers.
• Bilingual reading: translate incrementally with terminology, citation, number, and formula protection.
• Grounded Q&A: retrieve paragraph-level evidence with BM25, draft atomic claims, run deterministic and model-based attribution checks, then link every accepted claim to its source.
• Adaptive Paper Agent: route quick questions through the fast reader and plan 3–8 evidence, method, experiment, reproduction, or critic capabilities for deeper questions. Findings distinguish fact, inference, assessment, and unknown states.
• Cross-paper comparison: compare two or three paper versions, distinguish missing evidence from confirmed absence, and avoid ranking incomparable metrics or datasets.
• Installable terminology: domain term packs participate in translation automatically; personal entries are explicit overrides rather than a separate terminology product.
• Reference workflow: extract references, lint citation formatting, resolve identities through scholarly APIs, and import public arXiv sources.

Quick start

Requirements

• Python 3.10 or newer
• Bun 1.3 or newer
• An OpenAI-compatible chat-completions endpoint for translation and analysis features; parsing and offline tests do not require a model

Run locally

git clone https://github.com/ZilaiWang/paperlens.git
cd paperlens

python3 -m venv .venv
.venv/bin/pip install -e "core[server,dev]"
# Optional enhanced geometry/table path; review its AGPL/commercial license.
.venv/bin/pip install -e "core[pymupdf]"
cp .env.example .env
.venv/bin/uvicorn server.app.main:app --reload --port 8700

In a second terminal:

cd web
bun install --frozen-lockfile
NEXT_PUBLIC_API_BASE=http://127.0.0.1:8700 bun run dev

Open http://127.0.0.1:3000. The API schema is available at http://127.0.0.1:8700/docs.

The default .env.example points to a local OpenAI-compatible endpoint. Update OPENAI_BASE_URL, OPENAI_API_KEY, and PAPERLENS_MODEL for your provider. See the configuration reference for every supported setting.

Architecture

flowchart LR
    UI["Next.js web"] --> API["FastAPI API"]
    API --> JOBS["Jobs and SSE events"]
    JOBS --> CORE["paperlens_core"]
    CORE --> IR["DocumentIR"]
    IR --> RETRIEVAL["BM25 and evidence ledger"]
    RETRIEVAL --> LLM["Bounded LLM workflows"]
    API --> DB["SQLite and local files"]

Loading

The repository is a small polyglot monorepo:

core/       Python domain library: parsing, DocumentIR, retrieval, analysis
server/     FastAPI transport, persistence adapters, jobs, and application services
web/        Next.js reader and comparison interface
scripts/    Reproducible evaluation and release utilities
tests/      Offline regression tests and downloadable corpus manifest
docs/       User, operator, contributor, architecture, and decision documentation

Read Architecture for data flow, module boundaries, and the rules intended to keep the project evolvable.

Development

.venv/bin/ruff check core server scripts tests
.venv/bin/python -m pytest
cd web && bun run build

The evaluation PDFs are intentionally not stored in Git. Download and verify them from the checked-in manifest when needed:

.venv/bin/python scripts/fetch_eval_corpus.py
.venv/bin/python scripts/eval_parse.py --corpus tests/eval_corpus
PYTHONPATH=core .venv/bin/python scripts/agent_bench.py

See Development, Testing, and Contributing before opening a pull request.

Project status and limitations

PaperLens 1.3 is usable as a self-hosted, single-process application. Anonymous workspaces use an opaque HttpOnly session cookie and storage-level scoping, and development CORS is restricted to the configured frontend origins. This is identity isolation, not user-account authentication: it is not yet a multi-tenant cloud service. SQLite and in-process jobs remain deliberate constraints. Do not expose the API directly to the public internet without an authentication layer, TLS, and a restrictive reverse proxy.

Complex borderless tables, scanned PDFs, formula OCR, and unusual multi-column layouts can still produce partial parses. PaperLens reports parse and evidence gaps rather than treating missing extraction as proof that a paper omitted something.

See Roadmap for the planned modularization and scaling path.

Community and security

• Use GitHub Issues for bugs and feature proposals.
• Read SECURITY.md before reporting a vulnerability.
• Participation is governed by CODE_OF_CONDUCT.md.

License

PaperLens is released under the MIT License. Papers imported by users remain subject to their original licenses and terms; PaperLens does not grant redistribution rights for third-party documents. Optional PyMuPDF support is dual-licensed under AGPL/commercial terms; review third-party notices before distribution or hosted use.