paperpilot 1.2.0

A CLI research agent for AI-related paper search, code discovery, PDF collection, and bilingual reports.

PaperPilot

English | 中文

PaperPilot is a CLI research agent for AI-related literature review. It turns a natural-language research request into a verified paper corpus, code/PDF collection, evidence-grounded synthesis, and bilingual reports in Markdown, HTML, and PDF.

It is designed as a file-system based research workflow, not a chatbot. Each run creates a self-contained run folder with state, logs, intermediate artifacts, evidence checks, and final reports.

Highlights

• Natural-language research intake with LLM-assisted query understanding.
• Layered Source Registry with arXiv, Semantic Scholar, OpenAlex, Crossref, OpenReview, PubMed, Europe PMC, bioRxiv, medRxiv, DBLP, ACL Anthology, and optional API-key sources.
• Local corpus import with --user-corpus for PDF, BibTeX, RIS, Markdown, and text files.
• Research protocol generation with inclusion/exclusion criteria and negative keywords.
• Corpus normalization, DOI/arXiv/title-similarity deduplication, ranking, and relevance screening.
• Code repository detection for GitHub, GitLab, Hugging Face, and project pages.
• Open-access PDF download only; no paywall bypassing.
• Full-text extraction for downloaded PDFs.
• Prompt Registry, Tool Registry, Capability Registry, and event logging.
• Evidence ledger that maps report-level claims to numbered paper citations.
• Review-agent checks for source verification, relevance, citation compliance, and overclaiming risk.
• Canonical bilingual report model with aligned Chinese/English Markdown, HTML, and PDF outputs.

Installation

From PyPI:

python -m pip install paperpilot -i https://pypi.org/simple

For local development:

git clone https://github.com/CHB-learner/PaperPilot.git
cd PaperPilot
python -m pip install -e .

LLM Configuration

PaperPilot requires an OpenAI-compatible LLM configuration for query understanding, planning, screening, synthesis, and report generation.

Interactive setup:

PaperPilot

Manual setup:

PaperPilot config set --base-url https://api.deepseek.com --model deepseek-chat
PaperPilot config import ./api.json
PaperPilot config list
PaperPilot config use deepseek
PaperPilot config show

Optional source API keys:

PaperPilot sources list
PaperPilot sources config core
PaperPilot sources config lens
PaperPilot sources enable core
PaperPilot sources test core

Configuration is stored in:

~/.paperpilot/config.json

Configuration priority:

1. Environment variables: OPENAI_API_KEY, OPENAI_BASE_URL, OPENAI_MODEL
2. User config: ~/.paperpilot/config.json
3. Legacy project file: llmapi.txt

Do not commit api.json, llmapi.txt, .env, or any file containing API keys.

Quick Start

Interactive mode:

PaperPilot

Command mode:

PaperPilot "RNA inverse folding sequence design" \
  --auto-confirm \
  --max-papers 50 \
  --since-year 2021 \
  --github-filter required \
  --sources auto \
  --mode apa \
  --quality balanced

Use local papers as seed corpus:

PaperPilot "RNA inverse folding sequence design" \
  --auto-confirm \
  --user-corpus ./papers \
  --user-corpus references.bib

Skip PDF downloads:

PaperPilot "vision language model" --auto-confirm --no-download

Inspect or rerun a task:

PaperPilot inspect runs/<task-id>
PaperPilot resume runs/<task-id>

Architecture

PaperPilot follows a state-machine workflow:

Intake -> Protocol -> Search -> Corpus -> Screening -> Verification -> Synthesis -> Review -> Report

flowchart LR
  U[User request<br/>topic + params + local corpus] --> C[Run context<br/>task/state/events]
  C --> P[Prompt Registry]
  P --> QA[Query Understanding Agent]
  QA --> PL[Planner Agent]
  PL --> RP[Research Protocol Agent]
  RP --> ST[Source Registry<br/>arXiv / S2 / OpenAlex / Crossref / OpenReview<br/>PubMed / Europe PMC / bioRxiv / medRxiv / DBLP / ACL]
  U --> LC[Local Corpus Import]
  LC --> CB[Corpus Builder]
  ST --> CB
  CB --> RJ[Relevance Judge<br/>core / adjacent / exclude]
  RJ --> VF[Verification + PDF Tools]
  VF --> LM[Literature Matrix]
  LM --> SA[Synthesis Agent]
  SA --> QG[Quality Gate + Reflection]
  QG --> EL[Evidence Ledger<br/>claim -> citation]
  EL --> RA[Review Agents<br/>source / citation / overclaiming]
  RA --> CR[Canonical Report]
  CR --> OUT[ZH/EN Markdown<br/>ZH/EN HTML<br/>ZH/EN PDF]

Default free sources include arXiv, Semantic Scholar, OpenAlex, Crossref, OpenReview, PubMed, Europe PMC, bioRxiv, medRxiv, DBLP, and ACL Anthology. Optional API-key sources include CORE, Lens.org, IEEE Xplore, Springer Nature, Elsevier/Scopus, and Dimensions.

The repository also includes an HTML architecture overview:

• paperpilot_agent_flow.html

Output Artifacts

Each run writes a folder under runs/<task-id>/ unless --output-dir is provided.

Core run files:

• task.json: task metadata and parameters.
• state.json: stage status.
• events.jsonl: stage event stream.
• manifest.json: generated artifact list.
• prompt_manifest.json: versioned prompt roles and required JSON keys.
• registries.json: built-in ToolRegistry and CapabilityRegistry.
• source_diagnostics.json: enabled sources, returned counts, and source-level errors.

Search and corpus files:

• query_understanding.md: keyword interpretation and ambiguity analysis.
• plan.json: search plan and diversified queries.
• protocol.json: research question, scope, inclusion/exclusion criteria, negative keywords.
• metadata.json: normalized raw search candidates.
• user_corpus_log.json: local corpus import log.
• corpus.json: screened full corpus.
• core_papers.json: core papers.
• adjacent_papers.json: adjacent papers.
• excluded_papers.json: excluded papers and reasons.
• ranked_papers.json: final report-view papers.

Evidence and quality files:

• verification.json: DOI, URL, PDF, and code status.
• download_log.json: PDF download status.
• fulltext/: extracted PDF text.
• paper_notes.json: full-text extraction metadata.
• literature_matrix.json: method/task/evidence matrix.
• synthesis.json: field overview, method taxonomy, paper summaries, trends, gaps.
• quality_gate.json: pass/retry/needs-user-attention verdict.
• reflection.json: search quality reflection and retry hints.
• evidence_ledger.json: claim-level evidence ledger.
• review_agent_findings.json: review-agent checks.

Final reports:

• report.canonical.json: shared bilingual report model and citation map.
• report.zh.md
• report.en.md
• report.zh.html
• report.en.html
• report.zh.pdf
• report.en.pdf
• pdfs/: downloaded open-access PDFs.

GitHub / Code Filter

PaperPilot "retrieval augmented generation" --auto-confirm --github-filter required

Filter modes:

• any: keep all papers and annotate code availability.
• required: final report view keeps papers with detected public code links; full screened corpus is still saved.
• none: final report view keeps papers without detected public code links.

CLI Options

--max-papers INT                 maximum papers in final report view
--since-year INT                 prefer papers since this year
--github-filter any|required|none
--github-search-limit INT        active GitHub search limit
--no-download                    skip PDF downloads
--pdf-limit INT                  maximum PDFs to download
--user-corpus PATH               import local corpus path; repeatable
--mode quick|apa|systematic
--interaction auto|gated
--quality fast|balanced|strict
--include-adjacent               include adjacent papers in matrix/appendix
--sources auto|all|core|biomed|cs|configured
--enable-source SOURCE           enable one additional source; repeatable
--disable-source SOURCE          disable one source; repeatable

Development

Run tests:

python -m unittest discover -s tests
python -m compileall literature_agent

Build locally:

python -m pip install build twine
python -m build
python -m twine check dist/*

Publish to PyPI:

python -m twine upload dist/*

Open Source Notes

Before pushing to GitHub:

• Make sure .gitignore is present.
• Do not commit API keys, local run outputs, build artifacts, or virtual environments.
• Add a LICENSE file before calling the project open source in a strict legal sense.
• If any PyPI or LLM token was ever committed, revoke it immediately and create a new one.

Suggested first push:

git init
git add README.md README.zh-CN.md pyproject.toml literature_agent tests paperpilot_agent_flow.html .gitignore LICENSE
git commit -m "Initial open source release"
git branch -M main
git remote add origin https://github.com/CHB-learner/PaperPilot.git
git push -u origin main