Skip to content

Latest commit

 

History

History
149 lines (109 loc) · 8.61 KB

File metadata and controls

149 lines (109 loc) · 8.61 KB
LitAnchor logo

LitAnchor Notes

Anchor every note to its source.

LitAnchor = Literature + Anchor — start from the literature and keep every note anchored to its source.

Release Windows CI Python 3.10+ Agent Skills License

Quick start · Why it is different · Reading modes · Note template · Docs · 中文

LitAnchor Notes is an evidence-first note brand. Its current public module, litanchor-paper-reading, handles one academic paper at a time. Give it a PDF for a standalone Markdown note, or use Zotero acquisition and safe Obsidian export.

Current public candidate: v1.0.0-rc1. It remains a Pre-release until public usage feedback supports promotion to the stable release.

Start in two prompts

Give the repository URL to an agent that can run local commands and access files:

Install this Skill for me: https://github.com/xieyf1024/Litanchor-paper-reading

Then provide a PDF directly:

Deep-read this PDF and create a Markdown note next to it.

Or use the integrated Zotero → Obsidian route:

Deep-read “Paper title” and save the note to my Research Vault.

The wording is not literal. Agents should recognize equivalent install, skim, deep-read, internalize, and export requests. The direct-PDF route needs neither Zotero nor Obsidian. The integrated route may confirm the Obsidian Vault, Literature Inbox, and MinerU consent policy on first use.

What you get

📌 Traceable 🧠 Complete 🖼️ Visual 🛡️ Safe
Important facts link to verified physical PDF pages Methods, equations, experiments, results, discussion, and limitations stay distinct Key method or result figures are selected and cropped from the original PDF Writes stay inside an authorized directory and refuse overwrite by default

Reader notes stay clean: they show compact linked p.x Zotero locators, while Evidence, Claims, full quotations, and validation records remain in private sidecars.

More than an AI summary

Generic summary tools LitAnchor
Generate fluent prose directly from the document Build Evidence → Claims → section synthesis before composing the note
Check only sentences already written Review both factual fidelity and important-content recall
Emit page links even when locations are uncertain Reject unverified PDF pages as formal evidence
Often skip equations, experiments, and figures Run dedicated method, metric, experiment-chain, and visual passes
Fill gaps with model knowledge Use the supplied paper as the only factual source

The workflow at a glance

flowchart LR
    source["PDF / Zotero"] --> parse["Parse and map pages<br/>PyMuPDF + optional MinerU"]
    parse --> ground["Read and ground<br/>Evidence → Claims"]
    ground --> review["Review twice<br/>Fidelity + recall"]
    review --> note["Markdown / Obsidian<br/>Page locators + key figures"]

    classDef sourceNode fill:#E8F1FF,stroke:#2563EB,color:#172554
    classDef processNode fill:#FFF7E6,stroke:#D97706,color:#451A03
    classDef outputNode fill:#ECFDF5,stroke:#059669,color:#064E3B
    class source sourceNode
    class parse,ground,review processNode
    class note outputNode
Loading

PyMuPDF remains authoritative for physical pages, quotations, coordinates, and original-image crops. For eligible files with user consent, MinerU automatically improves heading structure, reading order, captions, and complex-layout candidates. Its output must align back to the original PDF before supporting evidence.

Three reading modes

Mode Use it when Main output
skim You need a fast decision on what the paper says and whether to continue One-sentence summary, question, method skeleton, main results, conclusion boundary, and essential locators
deep (default) You need an auditable graduate-level paper note Background/gap, data, methods, equations, experiments, results, visuals, interpretation, limitations, and conclusions
internalize You want to turn reading into testable research action Everything in deep, plus research connections, falsifiable hypotheses, validation design, failure conditions, and follow-up reading

All modes share the same factual boundary. Learning-layer content in internalize is labelled [analysis], [hypothesis], or [user]; it never masquerades as an author conclusion. See Paper Template v1.0 for the complete structure.

Installation and requirements

The current public build targets local-capable agents on Windows. The user asks for installation; the agent downloads the Release, creates an isolated environment, installs dependencies, runs doctor, and completes first-run setup.

Environment requirements and agent entry points
  • Windows 10/11 x64;
  • Python 3.10 or newer;
  • Zotero 7 or newer (only for the Zotero integration route);
  • Obsidian Desktop with a local Vault (only for Obsidian export);
  • an agent that can run local commands, write to authorized paths, and download dependencies.

Python and Zotero use minimum versions only. Newer untested versions are capability-probed by doctor instead of being rejected by an arbitrary maximum.

.\install.ps1 -Action Install
.\litanchor.ps1 doctor
.\litanchor.ps1 setup -Vault "Vault name" -Inbox "LitAnchor\00_Inbox" -MinerUConsent ask_each_time -CreateInbox
.\litanchor.ps1 run-plan -Paper "Paper title" -Vault "Vault name"
.\litanchor.ps1 run-plan -PdfPath "D:\Papers\paper.pdf" -OutputNote "D:\Notes\paper.md"

These are agent and contributor interfaces, not mandatory manual steps for end users. See installation requirements and integrations.

Reliability contract

  • Use only the supplied paper for formal paper facts.
  • Preserve author uncertainty; never upgrade interpretation or speculation to fact.
  • Preserve numbers, units, variables, ranges, errors, and applicability conditions.
  • Block formal export when parsing, evidence, or important-content recall fails.
  • Keep Zotero read-only and Obsidian writes inside the authorized Inbox.
  • Require local consent for external parsing; MinerU is never an authoritative evidence source.

Current boundaries

  • One paper per run; no batch review or knowledge graph.
  • Native-text PDFs are the stable path; scanned and pathological layouts may degrade or block.
  • Markdown is the sole formal note output; LitAnchor does not generate a separate PDF note.
  • No Zotero write-back, bidirectional sync, or silent overwrite of existing notes.
  • Review key visuals, core numbers, and source locators before long-term use.

Documentation and contributing

Documentation · Workflow · Data Schema · Evaluation · Roadmap · Contributing · Security

Use the matching installation, PDF runtime, or note-quality Issue form. Never upload paper PDFs, private Zotero data, Obsidian Vault contents, API keys, personal annotations, or runtime Evidence/Claim artifacts.

License

GNU Affero General Public License v3.0 only. See THIRD_PARTY.md and NOTICE.md for PyMuPDF licensing, MinerU service boundaries, and design references.