Skip to content

Releases: marmot1123/LumenCite

LumenCite v1.0.0

Choose a tag to compare

@github-actions github-actions released this 17 Aug 14:10
1853b55

LCIR — the machine-readable representation of a paper that LumenCite has been assembling since v0.8.0 — is finished, and it is on by default. Papers are analysed as you add them, an existing library fills in gradually on launch, full-text search reads its text from the result, and the in-app assistant can read a paper by its structure instead of as one undifferentiated block of text.

There is no migration — this release adds no schema change at all — and the Web Clipper extension is unchanged from v0.8.0 (v0.2.0).

Read this before updating

  • LCIR is now on by default. It counts as "off" only if you actually turned it off. A library that simply never touched the toggle has no setting recorded, and that is what now counts as on — so if you have been using LumenCite since before this release without going near LCIR, it is on for you too. You can switch it off in Settings → Data.
  • That means disk usage grows without you doing anything. Each analysed paper writes a cropped PNG per figure under attachments/<entry>/.lcir/, and those crops are included in backup archives. On a real 138-paper library the crops in use come to about 554 MB, and one backup generation is about 750 MB — up from 531 MB before any of this existed. The launch-time backfill is deliberately unhurried (a time budget per run, checked between papers, standing aside whenever you start a batch yourself or a backup begins).
  • Existing papers are not re-analysed automatically. The extractor moved from 0.6.0 to 0.14.0, so every paper whose LCIR was built from a PDF by v0.10.0 or earlier counts as out of date. Raising a version never triggers a rebuild — you have to press "rebuild outdated LCIR" in Settings → Data.
    • What it costs, measured on that 138-paper library: about 20 minutes, plus a further 30 minutes and roughly $0.80 of API calls if you have figure descriptions enabled (only genuinely new figures are billed — existing descriptions follow their figure by image fingerprint).
    • What skipping it leaves behind: figure regions stay at 1,198 instead of 1,629 (+431), and figure–caption pairing stays at 262 instead of 662.
    • The separate reader for arXiv TeX sources is unchanged at 0.5.0 and stays current.

Windows and Linux: this one needs a manual download

Structural analysis and Vision OCR both read PDFs through the same native library (pdfium), and until now only the macOS build shipped a copy of it. v1.0.0 is the first release that bundles it on Windows and Linux, so on those platforms this is the first version where either actually runs.

In-app auto-update is macOS-only, so Windows and Linux only get told that a new version exists. Until you download and run the new installer from this page, structural analysis and Vision OCR do not work there at all.

Install

  • macOSLumenCite_1.0.0_universal.dmg (signed + notarized), or via Homebrew:
    brew tap marmot1123/lumencite
    brew trust marmot1123/lumencite
    brew install --cask lumencite
    Existing macOS installs update in place via the built-in updater.
  • Windows.msi or -setup.exe (signed with a Certum Open Source Code Signing certificate). SmartScreen may still warn until download reputation builds; choose "More info" → "Run anyway".
  • Linux — AppImage / .deb / .rpm.
  • Web Clipper extension — unchanged from v0.8.0. If you already have v0.2.0 loaded, no action is needed; otherwise grab lumencite-clipper-0.2.0.zip from this release and load it unpacked (see the README "Browser extension" section).

See CHANGELOG.md for the full list of changes.

Full changelog: v0.10.0...v1.0.0

LumenCite v0.10.0

Choose a tag to compare

@github-actions github-actions released this 27 Jul 21:43
5d955b4

This release completes Phase 8 of LCIR — the experimental, machine-readable intermediate representation for papers — and adds its first export path. Papers can now be written out as LCIR JSON or as structured Markdown; figures are detected and cropped out of PDFs, TeX tables are structured cell by cell, and figures can optionally be described by an LLM Vision model as alt text.

Everything stays gated behind the off-by-default lcir.enabled flag, so default behaviour is unchanged. Two additive migrations (0019, 0020) create new tables that remain empty unless the flag is on — no data migration, and existing libraries upgrade unchanged.

Added

  • LCIR — export (Phase 9a, experimental) — write an entry out as LCIR JSON (the validated structure, with provenance, coordinates and confidence) or as structured Markdown (sections, raw LaTeX math, theorems and proofs, GFM tables, and a reference list carrying cite keys). Available from two buttons in the detail panel, and from the CLI as lumencite export-lcir <id|key> [--format json|md] [--source tex|pdf].
  • LCIR — figures from PDFs (Phase 8a, experimental) — detects figure regions on each page, saves a cropped PNG per figure, records them as figure nodes, and links each one to its caption (caption_of). Adds migration 0019 (assets, node_assets).
  • LCIR — structured tables from TeX (Phase 8b, experimental) — parses tabular / tabular* / tabularx into a row-by-cell grid (table nodes) with column spec, alignments, \multicolumn spans and rules, keeping LaTeX inside cells intact. Markdown export renders these as GFM pipe tables.
  • LCIR — figure alt text via LLM Vision (Phase 8c, opt-in, incurs API cost) — describes figure crops with a Vision model and stores the result in node_alt_texts. Gated behind its own consent flag, lcir.vision_alt_text.enabled, which is separate from lcir.enabled and also off by default; both must be on. Runs as an explicit batch (never during a build), can be scoped to a single entry or attachment, skips very small crops, and shows the target count before you start. Generated text is stored as origin='llm_inference' with a confidence and model name — it never overwrites the author's caption and never enters full-text search. Descriptions carry across extractor versions by crop fingerprint, so a rebuild does not re-bill. Adds migration 0020.
  • Rebuilding LCIR for an existing library — raising an extractor version no longer leaves old documents behind: Settings → Data gains a "rebuild outdated LCIR" batch (with progress and a guard against concurrent runs), and each attachment row in the detail panel gains a per-attachment build/rebuild button.
  • New LCIR read tools over MCPget_figures (bounding boxes, figure numbers, captions, crop paths and alt text) and get_tables (captions and cell grids), available when lcir.enabled is on.

Before you enable LCIR

  • Disk usage grows when LCIR is enabled and rebuilt. Figure crops are written as PNGs under attachments/<entry>/.lcir/ and are included in backup archives. On the development library, 888 figures came to roughly 529 MB, taking the backup set from 531 MB to 726 MB. Check free space before enabling and rebuilding a large library.
  • Set the OCR model to claude-sonnet-5 before generating alt text. Vision alt text reuses the OCR provider/model settings (llm.ocr_provider / llm.ocr_model), and description quality differs sharply between models — in our testing only claude-sonnet-5 was reliably accurate about shapes, mappings and edge labels, while smaller models misread figures or asserted relationships that were not there. For scale, generating alt text for 888 figures with claude-sonnet-5 cost about $6.
  • Default behaviour is unchanged. All lcir.* flags are off by default; full-text search, BibTeX, the Web Clipper and existing CLI commands are unaffected.

Install

  • macOSLumenCite_0.10.0_universal.dmg (signed + notarized), or via Homebrew:
    brew tap marmot1123/lumencite
    brew trust marmot1123/lumencite
    brew install --cask lumencite
    Existing macOS installs update in place via the built-in updater.
  • Windows.msi or -setup.exe (signed with a Certum Open Source Code Signing certificate). SmartScreen may still warn until download reputation builds; choose "More info" → "Run anyway".
  • Linux — AppImage / .deb / .rpm.
  • Web Clipper extension — unchanged from v0.8.0. If you already have v0.2.0 loaded, no action is needed; otherwise grab lumencite-clipper-0.2.0.zip from this release and load it unpacked (see the README "Browser extension" section).

Full changelog: v0.9.0...v0.10.0

LumenCite v0.9.0

Choose a tag to compare

@github-actions github-actions released this 23 Jul 04:19

See CHANGELOG.md for the full list of changes.

Downloads

  • macOS (Universal — Apple Silicon + Intel).dmg (signed + notarized)
  • Windows.msi / .exe installer (signed with Certum; added at release time)
  • Linux.AppImage, .deb, .rpm

Notes

  • The auto-updater is enabled on macOS (signed updates). Windows and Linux update by manual download from this page.

LumenCite v0.8.0

Choose a tag to compare

@github-actions github-actions released this 22 Jul 04:03

See CHANGELOG.md for the full list of changes.

Downloads

  • macOS (Universal — Apple Silicon + Intel).dmg (signed + notarized)
  • Windows.msi / .exe installer (signed with Certum; added at release time)
  • Linux.AppImage, .deb, .rpm

Notes

  • The auto-updater is enabled on macOS (signed updates). Windows and Linux update by manual download from this page.

LumenCite v0.7.0

Choose a tag to compare

@github-actions github-actions released this 04 Jul 15:31

See CHANGELOG.md for the full list of changes.

Downloads

  • macOS (Universal — Apple Silicon + Intel).dmg (signed + notarized)
  • Windows.msi / .exe installer (signed with Certum; added at release time)
  • Linux.AppImage, .deb, .rpm

Notes

  • The auto-updater is enabled on macOS (signed updates). Windows and Linux update by manual download from this page.

LumenCite v0.6.0

Choose a tag to compare

@github-actions github-actions released this 04 Jul 08:53
11d41f8

See CHANGELOG.md for the full list of changes.

Downloads

  • macOS (Universal — Apple Silicon + Intel).dmg (signed + notarized)
  • Windows.msi / .exe installer (signed with Certum; added at release time)
  • Linux.AppImage, .deb, .rpm

Notes

  • The auto-updater is enabled on macOS (signed updates). Windows and Linux update by manual download from this page.

LumenCite v0.5.0

Choose a tag to compare

@github-actions github-actions released this 03 Jul 04:33
12290a5

See CHANGELOG.md for the full list of changes.

Downloads

  • macOS (Universal — Apple Silicon + Intel).dmg (signed + notarized)
  • Windows.msi / .exe installer (signed with Certum; added at release time)
  • Linux.AppImage, .deb, .rpm

Notes

  • The auto-updater is enabled on macOS (signed updates). Windows and Linux update by manual download from this page.

LumenCite v0.4.0

Choose a tag to compare

@github-actions github-actions released this 29 Jun 20:26
90e2323

See CHANGELOG.md for the full list of changes.

Downloads

  • macOS (Universal — Apple Silicon + Intel).dmg (signed + notarized)
  • Windows.msi / .exe installer (signed with Certum; added at release time)
  • Linux.AppImage, .deb, .rpm

Notes

  • The auto-updater is enabled on macOS (signed updates). Windows and Linux update by manual download from this page.

LumenCite v0.3.0

Choose a tag to compare

@github-actions github-actions released this 20 Jun 01:06
b5466d7

See CHANGELOG.md for the full list of changes.

Downloads

  • macOS (Universal — Apple Silicon + Intel).dmg (signed + notarized)
  • Windows.msi / .exe installer (signed with Certum; added at release time)
  • Linux.AppImage, .deb, .rpm

Notes

  • The auto-updater is enabled on macOS (signed updates). Windows and Linux update by manual download from this page.

LumenCite v0.2.1

Choose a tag to compare

@github-actions github-actions released this 18 Jun 03:18

See CHANGELOG.md for the full list of changes.

Downloads

  • macOS (Universal — Apple Silicon + Intel).dmg (signed + notarized)
  • Windows.msi / .exe installer (signed with Certum; added at release time)
  • Linux.AppImage, .deb, .rpm

Notes

  • The auto-updater is enabled on macOS (signed updates). Windows and Linux update by manual download from this page.