Nassila

Changelog

Notable releases of the desktop app, with worker-themed codenames where applicable.

[1.10.1] — 2026-08-23 · Packaged network parity

Fixed

  • "Keep my title — find correct DOI" works in installed builds — the DOI↔title repair searched Crossref / PubMed / OpenAlex from the sandboxed renderer, which the production security policy blocks, so the button silently did nothing in the installed app (dev runs were unaffected). The registry search now runs in the main process with the same semantics: your title is the trusted anchor, the replacement DOI must genuinely belong to it, and only empty fields fill from the new record.
  • DOI lookup works in installed builds — the row-level "Find DOI" action and the "Find Missing DOIs" menu used the same blocked path.
  • Autocorrect's online step and the input-bar "Resolve" button work in installed builds — both were silently blocked by the same security policy; both now resolve through the main process.

Full history on GitHub

[1.10.0] — 2026-08-22 · Masdar Papers

Added

  • Bibliography dedupe — duplicate references (same DOI, or same title + year) collapse into one audit finding, with citation sites from both numbers preserved. Title-only look-alikes are flagged ambiguous instead of silently merged; Sharh-lite shows the dedupe outcome when it did work.
  • Attach papers by folder — point the Manuscript loop at a folder of source PDFs: Nassila scans it (up to 200 files, 50 MB each), matches each PDF to your references by first-page DOI or title, and after your confirmation attaches and re-audits only the matched references.
  • Targeted re-audit for multiple refs — after attaching papers, one run re-grounds exactly the selected entries while other findings stay in place.

Changed

  • Raqim Resolve works again for titles and URLs — Resolve previously returned only a placeholder card for paper landing pages and silently dropped moderate title matches. URL rows now search Crossref / PubMed / OpenAlex / DataCite, and title matches with reasonable similarity surface again.
  • Resolve panel field prefill — switching between DOI / PMID / PMCID / URL / title auto-copies that value from the row into the search box (without overwriting what you typed), and Verify shows exactly which identifier it searched.
  • Honest Wayback links — the unconditional [Wayback ↗] on every URL row is gone. The Resolve panel now links to the Wayback Machine only when a snapshot actually exists, straight to that snapshot.

[1.8.0] — 2026-08-13 · Sanad 9B

Added

  • Qwen3.5 thinking handling — stripQwenThinkingTraces runs first in the grounding JSON repair pipeline, so thinking-wrapped model output (with braces/quotes inside the reasoning) parses identically to clean JSON, and clean output passes through byte-identical.
  • Token budget — max_tokens: 2048 is sent for Sanad 9B grounding calls to prevent mid-JSON truncation from thinking traces.
  • No-thinking template on the web — the --chat-template-file template for llama.cpp users lives in the Sanad setup guide; the Custom provider preset links there.
  • Sole-tier Sanad registry — nassila-sanad-9b (9B FT-5) is the only Sanad tier: single 9B tier chip; setup modal, presets, hints, and defaults all point at the 9B model. 4B S15 / 12B S14 / E4B S12 retired (abstract-era).

Changed

  • Sanad default model — fresh installs default to nassila-sanad-9b in LM Studio / Ollama / vLLM presets.

[1.7.0] — 2026-08-10 · Integrity Bundle

Added

  • Structured DOCX manuscript import — Word documents now import with real structure preserved: paragraph blocks plus a heading side-channel (Heading1–6) for smarter segmentation. Bibliography DOCX parsing is unchanged.
  • Dirty-close warning — closing the window with unsaved work asks for confirmation before the session is lost.
  • Preflight+ mapping breakdown — Sharh-lite shows matched / ambiguous / unmatched citation counts alongside the coverage percentage.
  • Trust & packaging hardening — audit rate limiting under concurrency, mid-LLM cancel via AbortSignal, and packaged smoke gates on release trains.

[1.6.0] — 2026-08-05 · Maktab Loop

Added

  • Maktab OCR golden fixtures — generated PDF suite (text + scan pages) with byte-level goldens and a real-Tesseract OCR probe in the Windows CI packaging job.
  • Cache controls — Settings → Storage: inspect and clear the source-artifact and Maktab extraction caches.
  • Review-after-import banner — a dismissible notice in the loop when PDF import extraction needs human attention.
  • Deterministic Sharh summaries — richer aggregate coverage/passage/claim summaries with a localized headline and per-finding copy (EN/AR).
  • Masdar source-PDF attach — attach a local PDF for any cited source and re-ground just that reference (single-reference re-audit).
  • RTL acceptance pass — swept layout-direction strays to logical utilities across panels and settings.

[1.5.0] — 2026-07-29 · Raqim Web

Added

  • Maktab Tier A Rust WASM Engine — integrated @firecrawl/pdf-inspector-wasm into PDF manuscript extraction. Accelerates text-based PDF ingest by 10–20× (~10–50ms execution), extracts Markdown headings (# Intro) and vector/heuristic Markdown tables (| ... |), with fallback to pdfjs-dist and Tesseract OCR.
  • Raqim Web metadata resolver — deterministic webpage metadata resolution (webpage-metadata.ts) extracting Open Graph (og:*), Dublin Core (dc.*), Schema.org JSON-LD, and HTML meta tags.
  • Host-specific platform extractors — structured title/publisher extraction for GitHub repositories (owner/repo), Kaggle datasets, Hugging Face models/datasets, Substack, Medium, and YouTube videos.
  • Wayback Machine archive integration — automatic fallback link generation (web.archive.org/web/*/${url}) and [Wayback ↗] action on webpage URL citations.
  • IPC bridge — registry:resolveWebpageMetadata exposed via window.api for packaged Electron app parity (bypasses renderer CORS).
  • Raqim Resolve UX — "Fetch webpage metadata" button and inline Wayback Machine archive link.
  • Parser & access date support — plain-text parser extracts accessed dates and auto-classifies URL-only references as type: 'webpage'.

[1.4.0] — 2026-07-21 · Raqim Statute

Added

  • Legislation Resolve — US federal (congress.gov, govinfo.gov, uscode.house.gov), UK legislation.gov.uk, and generic .gov official catalogue URLs; EU ELI from 1.3.x retained.
  • Raqim Resolve UX — legislation-specific panel hint; provider labels for US federal, UK legislation, official catalogue, and grey-web stubs.
  • Statute import — merge statute numbers split across PDF/DOCX lines when importing bibliography plain text.
  • Masdar chunking — paragraph/page-aware excerpt selection with page hints on cite sites.
  • Preflight+ — citation-mapping coverage warnings; abstract-only and no-source flags.
  • Submission integrity bundle — export JSON from Sharh-lite (preflight + evidence summary + provenance index; no manuscript body).

[1.3.1] — 2026-07-20 · Maktab OCR hardening

Fixed

  • OCR packaging — on-device Enhanced OCR works reliably in Windows installers (canvas / Tesseract natives).
  • Character-reversed Arabic PDFs — detect broken ToUnicode encoding; prefer DOCX instead of OCR.

Changed

  • Arabic Tesseract deferred — Arabic-heavy PDFs keep embedded text and warn to use DOCX. Enhanced OCR runs eng/fra only for Latin scans until vision/LLM OCR.
  • Installer size — unused Arabic tessdata pack removed (~12 MB); Latin packs remain.

[1.3.0] — 2026-07-20 · Sharh-lite

Added

  • Faster manuscript audits — main-process scheduling with live progress and reliable cancel.
  • Quote-verification chips — see whether a supporting quote was found, separately from the Sanad claim verdict.
  • Raqim Repair — stronger verify for PMCID, arXiv, publisher URLs, and common parser edge cases; safer handling when DOI and title disagree.
  • Raqim Resolve — repair panel: search by title, DOI, PMID, PMCID, or URL; ranked suggestions from scholarly registries plus Hugging Face, Kaggle, and GitHub (you choose what to apply).
  • Masdar attach — attach your own PDF for a cited reference and re-ground that source offline.
  • Projects + Help — save/open .nassila projects; first-run bibliography tip; Help menu links to this website.
  • OCR language packs — bundled English and French for scanned Latin PDFs (Arabic pack later deferred in 1.3.1).
  • Sharh-lite — deterministic evidence summary (claim counts and suggested next actions).

Fixed

  • Windows installer / taskbar / shortcut icons show Nassila branding correctly.
  • DOCX and PDF bibliography import counts align more closely with paste (fewer merged or false-split references).

[1.2.1] — 2026-07-17 · Masdar UX

Added

  • Audit in-progress panel — cited-sources table grows during the run; detail panel stays locked until the audit finishes.
  • DOI↔title conflict — manual-only — Verify no longer auto-patches conflict rows; choose keep-title (find DOI) or keep-DOI (update title). Predatory-list sync no longer clears the yellow conflict panel.
  • Shortcuts — Ctrl/Cmd+Enter runs audit; Copy evidence and Jump to Bibliography on findings.
  • Icon polish — toast, network, dropdown, toolbar, and source-link affordances.

Deferred

  • Per-reference attach PDF and quote-verification chips → shipped in 1.3.0.

[1.2.0] — 2026-07-15 · Masdar-lite

Added

  • Open-access PDF text in grounding — when Unpaywall finds a PDF, Nassila extracts text for passage grounding (not abstract-only).
  • Incremental audit progress — cited-sources table fills as each reference completes; live N / M progress.
  • Scanned-PDF OCR (O1) — on-device OCR fallback for hard-to-extract PDFs (English / French; Arabic later deferred in 1.3.1).
  • Icon system — consistent Lucide icons for issue severity and common controls.

Fixed

  • Bibliography PDF import in development no longer fails on the pdf.js worker.
  • Duplicate Online indicator removed from the status bar (header only).

[1.1.3] — 2026-06-29 · Polish

Added

  • Notifications — in-app toasts for verify, autocorrect, export, and related actions; optional OS notification when a manuscript audit finishes in the background.
  • Sanad setup — lighter in-app modal with links to the full website guide.

Changed

  • Arabic UI glossary-aligned copy across workers, modes, and panels.
  • Bibliography busy state shows a clearer task-specific status strip.

Fixed

  • Windows app icon in taskbar/titlebar for dev and builds.
  • Network status false Offline flapping; Retry connection when Offline.
  • DOI↔title identity conflicts surface clearer resolution choices.
  • Placeholder webpage titles like () replaced from fetched metadata during autocorrect.

[1.1.2] — 2026-06-27 · Raqim Bridge

Added

  • Bibliography bridge — send manuscript References to Bibliography; audit from curated library with manuscript-ref:N cite-key preservation.

Fixed

  • Bibliography PDF import — manuscript-grade PDF text layout so numbered reference lists split correctly (DOCX parity).
  • Verify references (packaged app) — unified L1+L2 runs in the main process so verification works in release builds.
  • PDF 9. References heading — IMRAD-style numbered reference headings on PDF export.
  • L3 rollup — deduplicated insufficient-evidence reasons across cite sites.
  • Cited-sources table — opaque sticky header on scroll.

Changed

  • Loop audit detail — compact layer summary and cite-site list.

[1.1.1] — 2026-06-27 · Bibliography-first

Added

  • Bibliography-first workflow — loop hints when references should be curated before audit.
  • Bibliography DOCX import — shared manuscript-grade reference extraction.
  • Journal search (CrossRef) — online journal target search from Bibliography.

Fixed

  • DOI from https://doi.org/... URLs — plain-text parser populates DOI when only a doi.org link is present.
  • Manuscript segmentation — numbered section headings no longer swallowed as bibliography.

Changed

  • Settings clarifies Unpaywall email is not university paywall login.

[1.1.0] — 2026-06-27 · Sanad

Added

  • Manuscript loop — upload DOCX/PDF, run full audit (L1 registry + L2 metadata + L3 passage grounding), cited-sources table, and evidence detail.
  • Passage grounding (Sanad) — local runners (LM Studio, Ollama, vLLM, Custom) and Cloud API; E4B / 12B tier chips; Set up Sanad guide modal.
  • Settings → General → Manuscript source fetch — one-time Unpaywall email for OA full-text lookups.
  • L1 multi-registry fallback — DOI: Crossref/DataCite → OpenAlex → PubMed; PMID: PubMed → OpenAlex.
  • DOCX references fallback — numbered bibliography block when no References heading.

Changed

  • Hydra worker tabs removed — primary surfaces are Manuscript loop + Bibliography; Tasnif / Sharh inline in loop detail.
  • External Marker PDF CLI removed — PDF ingest uses bundled pdf.js only.
  • OA fetch — improved Unpaywall location handling for open-access full text.
  • L3 grounding engine — JSON repair, retry on parse failure, passage/excerpt caps.

[1.0.1] — 2026-06-03

Fixed

  • More menu closes when clicking outside or choosing an action.
  • Import hints dismiss when the pointer leaves the button.
  • Vancouver — citation cards show six authors then “et al.” when there are seven or more.

[1.0.0] — 2026-05-24

Added

  • Desktop app: import and validate bibliographies (BibTeX, RIS, CSL-JSON, plain text, DOCX/PDF extraction where supported).
  • Unified registry verify (Crossref, PubMed, OpenAlex): L1 resolution + L2 metadata alignment (up to 200 rows per run).
  • Autocorrect, duplicate detection, predatory-journal flags, CSL export (thousands of styles).
  • English and Arabic UI with bilingual product positioning.

Full history on GitHub