Skip to content

CHANGELOG.md

Latest commit

 

History

History
216 lines (179 loc) · 17.4 KB

CHANGELOG.md

File metadata and controls

216 lines (179 loc) · 17.4 KB
Raw

Changelog

All notable changes to PaperLens will be documented in this file.

[Unreleased]

Added

  • CrossRef metadata cache (core/work_cache.py): the post-search DOI enrichment queried CrossRef once per incomplete paper (~150 per search, ~2s each) and repeated the same lookups across searches. Now cached per DOI (30-day TTL, thread-safe, persisted to ~/.paperlens/crossref_work_cache.json, capped at 20k entries). Pure speed: cached payloads are byte-identical to live responses (verified 11/11 DOIs field-by-field against live CrossRef).
  • Citation graph redesign (three-column lineage): new CitationGraphV2 — referenced / current / citing three-column lineage layout (from the reviewed prototype), Top 10/20/all, flat / by-year / cluster arrangements (year bubbles expand on click), anchored physics drag, equal-ratio zoom, "+N more · expand all". The AI toolbar (summary / detail / novelty / export) is preserved. Floating mini-graphs keep the legacy renderer.
  • Keyword co-occurrence / author collaboration graph redesign: new NetworkGraphV2 renderer — three deterministic layouts (circular / concentric / cluster), anchored physics (drag snaps back to its anchor), equal-ratio wheel zoom, community detection (weighted label propagation) for coloring, dashed edges for cross-cluster links, hover neighbor highlight with click-to-pin, and a segmented toolbar (data / arrange / count / fit). Replaces the temporary rendering path where these two networks were drawn by the citation graph class.
  • /api/keyword-network and /api/author-network now return the top 120 nodes by frequency (was: ad-hoc ≥2 filter / top 20-30), so the frontend Top-N control can switch without re-requesting.

Changed

  • Faster search, no quality trade-off: DOI enrichment concurrency 10 → 20 (CrossRef's public pool allows 50 req/s; this stays at ~13), connection-pooled requests.Session instead of bare requests.get (eliminates ~150 TCP/TLS handshakes per search), 429 backoff retry, and a mailto: User-Agent when any source has an email configured (polite pool, no rate limit). Measured on "deep learning microscopy" / 6 sources: 26.1s cold → 19.7s warm → ~15.5s with a fully warm cache (was 30.8–40.2s). Nothing was removed — the enrichment fills title/authors/journal/abstract, which drive dedup, IF sorting and relatedness filtering, so cutting it would trade away result quality.

Fixed

  • PubMed results carried other papers' DOIs (P0, silent data corruption): the XML parser used article.findall(".//ArticleId") and took the last IdType="doi" match — but ReferenceList/Reference/ArticleIdList uses the same element name, so the recursive search swept in every reference's IDs and the last reference's DOI overwrote the article's own. Measured 25–40% of PubMed results: e.g. PMID 35594500 ("Lipid-Nanoparticle-Based Delivery of CRISPR/Cas9") was tagged with 10.3390/vaccines9010065 (an mRNA vaccine paper). Impact: original-link/PDF buttons, exported RIS/BibTeX, citation graphs and DOI-based dedup all pointed at the wrong paper. Now scoped to ./PubmedData/ArticleIdList (the article's own IDs) with an ELocationID fallback. After the fix: 0/10 mismatches (was 3/3), and title duplicates dropped from 8 groups to 0 — duplicates were a downstream symptom of DOIs being corrupted, which defeated dedup.
  • Missing import os in search_engine.py — SearchEngine.__init__ crashed with NameError (introduced with the online journal-metrics fallback).
  • Author search returning irrelevant results ("搜某作者的论文全是无关结果") — three-layer fix:
    1. Query understanding: added the most common patterns that were silently unrecognized — "张伟的论文" (no verb), bare Chinese names ("周金华"), "papers by X" (full surname was truncated to given name), "Zhang Wei's papers", honorifics (王五教授的论文), "写的"; blacklist switched to substring matching; lookbehind prevents mid-word splits ("深度学习" sawed into "度学习" evading the blacklist).
    2. Dispatch: /api/search (keyword mode) never ran author detection at all — only the AI-search path did — and engine.search's source dispatcher dropped the author parameter entirely, so the author name was sent to sources as a full-text keyword. Keyword mode now runs the same detection; the dispatcher routes the author name as the query subject (OpenAlex native author-ID lookup with raw_author_name.search fallback for Chinese names — 27k hits; PubMed [au] for Latin names; CNKI/Wanfang get the original Chinese name).
    3. Result gate: au results whose author list has zero name match are demoted to the end (not dropped). The relatedness filter already exempted au mode — it was the "22 filtered → 0 cached" killer when field arrived wrong.
  • Author + topic combined queries ("张伟发表的CRISPR研究"): the extracted author was never applied to topic-search results, and deduplicate_papers reshuffled the ranked order into a "DOI group + title group" concatenation, silently destroying any upstream ordering. Now: author detection feeds a merged sort (author-hit first, then BM25 relevance), dedup preserves upstream order, and pypinyin-based variants let Chinese names match Latin signatures ("张伟" ↔ "Wei Zhang"/"Zhang W", 复姓 "欧阳文清" ↔ "Wenqing Ouyang"; full-name vs initials disambiguation keeps 张伟兴/张伟华 out of 张伟's results). pypinyin added to the project venv.
  • AI translation fallback never worked: translator called chat(prompt, max_tokens=100) but AIAssistant.chat has no max_tokens parameter — TypeError on every call, silently degrading to dictionary-only translation.
  • Citation graph reopening after viewing a network graph kept legendCenterRow hidden (row visibility was set but never restored).
  • OpenAIRE data source (46th): EU open-access aggregation, 1M+ research products via Graph API /graph/v1/researchProducts. Complements CORE/DOAJ with European institutional repositories and theses.
  • Two-layer impact factor: built-in official JCR value first, OpenAlex 2yr_mean_citedness estimate as fallback (free, no key, tens of thousands of journals). Estimated values are marked with * in the UI and carry impact_factor_src=openalex.
  • 保留未收录 (keep unknown) toggle for impact-factor filter — previously papers without an IF were silently dropped, making the filter behave as a 102-journal whitelist.
  • 中科院分区「未收录」option — the CAS table contains 102 entries (1区 92 / 2区 9 / 3区 1 / 4区 0), so selecting 4区 always returned 0 results. Added an explicit __unknown__ bucket plus a coverage hint icon.

Fixed

  • 12 dead source switches — UI checkboxes existed and badges (CARSI/OA) were shown, but the flags were never sent in the search request. Refactored to a single source of truth SOURCE_SWITCHES in search.js, shared by both the normal and AI search paths.
  • Filter cache key collision — the cache stores filtered+sorted results, but the key omitted journal/field/mesh_term/pub_type/IF/CAS/OA/affiliation/author/sort/max_results. Changing any of these returned stale results. Extracted _compute_filter_seed() as the shared single source.
  • Title-only relevance check — relevance scoring only looked at the title (and had a DOI-substring bypass), so papers matching only in the abstract were treated as irrelevant. Now matches title OR abstract (title-only when the query field is explicitly ti).
  • Author field override — a residual topic word (e.g. 张伟发表的CRISPR研究) was force-passed as field=au, turning CRISPR into an author name. Now field=au is only forced when no residual topic word remains.
  • OpenAlex ISSN lookup — filter=issn: requires the hyphenated form (2156-7085); the normalized form (21567085) returned 0 results, silently disabling the whole IF fallback layer.
  • OpenAIRE DOI selection — an item can carry several doi PIDs (publisher DOI first, then the 10.48550/arXiv... preprint DOI). The old loop let the last one win, breaking dedup, citation lookup and DOI links. Preprint DOIs are now deprioritized.
  • OpenAIRE junk entries — EU project documents with no DOI/journal/authors leaked into results; added a quality floor (must have DOI, or journal+year, or authors+year).

Changed

  • Data-source count unified across README / README.zh-CN / ROADMAP: 46 retrieval entries = 29 independent upstream indexes + 17 CrossRef publisher slices (37 enabled by default). Previously README said 43 and ROADMAP said 45.

[1.3.0] - 2026-07-06

Added

  • 13 New Data Sources: DBLP (CS conferences), bioRxiv/medRxiv (preprints), AGRIS (agriculture), ACS Publications, Optica, IOP Publishing, AIP Publishing, RSC Publishing, Europe PMC, Springer Nature, Wiley, IEEE, Project MUSE — total now 26+ sources
  • Liquid Glass Effect: iOS-style frosted glass UI with CSS backdrop-filter across all major components
  • AI Result Count: moved to panel header for better visibility
  • Default Search Range: 10 years default for better initial results

Fixed

  • Search Stability (34 bugs): comprehensive audit fixing pub_type parameter, author search (OR queries, PubMed format conversion, OpenAlex filters), year filtering, journal extraction fallback
  • AI Search: stream output marker display, truncated JSON output, result count selector error
  • History Dropdown: height calculation, overlap with AI panel, click behavior (fills search box only, no auto-search)
  • CNKI Search: browser instance reuse to avoid repeated captcha verification
  • OpenAlex: journal URL length limit
  • Zotero: status check now directly calls backend test API
  • Liquid Glass: removed broken SVG filter causing purple tint and content blur

Changed

  • ROADMAP updated to v2.1 with search stability and UI optimization focus
  • README simplified — removed implementation details from roadmap section

[1.2.3] - 2026-06-09

Added

  • OpenAlex API Key support — OpenAlex now requires API key (free but mandatory). Without key: $0.10/day budget; with free key: $1/day (10x more). Added API key input field in settings panel with test button.
  • Semantic Scholar API Key support — Added config entry and frontend UI for Semantic Scholar API key. Without key: 100 req/5min; with key: 1 req/sec.

Changed

  • Updated OpenAlex search, citation enrichment, and DOI lookup to include API key in all requests
  • Updated server.py to pass API key for all direct OpenAlex API calls (citation graph, related papers, recommendations)
  • Added semantic_scholar section to config.yaml with enabled and api_key fields

[1.2.2] - 2026-06-09

Fixed

  • i18n: Backend error messages now use error codes — all API error responses use language-independent error codes (e.g., no_query, paper_not_found), frontend translates them via te() function
  • i18n: AI assistant error messages — AI error responses now use AI_ERROR: prefix with error codes, frontend translates them via teAI() function
  • i18n: Default collection group name — fixed Temporal Dead Zone bug where collectionsData referenced LANGS before initialization; default group now uses t('defaultGroup') when rendering
  • i18n: PDF download error detection — error checks now use error codes instead of matching translated text strings

Changed

  • Added 20+ error code translations to both Chinese and English LANGS dictionaries
  • Added te() helper function for translating backend error codes
  • Added teAI() helper function for translating AI error responses

[1.2.1] - 2026-06-09

Added

  • New Data Sources: Google Scholar (experimental), CNKI, Wanfang, VIP — all with β badge
  • Keyword Co-occurrence Network: visualize keyword relationships across search results
  • Author Collaboration Network: visualize co-authorship patterns
  • EndNote XML Export: additional export format for Mendeley/EndNote direct import
  • Missing i18n keys: noOALink, analyzingResearchGaps, noDOILoadGraph, querying, query
  • applyLang() now updates all citation graph panel buttons, hints, and legend text dynamically

Fixed

  • i18n: English version had hardcoded Chinese strings — fixed 15+ untranslated Chinese text in JS (error messages, status text, button labels)
  • i18n: Citation graph overlay buttons broken — static HTML used ${t(...)} template literals that don't evaluate in HTML context; replaced with static text + applyLang() dynamic updates
  • i18n: English prompt mixed Chinese — "compare" mode AI prompt contained Chinese text in English context
  • i18n: English translation used Chinese punctuation — enterGroupName used Chinese colon ： instead of :

[1.2.0] - 2026-06-09

Added

  • Chrome Extension: DOI quick-lookup and paper collection while browsing academic websites
    • Auto-detect DOIs on any webpage (meta tags + regex matching)
    • One-click paper lookup via PaperLens API
    • One-click collection to PaperLens
    • Real-time sync notification after collection
  • Collection Management: delete collection groups with confirmation dialog
    • Default collection group cannot be deleted
    • Deleting a group removes all items in it
    • Double confirmation before deletion
  • i18n Improvements: comprehensive translation for all UI elements
    • Citation graph buttons (fullscreen, close, AI analysis, export)
    • Collection panel (refresh, delete group, new group)
    • AI analysis panel (copy, fullscreen, status messages)
    • Graph hints and legend text

Fixed

  • Collection data now auto-refreshes when opening collection panel
  • Periodic polling (every 30 seconds) for collection data sync
  • Chrome extension notifies frontend after successful collection

Changed

  • Added DELETE /api/collections/groups endpoint for collection group deletion
  • Frontend uses translation functions for all dynamic content

[1.1.0] - 2026-06-08

Added

  • Local Paper Collections: star papers to save, custom collection groups, collections sidebar panel
  • Floating Citation Graph: new graph opens as draggable floating window with full toolbar (AI summary/detail/novelty/export)
  • Novelty Detection: AI analysis mode that identifies research gaps, unexplored areas, and promising directions
  • Fullscreen toggle for citation graph with zoom state preservation
  • Reset zoom button in citation graph legend
  • Zoom level percentage display in graph legend
  • Click overlay background to close citation graph
  • Markdown bold (**text**) rendering in AI analysis results

Removed

  • Related Work Discovery: removed due to persistent layout issues (nodes clustering together)

Fixed

  • "Open in new graph" no longer closes the current citation graph
  • Floating graph toolbar now shows automatically after loading
  • Citation graph expand: nodes no longer cluster — deterministic fan-shaped placement outside existing ring
  • Citation graph zoom now centers on mouse cursor position
  • Paper names fully display as zoom level increases (dynamic label length)
  • Main content area no longer hidden behind status bar (added bottom padding)
  • Graph panel fullscreen toggle with smooth CSS transition
  • D3 data join key functions prevent node/link misalignment on expand
  • Zoom behavior cleanup on graph reopen prevents duplicate handlers

Changed

  • Project repositioned from "Search Platform" to "AI-Powered Research Assistant"
  • README.md and README.zh-CN.md fully rewritten with architecture diagram
  • AI analysis prompt for "novelty" mode with 5 dimensions (both Chinese and English)

[1.0.2] - 2026-06-08

Added

  • Interactive citation graph with D3.js force-directed layout
  • Zoom, pan, and drag support in citation graph
  • Click nodes to expand citation relationships dynamically
  • Toolbar in citation graph: AI analysis, export, download, original link
  • AI analysis results displayed within citation graph panel
  • /api/paper-by-doi endpoint for single paper lookup by DOI

Fixed

  • Search button stuck on "searching" after results loaded
  • Language switch not updating search result labels
  • Default year_to in config.yaml (2025 → 2026)

[1.0.1] - 2026-06-08

Fixed

  • Fixed create_window() icon parameter error in pywebview
  • Fixed os.sys reference error in server.py
  • Fixed _last_keywords not initialized in OpenAlexSearch
  • Fixed max_tokens set to 384000 (now 4096)
  • Fixed hardcoded year 2026 (now dynamic)
  • Removed dead code get_icon_path() in main.py
  • Moved import json to top level in server.py

Added

  • Copy button for AI analysis results
  • Retry button on network errors
  • Reset settings to default button
  • Better error messages for PDF download failures

[1.0.0] - 2026-06-08

Added

  • PubMed + OpenAlex dual-source search with automatic deduplication
  • AI-powered natural language search (Chinese and English)
  • AI paper analysis: summary, detailed analysis, and multi-paper comparison
  • Streaming output for AI analysis
  • Dual AI model configuration (separate models for search and analysis)
  • Batch DOI import
  • RIS, BibTeX, and CSV export
  • Advanced filtering: journal, field tags, year range, publication type
  • In-result sorting by citations, date, and title
  • Search history with server-side persistence
  • User preference persistence (year range, sort, filters)
  • Dynamic placeholder with rotating suggestions
  • Clickable suggestion chips
  • OA paper PDF link verification
  • Settings panel with proxy configuration
  • XSS protection throughout
  • Escape key to close modals
  • Cross-page checkbox persistence
  • Auto-detection of current year for date range
  • Standalone Windows executable via PyInstaller