paku v1.1.0

New features

• Semantic recommendations: GET /api/recommendations powered by local Ollama LLM (gemma4:26b). Analyses your
  collection context (top rated, genres, studios, formats, watch status) and returns personalized suggestions resolved
  against AniList. SHA256 cache avoids re-querying for unchanged state. ?refresh=true forces regeneration.
• Notion CSV import: paku import-notion <csv> parses your Notion anime database export, maps 12 status values
  to canonical user_statuses, strips Notion page URLs from genre/studio fields, and merges watch statuses + personal
  scores into the dashboard. --dry-run previews matches without writing.
• Dashboard Recs tab: "For You" (Ollama personalized) and "Similar to…" (AniList graph, seed selector). Skeleton
  loading, Refresh, Add buttons. 5 tabs total.
• Branding: SVG logo, refined dark palette, status chip colors (green=Completed, blue=Watching, amber=Plan to
  Watch, red=Dropped, gray=On Hold).

Removed

• LangExtract: removed from config, CLI, docs, and tests. Ollama VLM proved sufficient on the 1287-image batch
  run.

Fixes

• Anime extractor: Instagram caption opinions ("NGL THIS ANIME IS UNDERRATED:") no longer extracted as titles.
  Sauce prefix stripping for comment-thread titles. Garbage filter runs before fallback ranking. "Join the
  conversation..." added to chrome filter.
• Ollama config split: ocr_model (gemma4-paku:latest, VLM) and recs_model (gemma4:26b, text LLM) with
  backward-compat fallback. Prevents VLM system prompt from corrupting text generation.
• Dashboard cache: Cache-Control: no-cache on root page prevents stale inline JS after updates.

Full Changelog: v1.0.1...v1.1.0

Install

pip install paku

Requires a config.yaml with a Google Cloud Vision key (see config.yaml.template).

License
Mozilla Public License 2.0

paku v1.0.1

What's changed

• AniList recommendations panel in Collection tab ("Recommended for you" — seeded from most-recently-added entry with an AniList ID)
• GET /api/collection/{id}/recommendations endpoint with has_anilist_id() saved-flag lookup
• Fixed _detect_multi_titles() to use finditer for wrapped carousel titles (e.g. "Si-Vis: The Sound of Heroes")
• CI publish job: auto-publish to PyPI on v* tag push via OIDC Trusted Publishing

Full Changelog: v1.0.0...v1.0.1

paku v1.0.0

CLI pipeline that extracts anime titles, GitHub/web URLs, and recipes from Instagram screenshots into structured JSON, CSV, and a local dashboard.

What's in v1.0.0

Three extractors

• Anime — AniList GraphQL enrichment, Levenshtein gating, multi-title carousel detection, 507/513 screenshots AniList-matched
• URL — direct URL + project keyword resolution, 4-tier confidence scoring
• Recipe — ingredient parsing with qty/unit split, structured JSON + CSV output

Pipeline

• Batch processing with checkpoint/resume (--resume)
• Auto-classifier routes screenshots to the right extractor
• Ollama VLM smart re-run (--smart) for low-confidence OCR results
• review_queue.json captures everything below threshold — nothing discarded

Dashboard

• Local web app (paku serve) — browse, filter, and triage extracted anime entries
• SQLite-backed, FastAPI + vanilla JS SPA

Quality

• 513 unit tests, 2 skipped (integration)
• Ruff lint clean across all modules
• GitHub Actions CI: lint + test matrix (Python 3.11 & 3.12) + wheel build

Install

pip install paku

Requires a config.yaml with a Google Cloud Vision key (see config.yaml.template).

License
Mozilla Public License 2.0