Skip to content

anymd 8.3.0

Choose a tag to compare

View all tags
@github-actions github-actions released this 01 Oct 11:08
v8.3.0
d4b5074
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Minor Changes

  • Add an opt-in local doc-VLM OCR route using PaddleOCR-VL-1.6 and PP-DocLayoutV3 on Candle, including Metal on Macs. anymd setup ocr installs SHA-256-pinned weights; ocr accepts auto, vlm and tesseract alongside the existing MCP booleans. Automatic OCR never downloads models. The worker has a hard page deadline, a per-region token cap, decode-time repetition stopping, shared OCR admission, and a fail-fast aggregate PDF deadline. Tables become Markdown and formulas become LaTeX. Experimental CPU q8/q4 decoder quantization is available through ANYMD_OCR_QUANTIZATION; Linux arm64 checks actual FP16 hardware and keeps the plain CLI on older machines.

  • Local transcripts now use bundled transcribe-cpp and one Qwen3-ASR-1.7B Q8 model for every language; the Whisper engine is removed. Pinned weights download only when explicitly requested and are SHA-256 verified. Long audio uses bounded 20-second chunks with source-relative segment timestamps. Optional Qwen3-ForcedAligner via standalone CrispASR supplies validated word timestamps where available. download_asr_model / --download-asr-model replaces the old spelling, which remains a Qwen-only compatibility alias. Japanese trails whisper-turbo on the 200-utterance FLEURS sample (5.93 vs 4.80 raw CER; 5.56 vs 4.59 with symmetric numeral/kana normalization), an accepted one-model trade-off. ASR benchmark tables publish both raw and explicitly named normalized scoring; they do not claim to reproduce Qwen's official scores.

  • New outline MCP tool and anymd outline <file> CLI command return a local, deterministic heading tree for PDF, DOCX, PPTX, EPUB, HTML and Markdown, as JSON or tree text. Nodes carry stable ids, title paths, page/slide/chapter ranges, Markdown byte ranges and child counts. read gains node (--node on the CLI), with the existing page selections, token budgets and cursors. Literal and ranked search hits carry node ids and title paths. Reads without a node keep their output unchanged.

  • The platform wheel includes a thin Python API: from anymd import convert returns a Document with Markdown text and source metadata. Optional langchain and llamaindex extras provide AnyMDLoader and AnyMDReader, using the same native converter without downloading another binary.

Assets 20
Loading