Changelog

All notable changes to smelt are documented here.

0.1.0 — 2026-03-05

First public release.

Universal convert tool — single entry point for PDF, DOCX, PPTX, XLSX, images, audio, web pages, YouTube, HTML, EPUB, and text files. 10 additional per-type tools available via expose_specialized_tools config
8 backends with automatic fallback: MinerU, Docling, PyMuPDF, Trafilatura, faster-whisper, YouTube transcript API, BeautifulSoup, Pandoc
4 output formats: Markdown, XML, JSON, DocTags — all tools accept a format parameter
Conversion cache with LRU eviction and 4 MCP resource endpoints (list, full text, metadata, per-page)
Image handling modes — reference, base64, or discard — configurable per format
Batch conversion — convert all supported files in a directory in one call
Backend and pagination controls — choose backend, page ranges, and truncation limits per request
YAML configuration with env-var overrides (SMELT_{SECTION}_{KEY}) and multi-path search
Docker support — CPU and GPU compose profiles with pre-built Dockerfiles

Fix exception type mismatch between pipeline and backends (proper NoBackendError)
Fix empty table crash when all cells are blank
Fix heading level parsing in backends
Fix cell index tracking for inline tables
Add InternalDoc.validate() to catch malformed documents early

Full CI pipeline: lint → typecheck → test → publish (Python 3.11–3.13 matrix)
81 test files covering unit, backend, formatter, integration, and adversarial edge cases
ruff + mypy strict mode
GPL-3.0-or-later license