v0.3.0 - Claude, Obsidian, MCP & Clean

What's New

Features

• Anthropic Claude Provider (#38): Native support for Claude models (claude-opus-4-8, claude-sonnet-4-6, claude-haiku-4-5). Use create_client("anthropic", ...) or llm="anthropic". Requires pip install 'hyperextract[anthropic]' and an OpenAI-compatible embedder.
• Obsidian Vault Export (#37): Turn any knowledge graph into an Obsidian vault with Markdown notes linked by [[wikilinks]]. CLI: he export obsidian <ka_path> -o <vault>.
• MCP Server (#40): Expose knowledge abstracts to MCP-capable assistants (Claude Desktop, IDE agents) via he-mcp. Tools: list_templates, info, search, ask (RAG), export_obsidian. Requires pip install 'hyperextract[mcp]'.
• he clean Command (#39): Remove a KA's vector index or the entire knowledge abstract with confirmation guards.
• README Refresh: New Trendshift trending badge, centered for-the-badge shields, and a "What's New / 最新动态" section.

Fixes

• Multi-chunk Embedding Mean (#36): Use true mean when averaging multi-chunk embeddings for better long-document quality.
• Embedding Batch Size Cap (#35): Cap embedding batch size for OpenAI-compatible providers to avoid provider limits.
• LLM Merge Strategies (#41): Fix silent None resolution for multi-word llm_* merge strategies (llm_prefer_incoming, llm_prefer_existing, llm_custom_rule).
• Batch None Handling (#30): Gracefully handle None results from batch() when LLM extraction fails.

Dependencies

• langchain-anthropic>=0.3.0 (optional extra hyperextract[anthropic])
• mcp>=1.2.0 (optional extra hyperextract[mcp])

Resolved Issues

• #38 Anthropic provider support
• #37 Obsidian export
• #40 MCP server
• #39 he clean command
• #36 / #35 Embedding reliability improvements
• #41 Template merge strategy fix
• #30 / #25 / #27 Batch None handling

Notes

• All new LLM/transport integrations are optional extras; the core package remains lightweight.
• Existing OpenAI, Bailian, and vLLM provider paths are unchanged.