●Change8

v0.12.50

Breaking Changes
📦 llamaindex
⚠ 2 breaking✨ 8 features🐛 10 fixes🔧 10 symbols

Summary

This release introduces several new integrations including Cloudflare AI Gateway, Service-Now, and S3 vector stores, alongside Llama 4 support in Bedrock. It also includes security improvements to cache directory handling and various bug fixes for agents and retrievers.

⚠️ Breaking Changes

  • The `get_cache_dir()` function now uses a more secure default location, which may change where your application stores or looks for cached data.
  • NLTK cache is now stored in a private folder, potentially requiring re-downloading of NLTK data if you relied on global paths.

Migration Steps

  1. If you rely on specific cache locations, verify the new path used by `get_cache_dir()`.
  2. Update Bedrock Converse implementations if using Llama 3.2 1b/3b for function calling, as these are no longer listed as supported for that feature.

✨ New Features

  • Added HTML table extraction support in MarkdownElementNodeParser.
  • Introduced Cloudflare AI Gateway integration (llama-index-llms-cloudflare-ai-gateway).
  • Added Google Search Tool support to GoogleGenAI LLM integration.
  • Introduced Service-Now reader (llama-index-readers-service-now).
  • Added S3 vector store support (llama-index-vector-stores-s3).
  • Added Llama 4 model support in Bedrock Converse.
  • Improved instrumentation span naming for better observability.
  • Added start event handling to workflow agents to support running as MCP servers.

🐛 Bug Fixes

  • Resolved race condition in SQLAlchemyChatStore using precise timestamps.
  • Fixed document store import in DocumentContextExtractor to use BaseDocumentStore.
  • Improved empty retrieval checks in vector index retrievers.
  • Fixed ID type mismatches across various retrievers.
  • Fixed structured output issues in multi-agent secondary constructors.
  • Removed duplicated session_id metadata filters in VectorMemoryBlock.
  • Ensured agents stop correctly when a function tool returns direct.
  • Updated ReAct agent parse error messages for clarity.
  • Removed load_kwargs from WikipediaToolSpec.load_data.
  • Fixed namespace resolution in Moorcheh Vector Store.

🔧 Affected Symbols

MarkdownElementNodeParserget_cache_dirSQLAlchemyChatStoreDocumentContextExtractorBaseDocumentStoreVectorMemoryBlockGoogleGenAIWikipediaToolSpec.load_dataBaiduVectorDBMoorchehVectorStore