●Change8

v0.14.15

Breaking Changes
📦 llamaindexView on GitHub →
⚠ 1 breaking✨ 17 features🐛 9 fixes🔧 8 symbols

Summary

This release introduces significant multimodal support across core components, integrates the AgentMesh agent layer, and includes numerous updates for LLM providers like Anthropic and Bedrock. Several packages also removed deprecated connection parameters.

⚠️ Breaking Changes

  • llama-index-embeddings-ibm and llama-index-llms-ibm removed support for the `persistent_connection` parameter in their respective clients/initializers. Users must update their initialization logic to remove this parameter.

Migration Steps

  1. Remove the `persistent_connection` parameter when initializing clients in llama-index-embeddings-ibm and llama-index-llms-ibm.

✨ New Features

  • Integration of AgentMesh: Trust Layer for LlamaIndex Agents.
  • Support for basic operations for multimodal types in llama-index-core.
  • Recursive LLM type support added to llama-index-core.
  • Multimodal template variable formatting support added to llama-index-core.
  • Multimodal prompt templates support added to llama-index-core.
  • Multimodal chat prompt helper added to llama-index-core.
  • Retry and error handling added to BaseExtractor.
  • Support for Claude Sonnet 4.6 added to llama-index-llms-bedrock-converse.
  • Addition of Sonnet 4-6 models in llama-index-llms-anthropic.
  • Support for new OCI DataScience endpoint /predictWithStream for streaming use cases in llama-index-llms-oci-data-science.
  • Flattening of dicts for improved OpenTelemetry data serialization.
  • Support for custom span processor and refactoring to use llama-index-instrumentation instead of llama-index-core in llama-index-observability-otel.
  • Sandbox LLM-generated code execution in EvaporateExtractor.
  • Enhancement to GitHubRepoReader allowing selective file fetching and deduplication.
  • Addition of LayoutIR reader integration.
  • Pagination support added for Microsoft Graph API calls in SharePoint reader.
  • Propagation of `partial_params` to `get_tools_from_mcp` utils.

🐛 Bug Fixes

  • Redundant `metadata_seperator` field removed from TextNode.
  • Mock prompt type updated in `mock_prompts.py` for tests.
  • Ensured at least one message/content block is returned by the old memory implementation.
  • Fixed thinking_delta not being populated in all chat modes for bedrock-converse.
  • Fixed mutable default argument usage in `load_all_file_paths` for llama-index-readers-bitbucket.
  • Hotfix for output_dir crash and Block extraction in layoutir reader.
  • Restricted `requires-python` to >=3.12 for llama-index-readers-layoutir to match dependency requirements.
  • WhatsAppChatLoader updated to retrieve DataFrame in pandas format.
  • Removed ORM Collection mix-usage with MilvusClient in Milvus vector store.

Affected Symbols

TextNodeBaseExtractormock_prompts.pyGitHubRepoReaderLayoutIR readerWhatsAppChatLoaderMilvusClientFaissMapVectorStore