●Change8

v0.12.44

📦 llamaindexView on GitHub →
✨ 10 features🐛 11 fixes⚡ 2 deprecations🔧 18 symbols

Summary

This release introduces several new integrations including IBM Db2 and OpenAI Realtime Conversation, while deprecating older agent architectures and query pipelines. It also includes significant bug fixes for ReAct agents and expanded batching support for embedding providers.

Migration Steps

  1. Update environment variables for Perplexity LLM to use PPLX_API_KEY.
  2. Review usage of older agent architectures and query pipelines in core as they are now marked for deprecation.
  3. If using Pinecone, ensure compatibility with the new v7 support.

✨ New Features

  • Added CachePoint content block for caching chat messages in core.
  • Added batch support for FastEmbed embeddings.
  • Added async batching for HuggingFace embeddings using asyncio.to_thread.
  • Added visualization functions for single/multi agent workflows.
  • Added IBM Db2 vector store integration.
  • Refactored DuckDB VectorStore.
  • Added support for Pinecone v7.
  • Added beta OpenAI Realtime Conversation integration for voice agents.
  • Added cross_encoder_kwargs parameter to SBertRerank for advanced configuration.
  • Enabled forwarding of arbitrary Azure Search SDK parameters in AzureAISearchVectorStore.

🐛 Bug Fixes

  • Fixed ReAct system header formatting and output parsing when 'Thought:' prefix is missing.
  • Fixed string stripping in ReAct output parser.
  • Fixed system prompt handling for CodeAct agent.
  • Fixed potential serialization issue in AgentStream event by excluding raw field.
  • Updated kwargs for Anthropic Bedrock.
  • Fixed system instruction setup for Google GenAI client.
  • Fixed image URL handling in Mistral AI.
  • Updated Perplexity LLM to use PPLX_API_KEY environment variable.
  • Fixed runtime modification of 'top_n' attribute in Bedrock Rerank.
  • Fixed type handling for IN/NIN filters in Qdrant.
  • Fixed Python version constraints for Pinecone vector store.

🔧 Affected Symbols

CachePointReActOutputParserCodeActAgentAgentStreamFastEmbedEmbeddingHuggingFaceEmbeddingAnthropicBedrockGoogleGenAIMistralAIPerplexityBedrockRerankSBertRerankAzureAISearchVectorStoreDb2VectorStoreDuckDBVectorStorePineconeVectorStoreQdrantVectorStoreOpenAIRealtimeConversation

⚡ Deprecations

  • Older agent architectures in llama-index-core are now deprecated.
  • Query pipelines in llama-index-core are now deprecated.