●Change8

v0.13.0

Breaking Changes
📦 llamaindexView on GitHub →
⚠ 3 breaking✨ 10 features🐛 5 fixes🔧 16 symbols

Summary

This major update to llama-index-core (v0.13.0) removes legacy agent classes and QueryPipeline in favor of a new Workflow-based agent architecture, while updating various provider integrations for LLMs and vector stores.

⚠️ Breaking Changes

  • Removed deprecated agent classes (FunctionCallingAgent, old ReActAgent, AgentRunner, step workers, StructuredAgentPlanner, OpenAIAgent). Users must migrate to workflow-based agents.
  • Removed QueryPipeline class and all associated code.
  • Default index.as_chat_engine() now returns CondensePlusContextChatEngine instead of agent-based engines. Agent-based chat engines are removed from this method.

Migration Steps

  1. Replace usage of FunctionCallingAgent, OpenAIAgent, and AgentRunner with FunctionAgent, CodeActAgent, or AgentWorkflow.
  2. Replace usage of the old ReActAgent with the new workflow-based ReActAgent.
  3. Remove dependencies on QueryPipeline as the class has been deleted.
  4. If an agent-based chat engine is required, manually instantiate one of the new agent classes instead of calling index.as_chat_engine().

✨ New Features

  • Introduced new workflow-based agents: FunctionAgent, CodeActAgent, ReActAgent, and AgentWorkflow.
  • Updated mixedbread embeddings and rerank for latest SDK.
  • Added Thought Summaries and signatures for Gemini LLMs.
  • Added support for kimi-k2-instruct (NVIDIA) and solar-pro2 (Upstage) models.
  • Enhanced Github Reader with file filtering and custom processing.
  • Added region_name support via client_kwargs in S3Reader.
  • Added Gemini Live beta implementation.
  • Added get nodes and delete nodes support for AstraDB vector store.
  • Added partition_names support in Milvus search configuration.
  • Added support for ANY/ALL postgres operators in Postgres vector store.

🐛 Bug Fixes

  • BaseDocumentStore no longer returns Nones in results.
  • Fixed FunctionTool parameter doc parsing and signature mutation.
  • Handled empty prompt in MockLLM.stream_complete.
  • Ensured wrapped exceptions bubble up in llama-index-instrumentation.
  • Reduced metadata keys in S3VectorStore to save space.

🔧 Affected Symbols

FunctionCallingAgentReActAgentAgentRunnerStructuredAgentPlannerOpenAIAgentQueryPipelineindex.as_chat_engineCondensePlusContextChatEngineFunctionAgentCodeActAgentAgentWorkflowBaseDocumentStoreFunctionToolMockLLM.stream_completeS3ReaderS3VectorStore