LlamaIndex changelog: what's new each month
Every stable LlamaIndex release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 12 months covered; the current month updates daily.
November 2026 · so far
1 release: 0.14.26
November brought critical security updates across LlamaIndex integrations, with Python dependency requirements raised to address known vulnerabilities.
- Python version requirement increased across integrations for security fixes—check your environment compatibility
Highlights
- Python security dependencies updated across all integrations
- Required Python version raised for security compliance
Fixed
- Security vulnerabilities addressed through Python dependency updates
September 2026
1 release: 0.14.25
September focused on security hardening with a major round of vulnerability fixes across the codebase.
Highlights
- Resolved numerous security alerts to improve overall system safety
Fixed
- Security alerts addressed across multiple components
August 2026
1 release: 0.14.24
August brought a focused release addressing edge cases and correctness issues across core indexing and evaluation components.
Highlights
- Fixed tool parameter handling to not incorrectly mark *args and **kwargs as required.
- Corrected MMR embedding search to respect zero threshold values.
- Improved document node preservation in ingestion pipeline upserts.
Improved
- HTMLNodeParser now skips empty nodes to avoid processing unnecessary content.
- MetadataFilters defaults to AND condition when none is specified, improving query consistency.
Fixed
- HotpotQA evaluator file handles are now properly closed.
- CodeSplitter preserves oversized leaf nodes instead of discarding them.
- Multiblock chat history writes are preserved during ingestion.
- IngestionPipeline upserts now keep all nodes per document as intended.
June 2026
1 release: 0.14.23
June brought routine maintenance updates to LlamaIndex with regular dependency refreshes across the project.
May 2026
1 release: 0.14.22
May brought routine maintenance updates with dependency improvements across the LlamaIndex project.
Improved
- Dependencies updated across 55 directories with pip group improvements
April 2026
2 releases: 0.14.20 → 0.14.21
April brought security updates and targeted bug fixes to core indexing and persistence components. The month focused on stability improvements across the LlamaIndex ecosystem.
- Message Block Buffer resolution behaviour changed; review any code relying on previous buffer handling
Highlights
- Fixed security vulnerability in nltk dependency
- Resolved KeyError crashes in DocumentSummaryIndex when deleting invalid nodes
- Improved structured output error handling for ValueError and TypeError cases
- Fixed UTF-8 encoding issues in file persistence operations
- Addressed breaking change in Message Block Buffer resolution
Improved
- Enhanced error resilience in structured output processing
- Improved file encoding consistency across persistence layer
Fixed
- Security vulnerability in nltk
- KeyError in DocumentSummaryIndex.delete_nodes with invalid node IDs
- Unhandled ValueError and TypeError from structured output operations
- UTF-8 encoding issues in persistence layer file operations
- Breaking change in Message Block Buffer resolution
March 2026
3 releases: 0.14.16 → 0.14.19
March brought rate limiting capabilities, multimodal reranking support, and numerous stability fixes across retrieval and indexing operations. Python 3.9 support was deprecated.
- Python 3.9 support deprecated; minimum version is now Python 3.10
Highlights
- Token-bucket rate limiter added for LLM and embedding API calls to manage request volume.
- Multimodal LLMReranker now available for ranking results using image and text inputs.
- Fixed async retry backoff to prevent event loop blocking in concurrent operations.
- Improved schema handling in SQL databases and knowledge graph models.
- Python 3.9 officially deprecated; users should upgrade to Python 3.10 or later.
New
- Token-bucket rate limiter for LLM and embedding API calls
- Multimodal LLMReranker for ranking with image and text
- Optional embed model parameter for SemanticDoubleMergingSplitterNodeParser
Improved
- Async retry backoff now avoids blocking the event loop
- Tool choice parameter correctly passed through FunctionCallingProgram
- Sync and async retrieval dedup keys now aligned for consistency
- CTE names preserved during schema prefixing in SQLDatabase queries
- Knowledge graph schema model generation corrected for additionalProperties
Fixed
- Document ID preservation in legacy JSON to document conversion
- DeleteFromDocstore parameter now properly passed in BaseIndex.deleteRefDoc
- StructuredPredict now raises ValueError instead of returning error string
- Removed incorrect per-node delete calls in index helpers
- Chonkie initialization documentation corrected
February 2026
2 releases: 0.14.14 → 0.14.15
February brought security improvements, better cost governance, and expanded multimodal support to LlamaIndex. The month focused on stability fixes and new integration capabilities for agent-based workflows.
Highlights
- Added TokenBudgetHandler for tracking and controlling token usage costs
- Introduced AgentMesh integration as a trust layer for LlamaIndex agents
- Expanded multimodal support with new template variables and prompt helpers
- Improved security defaults and fixed potential crashes in core components
- Added support for recursive LLM type handling
New
- TokenBudgetHandler for cost governance and token budget tracking
- AgentMesh integration providing a trust layer for agents
- Multimodal template variable formatting
- Multimodal prompt templates
- Multimodal chat prompt helper
- Recursive LLM type support
- Langchain 1.x support
Improved
- Enhanced security defaults in core components
- Better handling of empty strings versus None values in MediaResource.hash
- Fallback to bundled NLTK cache when environment variable is missing
- Improved edge case handling in text truncation function
- Extended multimodal type support for basic operations
Fixed
- Potential crashes in core components
- Pydantic ValidationError handling in VectorStoreQueryOutputParser
- DeprecationWarning for asyncio.iscoroutinefunction
- Redundant metadataseperator field removed from TextNode
- Mock prompt type updates in test utilities
January 2026
1 release: 0.14.13
January brought workflow enhancements and distributed processing capabilities to LlamaIndex. The release focused on agent flexibility, data handling improvements, and memory system updates.
- ChatMemoryBuffer replaced by Memory; update references accordingly
Highlights
- Early stopping control added to agent workflows for better execution management.
- Token-based code splitting support enables more precise document segmentation.
- RayIngestionPipeline integration allows distributed data ingestion at scale.
- Multi-modal Condensed Conversation & Context support expands agent capabilities.
- Configurable empty response messages in synthesizers improve output handling.
New
- Early stopping method parameter for agent workflows
- Token-based code splitting in CodeSplitter
- RayIngestionPipeline integration for distributed ingestion
- Multi-modal version of Condensed Conversation & Context
- Configurable empty response message option in synthesizers
Improved
- ChatMemoryBuffer replaced with Memory system
- ReActChatFormatter classmethod implementation corrected for better inheritance
Fixed
- MeanAgg now raises value error on empty list input instead of returning float
- ReActChatFormatter now correctly uses cls parameter in classmethod
December 2025
3 releases: 0.14.9 → 0.14.12
December brought improvements to multi-modal indexing, embeddings support, and tool integration capabilities. The month focused on stability fixes and expanding the platform's functionality for different use cases.
Highlights
- MultiModalVectorStoreIndex now returns a multi-modal ContextChatEngine for better multi-modal search experiences.
- Added support for Amazon Bedrock Application Inference Profiles in the embeddings integration.
- Introduced Airweave tool integration with advanced search features.
- Added async tool spec support for more flexible tool definitions.
- New Element node parser for improved document parsing.
New
- Mock function calling LLM for testing without real model calls.
- Airweave tool integration with advanced search capabilities.
- Async tool spec support for non-blocking tool specifications.
- Element node parser for document parsing.
- Amazon Bedrock Application Inference Profiles support.
Improved
- MultiModalVectorStoreIndex now properly returns multi-modal ContextChatEngine.
- Node content ingestion to vector store is now more readable and accessible.
- MockFunctionCallingLLM functionality improved.
- Bedrock embeddings version bumped with better provider extraction from model names.
- Development logging for llama improved.
- OpenAI integration now sanitizes generic Pydantic model schema names.
Fixed
- Fixed memory handling to ensure first message in queue is always a user message after flush.
- Fixed context copying in async utilities.
- Fixed crash in log vector store query result when result.ids is None.
- Fixed provider extraction from model names in Bedrock embeddings.
- Pinned azure-ai-projects version to prevent breaking changes.
November 2025
1 release: 0.14.8
November brought bug fixes and improvements to agent handling and tool outputs. The month focused on stability improvements for ReAct agents and better preservation of error information.
Highlights
- Fixed ReActOutputParser getting stuck when answer text contains the word Action
- Improved ReActAgent to properly handle multi-block chat messages
- Added buffer support to image, audio, video and document blocks
- Better exception preservation in ToolOutput for error debugging
Improved
- ReActAgent now handles multi-block ChatMessage correctly
- Image, audio, video and document blocks now have buffer support
- Exception information is now preserved in ToolOutput
Fixed
- ReActOutputParser no longer gets stuck when answer text contains 'Action:'
- Pydantic warning resolved
October 2025
4 releases: 0.14.4 → 0.14.7
October brought steady improvements to model integrations and document processing, with new support for Anthropic's latest models and better handling of edge cases in text chunking and streaming.
Highlights
- Added Anthropic Claude 3.5 Sonnet 4.5 model support
- Introduced progress bar for multiprocess document loading
- Fixed duplicate node positions when documents contain identical text
- Added Isaacus embeddings integration
- Improved streaming and tool calling with Anthropic models
New
- Anthropic Claude 3.5 Sonnet 4.5 model support
- Isaacus embeddings integration
- SerpEx tool integration
- Progress bar for multiprocess document loading
- Prompt caching model validation utilities for Anthropic
- Allow parallel tool calls for non-streaming mode
Improved
- Anthropic prompt caching with model validation
- Cohere v4 response format support in BedrockEmbedding
- OCI GenAI Cohere model definitions updated
- SemanticSplitterNodeParser documentation
- VoyageAI integration updated
- Tool call block integration with Anthropic
Fixed
- Duplicate node positions when identical text appears multiple times in documents
- Rare cases when sentence splits exceed chunk size
- Double token stream in Anthropic LLM
- Streaming thinking and tool calling with Anthropic
- Invalid use of field-specific metadata
- Authorization header setup in text embeddings inference
Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. LlamaIndex: pricing, features and alternatives · All changelogs