pypi / llama-index
llama-index
What changed in llama-index 0.14, from its official release notes, upgrade guides and advisories. Sources checked 2026-10-19.
7 breaking1 security15 new APIs and options7 behavior and defaults54 fixes and other
See which of these apply to your code, ranked, in your agent:
npx @unstale/cli initllama-index 0.1484 changes
Breaking changes 7
Frontend tool results are no longer persisted server-side.
The deprecated ipex-llm and optimum-intel IPEX integrations were removed.
The Python version floor was raised to keep vulnerable dependency forks alive.
Raises ValueError instead of fabricating a tool_call_id.
Migrates the MCP tools integration to MCP 2.x.
The Exa tool was updated to current API conventions, so developers may need to align their usage with those conventions.
The llama-index-memory-mem0 integration upgrades to mem0 2.x.
Security 1
Stops hardcoding verify=False for S3 TLS.
New APIs and options 15
StructuredLLMRerank now supports native async operations.
Supports user input containing images, audio, video, and documents.
Adds Claude Opus 5 to the Anthropic and Bedrock Converse model allowlists.
Adds Claude Sonnet 5 support to the Anthropic integration, including function calling.
Adds GPT-5.6 models to the supported OpenAI models.
Adds async support.
Accepts any AsyncDBChatStore.
Adds raise_on_error to LLM path extractors.
Supports the thinking type 'disabled' in Bedrock Converse.
Expands support for the V2 API.
Adds Claude Fable 5 support for Anthropic and Bedrock Converse.
Adds Claude Opus 4.8 support for Anthropic and Bedrock Converse.
Adds multimodal query engines.
Adds multimodal synthesis capabilities to core.
Supports the vLLM reasoning field in OpenAI messages.
Behavior and default changes 7
The Perplexity integration adds an attribution header.
The Weaviate client is tagged with a LlamaIndex integration telemetry header.
Uses Gemini 3.7 Flash by default in the library.
Tool output parsing now handles DocumentBlock and VideoBlock.
The embeddings and LLM integrations allow google-genai 2.0+ SDKs.
The Gemini Live voice agent integration now allows google-genai SDK versions 2.0 and newer.
Serializes rich content blocks in tool results.
Fixes and other changes 54
Compact and refine streaming was restored.
The node_ids query filter now uses Filter.by_id().
Failed function tools are no longer retried.
Images are now kept on Gemini function responses.
Gemini tool request IDs were fixed.
Metadata replacement now falls back when the target value is None.
Images and files are now kept on OpenAI Responses API tool results.
Persists frontend tool messages.
Isolates copies of the initial state.
Honors an agent's structured_output_fn and output_cls.
Makes aioboto3 optional for the Bedrock embeddings integration.
Makes aioboto3 optional for the Bedrock Converse integration.
Preserves falsy metadata values when indexing nodes.
Validates that the embedding payload is non-empty before making an AWS call.
Preserves multiblock chat history writes.
Gives each citation node its own ID and offsets.
Preserves oversized leaf nodes instead of dropping them.
Sets chunk_size to 32000 and fixes the input format.
Reports the model's effective context window in the llama-cpp integration.
Corrects schema, vector ranking, filters, and batching.
Stops sending default parameters.
Returns None instead of a list when an error is swallowed.
Closes evaluator file handles.
Skips empty nodes during parsing.
Keeps all nodes per document when performing upserts.
Repairs an image ingestion crash and an embedding_model typo.
Fixes the default endpoint and metadata.
Defaults the condition to AND when it is None.
Respects mmr_threshold=0 rather than treating it as unset.
Corrects the Opus 4.6 context window to 1M.
Supports Pinecone client v8 and v9.
Restores fallback to supplied LLM metadata.
Preserves falsy shard_identifier values, such as 0.
Restores compatibility with qdrant-client 1.19.0.
Sends query_history in the DashScope retriever payload.
Uses word tokenization for stopword removal.
Persists non-ASCII characters without escaping them.
Populates response text when writing a streaming chat response to memory.
Avoids overwriting thought signatures.
Does not mark *args or **kwargs as required tool parameters.
Returns real collection properties as node metadata.
Preserves the IndexNode object during model dumping.
Uses the running event loop in the ingestion pipeline.
Deep-copies workflow initial state to prevent mutation leaks across runs.
Resolved an issue involving string_expr and filter precedence in the Milvus vector store.
Fixes matching of missing metadata for NE and NIN filters.
Guards against ZeroDivisionError when input sequences are empty.
Fixes PromptMode import compatibility in the RankLLM reranker integration.
Fixes reuse of kwargs when refreshing reference documents.
Fixes RecursionError when units are larger than chunk_size.
Streaming tool_kwargs are parsed from a string into a dictionary.
Preserves source nodes returned by TreeSelectLeafRetriever.
Preserves URL-backed video and document memory blocks.
The Weaviate vector store now supports newer batch context managers.