run-llama/llama_index
57.7
Adequate · 14 November 2026
256.6k
lines of production code
Python
with JavaScript
6
measurements over time
What this system is
Features
Add chat UI events and models to core package
The core package now includes the foundational data structures for the chat UI, specifically introducing event classes (UIEvent, SourceNodesEvent, ArtifactEvent) and model definitions. This provides the necessary types for handling chat interactions and artifacts within the application's event-driven architecture.
_llama-index-core/llama\_index/core/chat\ui · high confidence
Add multi-modal faithfulness and relevancy evaluators
New \MultiModalFaithfulnessEvaluator\ and \MultiModalRelevancyEvaluator\ classes are now available in the \llama\_index.core.evaluation.multi\_modal\ module. These evaluators allow users to assess whether LLM responses are faithful to or relevant against combined textual and visual contexts (images provided via paths or URLs). The evaluators default to using \OpenAIResponses\ with the \gpt-4.1\ model for judgment, but support custom multi-modal LLMs and prompt templates.
_llama-index-core/llama\_index/core/evaluation/multi\modal · high confidence
Added mock sparse embedding implementation for testing
A new MockSparseEmbedding class has been introduced in the sparse\_embeddings module to support testing workflows. This implementation provides a configurable mock that returns a default embedding or looks up specific text-to-embedding mappings, enabling developers to test sparse embedding logic without relying on external models or services.
_llama-index-core/llama\_index/core/sparse\embeddings · high confidence
Base embedding class now supports caching and rate limiting
The \BaseEmbedding\ class in \llama\_index/core/base/embeddings/base.py\ has been updated to include optional \embeddings\_cache\ and \rate\_limiter\ fields. Users can now configure a \BaseKVStore\ instance to cache embedding results, avoiding redundant API calls for repeated queries, and attach a rate limiter to throttle API usage. The class also emits instrumentation events (\EmbeddingStartEvent\, \EmbeddingEndEvent\) for observability and validates that the cache is of the correct type.
_llama-index-core/llama\index/core/base/embeddings · high confidence
Introduce Azure Table Storage utility package
Added a new \llama-index-utils-azure\ package containing utilities for Azure Table Storage and Cosmos DB. This includes a \ServiceMode\ enum to distinguish between the two services, and helper functions to sanitize table names, validate property counts and sizes against service limits, and serialize/deserialize complex data types (splitting large values into parts to respect Azure's per-property size constraints).
llama-index-utils · high confidence
Introduce Document Summary Index with LLM and Embedding Retrievers
Added a new Document Summary Index that generates summaries for documents and maps them to underlying nodes, enabling retrieval via two modes: an LLM-based retriever that selects relevant summaries using a choice-select prompt, and an embedding-based retriever that uses vector similarity. The index supports configurable summary queries, progress tracking, and integrates with the existing document store and vector store infrastructure.
_llama-index-core/llama\_index/core/indices/document\summary · high confidence
Introduce EmptyIndex and KeywordTableIndex for specialized retrieval patterns
This change adds two new index types to the core library. The EmptyIndex allows users to create an index with no documents, enabling pure LLM calls and composition with other indices while supporting only generation-based query modes. The KeywordTableIndex provides a keyword-based retrieval structure that extracts keywords from documents and queries to map and retrieve relevant text chunks, supporting three retrieval modes: default (GPT-based extraction), simple (regex-based), and RAKE (using the RAKE algorithm). These indices are available via the standard \llama\_index.core.indices\ module.
_llama-index-core/llama\_index/core/indices/empty, llama-index-core/llama\_index/core/indices/keyword\table · high confidence
Introduce MultiModalVectorStoreIndex for combined text and image retrieval
Users can now build a single index that manages both text and image nodes using separate vector stores. The new MultiModalVectorStoreIndex automatically handles embedding and storing images (defaulting to a CLIP-based model) alongside standard text, while the MultiModalVectorIndexRetriever allows querying across both modalities simultaneously. This enables applications to retrieve relevant text and image content in a unified search operation.
_llama-index-core/llama\_index/core/indices/multi\modal · high confidence
Introduce Property Graph Sub-retrievers
Adds a new \sub\_retrievers\ package for the Property Graph Index, providing a \BasePGRetriever\ foundation and specialized implementations: \VectorContextRetriever\ for hybrid vector-graph retrieval, \TextToCypherRetriever\ for LLM-generated Cypher queries, \CypherTemplateRetriever\ for structured LLM-to-Cypher templating, \LLMSynonymRetriever\ for query expansion, and \CustomPGRetriever\ for user-defined logic.
_llama-index-core/llama\_index/core/indices/property\_graph/sub\retrievers · high confidence
Introduce PropertyGraphIndex and PGRetriever for graph-based retrieval
This change introduces the \PropertyGraphIndex\ and \PGRetriever\ components within the \llama\_index.core.indices.property\_graph\ module, enabling users to build and query indices based on property graphs. The \PropertyGraphIndex\ manages the storage of labeled nodes and relations, supports embedding of knowledge graph nodes, and utilizes configurable extractors (such as \SimpleLLMPathExtractor\ and \ImplicitPathExtractor\) to derive graph structures from input data. The \PGRetriever\ orchestrates multiple sub-retrievers (e.g., vector, cypher, LLM synonym) to fetch relevant nodes, handling deduplication and async execution. Utility functions like \default\_parse\_triplets\_fn\ are also provided to assist in parsing LLM-generated triplets.
_llama-index-core/llama\_index/core/indices/property\graph · high confidence
Introduce SimplePropertyGraphStore and graph store utilities
This change adds a new in-memory labelled property graph store implementation (\SimplePropertyGraphStore\) along with supporting types, prompts, and utilities. Users can now persist and load graph data with explicit UTF-8 encoding, query nodes and triplets with filtering by properties or IDs, and traverse relationships with depth-aware mapping. The store also includes utilities to sanitize large list values and clean string inputs to reduce context noise for LLMs.
_llama-index-core/llama\_index/core/graph\stores · high confidence
Introduce TreeIndex with multiple retrieval modes
The \llama\_index.core.indices.tree\ module now provides a \TreeIndex\ that organizes text documents into a hierarchical structure built bottom-up, where each parent node summarizes its children. Users can query this index using four distinct modes exposed via \as\_retriever\: \select\_leaf\ (recursive top-down traversal using LLM prompts to choose nodes), \select\_leaf\_embedding\ (traversal using embedding similarity), \all\_leaf\ (retrieving all leaf nodes), and \root\ (synthesizing answers directly from root nodes). The implementation includes the core index logic, an inserter for building and updating the tree structure, and specific retriever classes for each mode.
_llama-index-core/llama\index/core/indices/tree · high confidence
Introduce VectorIndexAutoRetriever for LLM-driven vector store queries
Users can now use the new VectorIndexAutoRetriever to automatically generate vector store query parameters (such as filters and top\_k) by leveraging an LLM. This component takes a VectorStoreIndex and VectorStoreInfo, uses a prompt template with examples to guide the LLM, and parses the output into a VectorStoreQuerySpec. It includes robust error handling, falling back to default query parameters if the LLM output cannot be parsed or validated.
_llama-index-core/llama\_index/core/indices/vector\_store/retrievers/auto\retriever · high confidence
Introduce base class and types for managed service indices
Added a new \llama\_index.core.indices.managed\ package containing \BaseManagedIndex\, an abstract base class for indices backed by managed services, along with \ManagedIndexQueryMode\ enum defining supported query modes (default, MMR). This establishes the foundational interface for future managed index implementations.
_llama-index-core/llama\index/core/indices/managed · high confidence
Introduce core embedding module with multi-modal support and model resolution
The embeddings package has been restructured into a dedicated core module, introducing a \MultiModalEmbedding\ base class that enables synchronous and asynchronous embedding of images alongside text, along with batch processing and progress tracking. A new \Pooling\ utility provides CLS and mean pooling operations for vector arrays, while \resolve\_embed\_model\ centralizes the logic for instantiating embedding models from strings (such as 'default' for OpenAI, 'clip' for CLIP, or 'local' for HuggingFace) or LangChain adapters, falling back to \MockEmbedding\ when no provider is available or during testing. The \load\_embed\_model\ function allows serialization-based loading of recognized embedding classes, and \MockEmbedding\/\MockMultiModalEmbedding\ are provided for testing purposes.
_llama-index-core/llama\index/core/embeddings · high confidence
Introduce core ingestion pipeline, cache, and API utilities
The ingestion module now exposes a structured API for data processing. Users can utilize the new IngestionPipeline and associated transformation runners (run\_transformations, arun\_transformations) to process documents, and leverage IngestionCache to store and retrieve processed nodes via a configurable key-value store. Additionally, deprecated helper functions (get\_client, get\_aclient) are provided to instantiate LlamaCloud SDK clients, allowing users to pass custom httpx clients for platform integration.
_llama-index-core/llama\index/core/ingestion · high confidence
Introduce core selector implementations and factory utility
The \llama\_index.core.selectors\ module now provides a unified set of selector classes for choosing among multiple options. This includes \LLMSingleSelector\ and \LLMMultiSelector\ for LLM-based selection, \EmbeddingSingleSelector\ for embedding-similarity-based selection, and \PydanticSingleSelector\ and \PydanticMultiSelector\ for structured output selection. A new \get\_selector\_from\_llm\ utility is also available to automatically instantiate the best available selector (preferring Pydantic-based ones) for a given LLM.
_llama-index-core/llama\index/core/selectors · high confidence
Introduce llama-dev CLI for monorepo development and automation
Adds the \llama-dev\ package, providing an official CLI for development, testing, and automation within the LlamaIndex monorepo. The tool includes commands for package management (\pkg\), testing (\test\), and release checks (\release\), along with a \--debug\ flag for verbose output and a \--repo-root\ option to specify the repository path. It also introduces utility functions to detect changed packages, find integrations and utils, and manage version bumps based on semver rules.
_llama-dev/llama\dev · high confidence
Introduce llama-dev package management CLI
The \llama-dev\ tool now includes a new \pkg\ command group for managing packages within the monorepo. Users can retrieve package details (name, version, path) via the \info\ subcommand, execute arbitrary commands across specific or all packages using \exec\, and update package versions with the \bump\ subcommand, which supports patch/minor/major bumps and dry-run modes.
_llama-dev/llama\dev/pkg · high confidence
Introduce new chat store storage architecture
The chat store module has been restructured with a new base interface (BaseChatStore) that exposes both synchronous and asynchronous methods for managing chat history. A new in-memory implementation (SimpleChatStore) is provided, featuring improved persistence that writes non-ASCII characters as-is rather than escaping them. Additionally, a new SQLAlchemy-based store (SQLAlchemyChatStore) is introduced, supporting async database connections, message status tracking for FIFO queue management, and optional database schema configuration for PostgreSQL and other supported databases.
_llama-index-core/llama\_index/core/storage/chat\store · high confidence
Introduce new node postprocessors and async support
The postprocessor module now includes several new capabilities: LLM-based reranking via LLMRerank and StructuredLLMRerank, PII masking through PIINodePostprocessor and NERPIINodePostprocessor, and time-based filtering with FixedRecencyPostprocessor, EmbeddingRecencyPostprocessor, and TimeWeightedPostprocessor. Additionally, the base postprocessor interface now supports asynchronous processing via \_apostprocess\_nodes, allowing for non-blocking node transformations.
_llama-index-core/llama\index/core/postprocessor · high confidence
Introduce relational node parsers for hierarchical and element-based document splitting
This change introduces a new \relational\ node parser module that provides parsers capable of preserving structural relationships between document elements. It adds \HierarchicalNodeParser\ for recursively splitting documents into parent-child node hierarchies using configurable chunk sizes and sub-parsers, and several \ElementNodeParser\ variants (\MarkdownElementNodeParser\, \UnstructuredElementNodeParser\, \LlamaParseJsonNodeParser\) that extract specific content types like tables, code blocks, and headings into separate nodes while maintaining source relationships. These parsers enable more granular retrieval by keeping related content (e.g., table summaries and their source text) linked.
_llama-index-core/llama\_index/core/node\parser/relational · high confidence
Introduce structured metadata filtering and node serialization utilities
The vector stores module now exposes a standardized metadata filtering API and robust node serialization logic. Users can apply structured filters (e.g., exact match, comparison, text match) to query results via the new \MetadataFilters\ and \FilterOperator\ types, with support for AND/OR/NOT conditions. Additionally, node data is now consistently serialized into a \\_node\_content\ metadata field, enabling reliable round-trip conversion between nodes and metadata dictionaries, which improves compatibility across different vector store backends.
_llama-index-core/llama\_index/core/vector\stores · high confidence
Introduce structured output parsers for Pydantic and selection tasks
This change introduces the \llama\_index.core.output\_parsers\ module, providing \PydanticOutputParser\ and \SelectionOutputParser\ to help users structure LLM responses into typed Pydantic models or specific choices. The \PydanticOutputParser\ generates JSON schemas from Pydantic classes and handles parsing back into model instances, while utility functions in \utils.py\ assist in extracting JSON and code blocks from LLM outputs, including fallback parsing via PyYAML for non-standard JSON.
_llama-index-core/llama\_index/core/output\parsers · high confidence
Introduction of BaseToolSpec for unified sync/async tool definitions
Users can now define tool specifications that support both synchronous and asynchronous functions through the new \BaseToolSpec\ class in \llama\_index.core.tools.tool\_spec\. The \to\_tool\_list\ method automatically detects whether a specified function is a coroutine and passes the appropriate sync or async handler to \FunctionTool\, allowing a single tool spec to expose both variants (via string or tuple definitions). This simplifies the creation of tools that need to leverage async execution without requiring separate implementations for sync and async paths.
_llama-index-core/llama\_index/core/tools/tool\spec · high confidence
Introduction of GPTTreeIndexBuilder for hierarchical tree index construction
A new GPTTreeIndexBuilder class has been added to the common tree indices module, providing a helper to construct tree-structured indexes from document nodes. This builder recursively consolidates text chunks into parent nodes by generating summaries via an LLM, supporting both synchronous and asynchronous execution modes, configurable tree branching via num\_children, and progress tracking. It integrates with the existing document store and prompt helper to manage node creation and text truncation during the bottom-up index building process.
_llama-index-core/llama\_index/core/indices/common\tree · high confidence
Introduction of LLM-based sub-question generation
The \question\_gen\ module now provides \LLMQuestionGenerator\ and \SubQuestionOutputParser\ to automatically break down complex user queries into structured sub-questions. The generator uses an LLM to analyze the original query and available tools, producing a list of \SubQuestion\ objects that can be used to guide downstream retrieval or tool execution. The output parser handles JSON extraction from LLM responses, including robustness for formats that wrap results in an 'items' key.
_llama-index-core/llama\_index/core/question\gen · high confidence
Introduction of RichPromptTemplate and multimodal chat prompt support
The prompts module now includes a new \RichPromptTemplate\ class that supports multimodal content (images, audio, video, documents) via the \banks\ library, enabling users to construct chat prompts with mixed media blocks. This change also introduces dedicated multimodal chat prompts (e.g., \CHAT\_CONTENT\_QA\_PROMPT\) that can render these media types in context, and adds a \SelectorPromptTemplate\ mechanism to automatically switch between standard and Cohere-specific templates when using Cohere models.
_llama-index-core/llama\index/core/prompts · high confidence
Introduction of StorageContext for unified storage management
The \llama\_index.core.storage\ module now exposes a \StorageContext\ class that serves as a central container for document, index, vector, graph, and property graph stores. This context allows users to initialize, persist, and manage these underlying storage components together, supporting both in-memory defaults and persistence to disk via configurable directories and file names.
_llama-index-core/llama\index/core/storage · high confidence
Introduction of new index struct types and registry
The core data structures module has been reorganized to introduce a comprehensive registry for index types. This change adds support for several new index structures, including ClickHouse, PostgresML, Property Graph (LPG), Document Summary, and Multimodal Vector Store indices, alongside existing types like Tree, List, and Keyword Table. Users can now utilize these new index types through the updated \IndexStructType\ enum and the central registry that maps them to their corresponding implementation classes.
_llama-index-core/llama\_index/core/data\structs · high confidence
Introduction of the new chat\_engine module with multimodal support
The \llama\_index/core/chat\_engine\ module has been introduced, providing a new set of chat engine implementations including \SimpleChatEngine\, \ContextChatEngine\, \CondenseQuestionChatEngine\, and \CondensePlusContextChatEngine\. This release also adds \MultiModalContextChatEngine\ and \MultiModalCondensePlusContextChatEngine\, enabling users to build conversational agents that can retrieve and process both text and image context from multimodal vector stores. The module includes a new \StreamingAgentChatResponse\ type to handle streaming responses and integrates with the new instrumentation events for better observability.
_llama-index-core/llama\_index/core/chat\engine · high confidence
Introduction of workflow-based agents and structured output utilities
The agent module has been restructured to introduce a new workflow-based architecture, exposing new agent types including FunctionAgent, CodeActAgent, and ReActAgent via the BaseWorkflowAgent interface. This change also adds event types (AgentInput, AgentStream, AgentOutput) for better state management and includes utility functions to support structured generation, specifically converting conversation messages to XML format and generating structured responses from LLMs.
_llama-index-core/llama\index/core/agent · high confidence
New BEIR and HotpotQA benchmark evaluators
Added \BeirEvaluator\ and \HotpotQAEvaluator\ classes to the \llama\_index.core.evaluation.benchmarks\ module, allowing users to evaluate retrieval and query performance on the BEIR and HotpotQA datasets respectively. The \BeirEvaluator\ supports downloading and evaluating against various BEIR datasets using standard retrieval metrics (NDCG, MAP, Recall, Precision), while the \HotpotQAEvaluator\ provides a way to assess answer quality on the HotpotQA distractor dataset using exact match and F1 scores.
_llama-index-core/llama\index/core/evaluation/benchmarks · high confidence
New CLI commands for automated release management
The \llama\_dev.release\ module introduces a new Click-based CLI with three commands to streamline the release process. The \release check\ command validates release prerequisites, such as ensuring the current branch is \main\ and verifying that the local \llama-index-core\ version matches or exceeds the latest PyPI release (with a \--before-core\ flag for pre-release checks). The \release changelog\ command automatically generates a markdown changelog by fetching PR details from GitHub since the last tag and grouping them by package. The \release prepare\ command bumps version numbers for the umbrella package and its core dependency, supporting configurable bump types (e.g., patch, minor) and dry-run modes.
_llama-dev/llama\dev/release · high confidence
New DocumentContextExtractor and extractor interface updates
The extractors module now includes a new DocumentContextExtractor that implements the Contextual Retrieval approach to generate succinct context for document chunks, improving RAG accuracy. The BaseExtractor interface has been updated to support configurable retry logic with exponential backoff and error handling policies, allowing users to control behavior when extraction fails. Additionally, the module exports these new components alongside existing extractors like KeywordExtractor and TitleExtractor.
_llama-index-core/llama\index/core/extractors · high confidence
New FLARE query engine with lookahead retrieval
Adds the FLARE (Active Retrieval Augmented Generation) query engine to llama-index-core, enabling iterative, lookahead-style retrieval where the model can insert \[Search(query)\] tags into a response draft and execute those queries in parallel before synthesizing the final answer. This new module includes the FLAREInstructQueryEngine, an answer inserter component (with both LLM-based and direct insertion strategies), and output parsers for detecting completion and extracting search tasks, allowing users to build more dynamic, self-correcting retrieval-augmented generation workflows.
_llama-index-core/llama\_index/core/query\engine/flare · high confidence
New LangChain helper modules for streaming and compatibility
The \llama\_index/core/langchain\_helpers\ package has been introduced, providing a \StreamingGeneratorCallbackHandler\ that enables token-by-token streaming responses with a configurable timeout (defaulting to 120 seconds) and a \text\_splitter\ module that re-exports core splitters for backward compatibility with existing LangChain-based integrations.
_llama-index-core/llama\_index/core/langchain\helpers · high confidence
New LoadAndSearch tool spec for splitting large data loads and queries
A new LoadAndSearchToolSpec is introduced to help agents handle tools that return large amounts of data. This spec wraps an existing tool to split its operation into two steps: a 'load' step that retrieves data and stores it in an index (defaulting to VectorStoreIndex), and a 'read' step that allows the agent to search that index using natural language queries. This prevents context window limits by separating data ingestion from retrieval.
_llama-index-core/llama\_index/core/tools/tool\_spec/load\_and\search · high confidence
New Object Index and Node Mapping API with Secure Pickle Deserialization
This change introduces the \ObjectIndex\ and \ObjectRetriever\ classes, allowing users to index and retrieve arbitrary Python objects (such as tools or SQL tables) alongside standard text nodes. It adds specific node mapping implementations (\SimpleObjectNodeMapping\, \SimpleToolNodeMapping\, \SQLTableNodeMapping\, \FnNodeMapping\) to handle the serialization of these objects into nodes. Crucially, the \SimpleObjectNodeMapping\ now uses a \RestrictedUnpickler\ when loading persisted mappings, restricting deserialization to a safe allowlist of built-in types and the mapping class itself to prevent arbitrary code execution (CWE-502) from untrusted pickle files.
_llama-index-core/llama\index/core/objects · high confidence
New base LLM interface and utility functions
The \llama\_index.core.base.llms\ package now provides the foundational \BaseLLM\ class and helper utilities. The \BaseLLM\ interface defines the core contract for LLM implementations, including abstract methods for chat, completion, and streaming operations, along with a \rate\_limiter\ field for throttling API calls and a \to\_payload\ method for safe observability logging. Additionally, \generic\_utils.py\ introduces functions to convert between chat messages and prompts, transform completion responses into chat responses (and vice versa), and decorate functions to bridge the gap between chat and completion interfaces, enabling consistent interaction patterns across different LLM types.
_llama-index-core/llama\index/core/base/llms · high confidence
New base classes for auto, multi-modal, and query engine retrieval
The \llama\_index/core/base\ module now introduces \BaseAutoRetriever\, \MultiModalRetriever\, and \BaseQueryEngine\ as foundational abstractions. \BaseAutoRetriever\ enables dynamic retrieval by generating and applying retrieval specifications synchronously and asynchronously. \MultiModalRetriever\ extends the base retriever to support text, image, and cross-modal (text-to-image, image-to-image) retrieval operations. \BaseQueryEngine\ provides a structured query interface with integrated instrumentation spans and callback management, while the existing \BaseRetriever\ continues to handle node retrieval and recursive object resolution. These classes standardize how users implement and extend retrieval and query capabilities across different data modalities.
_llama-index-core/llama\index/core/base · high confidence
New base classes for voice agent integrations
The \llama\_index/core/voice\_agents\ module now provides abstract base classes (\BaseVoiceAgent\, \BaseVoiceAgentInterface\, \BaseVoiceAgentWebsocket\, and \BaseVoiceAgentEvent\) that define the standard interface for building voice-enabled agents. These classes establish the core contract for managing audio input/output, handling websocket connections, and tracking conversation state, serving as the foundation for specific provider implementations like the newly added OpenAI Realtime integration.
_llama-index-core/llama\_index/core/voice\agents · high confidence
New bridge module for LangChain and Pydantic compatibility
A new \llama\_index.core.bridge\ package has been introduced to centralize compatibility layers for external dependencies. The \langchain.py\ bridge provides unified imports for LangChain components, supporting both v0.x and v1.x versions through conditional imports, and exposes specific LLM providers like ChatOpenAI, ChatFireworks, and ChatAnyscale. Additionally, the bridge includes modules for \pydantic\, \pydantic\_core\, and \pydantic\_settings\ to standardize access to their respective APIs, ensuring consistent behavior across different library versions.
_llama-index-core/llama\index/core/bridge · high confidence
New core abstractions for agents, sparse embeddings, and content blocks
This release introduces foundational classes that reshape how agents, embeddings, and multimodal content are handled. The \BaseWorkflowAgent\ class (in \llama\_index.core.agent.workflow.base\_agent\) provides a unified agent base that integrates with the workflow system, supporting structured outputs, tool retrieval, and configurable early stopping. A new sparse embedding abstraction (\llama\_index.core.base.embeddings.base\_sparse\) defines \BaseSparseEmbedding\ and the \SparseEmbedding\ type, enabling sparse vector retrieval alongside dense embeddings. The \llama\_index.core.base.llms.types\ module adds a \BaseContentBlock\ hierarchy and related message types, laying the groundwork for multimodal content handling. Additionally, new callback schemas (\llama\_index.core.callbacks.schema\) and a token-counting handler (\llama\_index.core.callbacks.token\_counting\) provide structured event tracking and token budgeting, while a \BaseSelector\ (\llama\_index.core.base.base\_selector\) offers an LLM-driven tool/choice selection interface.
llama-index · high confidence
New core utility modules for AWS, Gemini, SQL, and token counting
This change introduces several new utility modules in the core utilities package. It adds \aws\_utils.py\ to provide a standardized way to create AWS service clients with configurable retries and timeouts. It adds \gemini\_utils.py\ to handle role mapping and message merging for Gemini and Vertex AI models, ensuring compatibility with Gemini's chat mode constraints. It introduces \sql\_wrapper.py\, a wrapper around SQLAlchemy that provides enhanced table info retrieval, including support for table comments and optional view support. Finally, it adds \token\_counting.py\, which implements a \TokenCounter\ class to estimate token usage for chat messages and tools, supporting both synchronous and asynchronous estimation.
_llama-index-core/llama\index/core/utilities · high confidence
New documentation and package publishing automation scripts
This change introduces four new scripts to the \scripts/\ directory to automate documentation syncing, example conversion, and package publishing. \sync-docs-to-developer-hub.sh\ orchestrates the sync of framework docs, API references (built via mkdocs), and Jupyter examples to the developer hub, supporting incremental updates via \--since\ flags and skipping specific sections. \convert-examples.py\ handles the conversion of Jupyter notebooks to Markdown with Astro-compatible frontmatter, routing integration examples to a separate directory and detecting changes via git diffs. \integration\_health\_check.py\ provides a scoring mechanism for package health based on download trends and commit activity. Finally, \publish\_packages.sh\ manages the publishing of Python packages with retry logic and progress tracking.
scripts · high confidence
New evaluation module with specialized evaluators and batch runner
The \llama\_index.core.evaluation\ package is introduced, providing a suite of specialized evaluators for assessing RAG system quality. This includes \AnswerRelevancyEvaluator\ and \ContextRelevancyEvaluator\ for measuring relevance, \FaithfulnessEvaluator\ (which now includes an optimized prompt for Llama3:8B), \CorrectnessEvaluator\ for accuracy against reference answers, \RelevancyEvaluator\ for context-response alignment, \PairwiseComparisonEvaluator\ for comparing two responses, and \SemanticSimilarityEvaluator\ for embedding-based comparison. The module also introduces \BatchEvalRunner\ to execute evaluations in parallel with configurable workers and retry logic, along with utility functions for generating datasets and formatting results into DataFrames.
_llama-index-core/llama\index/core/evaluation · high confidence
New file-based node parsers for HTML, JSON, Markdown, and simple files
Users can now parse documents directly from files using four new node parsers: HTMLNodeParser, JSONNodeParser, MarkdownNodeParser, and SimpleFileNodeParser. HTMLNodeParser splits content by configurable HTML tags (defaulting to standard structural tags like p, h1-h6, section) and groups consecutive same-tag text into single nodes. JSONNodeParser performs a depth-first traversal of JSON structures, yielding leaf values with their key paths as metadata. MarkdownNodeParser splits documents on headers, supporting customizable header path separators and correctly handling indented code blocks to prevent header parsing inside code. SimpleFileNodeParser acts as a dispatcher that automatically selects the appropriate parser based on file extension (e.g., .md, .html, .json) extracted from the document's metadata or file path, falling back to a simple text split for unsupported extensions.
_llama-index-core/llama\_index/core/node\parser/file · high confidence
New in-memory key-value store with persistence support
A new \SimpleKVStore\ implementation has been added to the \llama\index.core.storage.kvstore\ module, providing an in-memory key-value store that supports serialization to and from JSON files via \fsspec\. This change introduces the \MutableMappingKVStore\ base class in \types.py\, which defines the interface for stores using Python's \MutableMapping\, and updates the \\\init\\_.py\ to export \SimpleKVStore\ alongside existing stores like \FirestoreKVStore\ and \RedisKVStore\. Users can now utilize this simple, file-backed store for scenarios requiring lightweight, persistent key-value storage without external database dependencies.
_llama-index-core/llama\index/core/storage/kvstore · high confidence
New modular retriever architecture with advanced fusion and routing capabilities
The retrievers module has been restructured to expose a comprehensive set of specialized retrieval strategies. Users can now leverage the QueryFusionRetriever, which supports multiple fusion modes including reciprocal rank, simple reordering, and a new relative score fusion that normalizes scores across different retrievers. The AutoMergingRetriever intelligently consolidates related chunks into parent contexts to reduce redundancy, while the RecursiveRetriever enables traversal of linked knowledge graphs by following IndexNode references. Additionally, the RouterRetriever allows dynamic selection among multiple candidate retrievers based on query context, and the TransformRetriever provides a wrapper to apply query transformations before retrieval.
_llama-index-core/llama\index/core/retrievers · high confidence
New multi-modal LLM utilities for image handling
The \llama\_index.core.multi\_modal\_llms\ package is introduced, providing core infrastructure for multi-modal language models. This includes the \MultiModalLLM\ base class and \MultiModalLLMMetadata\ for model configuration, alongside \generic\_utils\ which adds functions to load images from URLs, encode local image files to base64, and automatically infer MIME types from file paths or base64 content. These utilities enable users to seamlessly integrate image data into multi-modal LLM workflows.
_llama-index-core/llama\_index/core/multi\_modal\llms · high confidence
New node parser module with hierarchical utilities and LlamaParse integration
The \llama\_index.core.node\_parser\ package has been introduced, centralizing node parsing logic. This update adds the \LlamaParseJsonNodeParser\ for parsing JSON outputs from LlamaParse, along with \SemanticDoubleMergingSplitterNodeParser\ and \LanguageConfig\ for advanced semantic chunking. It also exposes hierarchical navigation utilities (\get\_leaf\_nodes\, \get\_root\_nodes\, \get\_child\_nodes\, \get\_deeper\_nodes\) via \HierarchicalNodeParser\ and provides a \load\_parser\ factory in \loading.py\ to instantiate parsers by class name. The \interface.py\ base class now includes improved post-processing for duplicate text handling and metadata merging, while \node\_utils.py\ adds support for building \ImageNode\ instances from \ImageDocument\ sources.
_llama-index-core/llama\_index/core/node\parser · high confidence
New prebuilt memory block components
The \llama\_index.core.memory.memory\_blocks\ package now exposes three prebuilt memory block implementations: \StaticMemoryBlock\ for returning constant text content, \VectorMemoryBlock\ for storing and retrieving conversation history via a vector store with session-based filtering, and \FactExtractionMemoryBlock\ for using an LLM to extract and condense key facts from conversations. These classes are now exported from the \memory\_blocks\ module for direct use.
_llama-index-core/llama\_index/core/memory/memory\blocks · high confidence
New property graph path extraction components and improved LLM schema compatibility
This change introduces a new \transformations\ subpackage for property graph indices, exposing four extraction components: \ImplicitPathExtractor\ (which converts existing node relationships like parent/child into graph edges), \SimpleLLMPathExtractor\, \SchemaLLMPathExtractor\, and \DynamicLLMPathExtractor\ (which use LLMs to extract entity-relation triplets, with the latter two supporting schema validation and dynamic ontology expansion). It also adds a \raise\_on\_error\ parameter to the LLM-based extractors to control exception handling during extraction failures, and introduces an \allow\_additional\_properties\ option in \SchemaLLMPathExtractor\ to clean auto-generated JSON schemas, ensuring compatibility with strict LLM providers like OpenAI and Google Gemini.
_llama-index-core/llama\_index/core/indices/property\graph/transformations · high confidence
New query transform components for HyDE, decomposition, and feedback-driven resynthesis
This change introduces a new \query\_transform\ module within the core indices package, providing several new query transformation strategies that users can apply before querying an index. It adds \HyDEQueryTransform\ to generate hypothetical documents for improved embedding retrieval, \DecomposeQueryTransform\ and \StepDecomposeQueryTransform\ to break down complex queries into sub-questions based on context and previous reasoning steps, and \FeedbackQueryTransformation\ to resynthesize queries using evaluation feedback. The module also includes the necessary base classes (\BaseQueryTransform\, \IdentityQueryTransform\) and default prompt templates to support these transformations.
_llama-index-core/llama\_index/core/indices/query/query\transform · high confidence
New query utility modules and backward-compatibility shims
The \llama\_index/core/indices/query\ package now includes \embedding\_utils.py\, which provides functions for retrieving top-k embeddings using similarity search, learner-based modes (SVM, linear/logistic regression), and Maximal Marginal Relevance (MMR) with configurable thresholds. Additionally, \base.py\ and \schema.py\ have been added as backward-compatibility shims that re-export \BaseQueryEngine\, \QueryBundle\, and \QueryType\ from their canonical locations to ensure existing imports continue to work.
_llama-index-core/llama\index/core/indices/query · high confidence
New response synthesizer modes: ContextOnly, NoText, and Accumulate
The response synthesizer module now exposes three additional response modes—ContextOnly, NoText, and Accumulate—alongside the existing Refine, CompactAndRefine, TreeSummarize, SimpleSummarize, Generation, and CompactAndAccumulate. ContextOnly returns the raw retrieved context without LLM generation, NoText returns an empty string, and Accumulate generates individual responses for each text chunk and concatenates them with a separator. These modes are registered in the factory and can be selected via the ResponseMode enum, giving users finer control over how retrieved information is processed into final answers.
_llama-index-core/llama\_index/core/response\synthesizers · high confidence
New response utility modules for Jupyter and CLI display
The \llama\_index.core.response\ package now includes dedicated utility modules to improve how responses are presented in different environments. \notebook\_utils.py\ adds functions like \display\_response\ and \display\_query\_and\_multimodal\_response\ to render text, source nodes, and images directly in Jupyter notebooks. \pprint\_utils.py\ provides corresponding \pprint\_response\ and \pprint\_source\_node\ functions for clean, wrapped text output in standard CLI terminals. Additionally, \utils.py\ introduces synchronous and asynchronous helpers (\get\_response\_text\, \aget\_response\_text\) to easily consume response generators into a single string.
_llama-index-core/llama\index/core/response · high confidence
New retrieval evaluation metrics and evaluators
Added new classes for evaluating retrieval performance, including \RetrieverEvaluator\ and \MultiModalRetrieverEvaluator\ to orchestrate evaluation with optional node post-processing, and introduced specific metrics such as \HitRate\, \MRR\, and \Precision\ in the retrieval metrics module.
_llama-index-core/llama\index/core/evaluation/retrieval · high confidence
New structured output program implementations and utilities
The \llama\index/core/program\ module now includes new program classes for generating structured outputs: \FunctionCallingProgram\ for LLMs that support function calling, \LLMTextCompletionProgram\ for generic text completion with Pydantic parsing, and \MultiModalLLMCompletionProgram\ for multimodal LLMs that accept image documents. These classes are exported from the module's \\\init\\_.py\ and are selected automatically via \get\_program\_for\_llm\ based on the \PydanticProgramMode\ (DEFAULT, OPENAI, FUNCTION, LLM, LM\_FORMAT\_ENFORCER). Supporting utilities include \streaming\_utils.py\ for incremental processing of streaming structured outputs, \utils.py\ for flexible model creation and JSON repair, and \llm\_prompt\_program.py\ as a base class for LLM prompt programs.
_llama-index-core/llama\index/core/program · high confidence
New text node parsers and splitters for code, semantics, and tokens
The \llama\_index/core/node\parser/text\ module now exposes a comprehensive suite of text splitting and node parsing strategies. Users can split code using \CodeSplitter\ with AST-based chunking and optional token-based size limits, or leverage \LangchainNodeParser\ to wrap Langchain's text splitters (supporting both v0.x and v1.x). For natural language, \SemanticSplitterNodeParser\ and \SemanticDoubleMergingSplitterNodeParser\ enable embedding- or Spacy-based semantic chunking with configurable thresholds, while \SentenceWindowNodeParser\ creates nodes with surrounding context windows. Additionally, \TokenTextSplitter\ provides token-aware splitting with metadata-reservation logic, and \SentenceSplitter\ handles basic sentence-level splitting. All parsers are exported via the \\\init\\_.py\ for easy import.
_llama-index-core/llama\_index/core/node\parser/text · high confidence
New tool execution engine and evaluation capabilities
The tools module now includes a dedicated \calling.py\ module that standardizes how tools are invoked, handling single-argument shortcuts and ensuring context variables are propagated correctly in async paths. A new \EvalQueryEngineTool\ allows wrapping query engines with evaluators to validate responses before returning them. Additionally, \FunctionTool\ now supports callbacks for post-processing results, accepts \partial\_params\ for pre-filled arguments, and correctly respects Pydantic \Field\ defaults and \typing.Annotated\ parameter descriptions in its generated schemas.
_llama-index-core/llama\index/core/tools · high confidence
New workflow-based agent architecture with CodeAct, Function, and ReAct implementations
The \llama\_index.core.agent.workflow\ module introduces a new workflow-based agent framework, exposing \AgentWorkflow\ for orchestrating multiple agents with handoffs, alongside specific agent implementations: \CodeActAgent\ (which executes Python code via \\<execute\>\ tags), \FunctionAgent\ (supporting parallel tool calls and initial tool forcing), and \ReActAgent\ (with updated reasoning and formatting). The module also provides \AgentContext\ and \SimpleAgentContext\ for state management, and a comprehensive set of streaming and control events (\AgentStream\, \AgentInput\, \ToolCall\, etc.) to expose agent internals to users.
_llama-index-core/llama\index/core/agent/workflow · high confidence
Standalone instrumentation package for LlamaIndex observability
The instrumentation logic has been extracted from the core LlamaIndex package into a new standalone \llama-index-instrumentation\ package. This package provides the foundational hooks for observability, including a \Dispatcher\ and \Manager\ to manage spans and events, a \DispatcherSpanMixin\ for automatic span creation on class methods, and base classes for custom event and span handlers. It also includes a \SimpleSpanHandler\ for in-memory span tracking and trace tree visualization, along with support for capturing and restoring propagation contexts across process boundaries.
llama-index-instrumentation · high confidence
Architecture
Span instrumentation module relocated to external package
The span instrumentation components (BaseSpan, SimpleSpan, and active\_span\_id) have been moved from the llama-index-core package to the separate llama\_index\_instrumentation.span package. This change updates the core library to re-export these classes from the new external location, effectively decoupling the span implementation from the core distribution while maintaining the same public API surface for existing users.
_llama-index-core/llama\_index/core/instrumentation, llama-index-core/llama\index/core/instrumentation/span · high confidence
Workflow module relocated to external package with backward-compatibility shims
The \llama\_index.core.workflow\ package no longer contains the workflow implementation itself; instead, it now acts as a compatibility layer that re-exports classes, decorators, and utilities from the external \workflows\ package. This change moves the core workflow logic out of the main \llama-index-core\ distribution, requiring users to install the separate \workflows\ dependency to use workflow features. The module maintains backward compatibility by exposing the same public API (e.g., \Workflow\, \Context\, \step\, \Event\ types) through re-exports, and includes deprecation warnings for visualization functions (\draw\_all\_possible\_flows\, \draw\_most\_recent\_execution\) directing users to the \llama-index-utils-workflow\ package for those capabilities.
_llama-index-core/llama\index/core/workflow · high confidence
Behavioural changes
2274 commits (676 fixes) modifying llama-index-integrations
A change to existing behaviour in llama-index-integrations — 2274 commits (676 fixs), 6268 files: “Add HuggingFace datasets reader integration (\#20468)”; “Support AG UI protocol (\#18898)”; “Opensearch test and documentation (\#11981)”; and 2269 more (refs \#20468, \#18898, \#11981, \#19954, \#19604, …).
llama-index-integrations · medium confidence · unverified
CLI functionality moved to standalone package
The command-line interface tools, including the upgrade script and import mappings, have been removed from \llama-index-core\ and are now maintained in a separate \llama-index-cli\ package. The \llama\_index.core.command\_line\ module is now deprecated and serves only as a placeholder, meaning users must install the dedicated CLI package to access upgrade and migration utilities.
_llama-index-core/llama\_index/core/command\line · high confidence
Core library restructured with new Settings API and deprecated legacy components
The core library has been reorganized to introduce a centralized \Settings\ object for managing LLMs, embeddings, tokenizers, and node parsers, replacing the previously used \ServiceContext\. The \ServiceContext\ class and its associated helper functions are now deprecated and raise errors when instantiated, directing users to the new \Settings\ API. Additionally, the \asyncio\_module\ utility is deprecated in favor of \get\_asyncio\_module\, and the \Playground\ component is now available for comparing index outputs across different retriever modes.
_llama-index-core/llama\index/core · high confidence
Deprecated import paths for text splitters
The \llama\_index.core.text\_splitter\ module now re-exports \SentenceSplitter\, \TokenTextSplitter\, and \CodeSplitter\ from their new locations in \llama\_index.core.node\_parser.text\. This change maintains backward compatibility for existing imports but marks the old paths as deprecated, signaling that users should migrate to the new \node\_parser\ package structure in future versions.
_llama-index-core/llama\_index/core/base/response, llama-index-core/llama\_index/core/text\splitter · high confidence
Instrumentation events module restructured to external package
The instrumentation events module has been refactored to delegate its core definitions to the external \llama\_index\_instrumentation\ package. The \BaseEvent\ class is now imported from \llama\_index\_instrumentation.base\, and the \SpanDropEvent\ is imported from \llama\_index\_instrumentation.events.span\, effectively moving the implementation details out of the local \llama\_index.core.instrumentation.events\ directory while maintaining the public API surface for these specific event types.
_llama-index-core/llama\index/core/instrumentation/events · high confidence
Introduce new reader base classes and resource management capabilities
The readers module has been restructured with a new \BaseReader\ class that provides default async implementations for \load\_data\ and \lazy\_load\_data\ using \asyncio.to\_thread\, ensuring non-blocking I/O for synchronous reader logic. A new \ResourcesReaderMixin\ has been added to support resource-specific operations, including \list\_resources\, \get\_permission\_info\, and \load\_resource\, along with their async counterparts. The \JSONReader\ now includes a \clean\_json\ option to preserve valid JSON structure when needed, and handles recursion errors gracefully with warnings. Additionally, \StringIterableReader\ and \BasePydanticReader\ are introduced to support serializable, Pydantic-based reader configurations.
_llama-index-core/llama\index/core/readers · high confidence
Introduction of VectorStoreIndex with asynchronous node insertion support
The \llama\_index.core.indices.vector\_store\ module now exposes \VectorStoreIndex\ and \VectorIndexRetriever\ as the primary vector-store-based data structures, replacing the legacy \GPTVectorStoreIndex\ alias. This new index class supports asynchronous operations for node insertion and embedding, allowing users to leverage \use\_async=True\ for non-blocking performance when adding nodes to the underlying vector store.
_llama-index-core/llama\_index/core/indices/vector\store · high confidence
KnowledgeGraphIndex and KGTableRetriever are deprecated
The KnowledgeGraphIndex class and its associated KGTableRetriever are now deprecated as of version 0.10.53. Users are advised to migrate to the new PropertyGraphIndex class and its associated retrievers. The codebase includes a warning during initialization if a previously saved KG index is being upgraded to the new storage format, ensuring backward compatibility during the transition.
_llama-index-core/llama\_index/core/indices/knowledge\graph · high confidence
LLM module restructured with new base classes and multimodal support
The \llama\_index/core/llms\ package has been reorganized into a new modular structure, introducing dedicated base classes for \CustomLLM\, \FunctionCallingLLM\, and \StructuredLLM\ to standardize LLM implementations. The module now natively supports multimodal inputs by exporting \AudioBlock\, \VideoBlock\, and \DocumentBlock\ types alongside existing text and image blocks. A new \chatml\_utils\ module provides utilities for formatting ChatML-style prompts, while \loading.py\ centralizes LLM instantiation logic. Additionally, the \mock.py\ module has been expanded with \MockFunctionCallingLLM\ and enhanced memory-tracking mock classes to improve testing capabilities for tool-calling and chat workflows.
_llama-index-core/llama\index/core/llms · high confidence
Memory module restructured with new base classes and deprecated legacy buffers
The memory module has been reorganized to introduce a new \Memory\ class and \BaseMemoryBlock\ components (including \StaticMemoryBlock\, \VectorMemoryBlock\, and \FactExtractionMemoryBlock\) as the primary memory abstractions. The legacy \ChatMemoryBuffer\, \ChatSummaryMemoryBuffer\, \SimpleComposableMemory\, and \VectorMemory\ classes are now marked as deprecated in favor of the new \Memory\ class. The \BaseMemory\ and \BaseChatStoreMemory\ base classes have been updated to support async operations (e.g., \aget\, \aput\) and multi-tenant chat history storage via \BaseChatStore\. \VectorMemory\ now uses a vector index for retrieval and batches messages by user input, while \SimpleComposableMemory\ filters repeated messages from secondary sources.
_llama-index-core/llama\index/core/memory · high confidence
New global callback handler system and Pythonic logging base
The callback subsystem now supports a global handler mechanism that automatically applies a specified handler (e.g., Langfuse, AgentOps, Opik, Wandb) to all new CallbackManagers, and introduces a PythonicallyPrintingBaseHandler that uses standard Python logging instead of raw print statements for more consistent output behavior.
_llama-index-core/llama\index/core/callbacks · high confidence
New modular index store architecture with async support
The index store module has been restructured into a new, modular package under \llama\_index/core/storage/index\_store\. This introduces a \BaseIndexStore\ abstract class that enforces both synchronous and asynchronous methods for managing index structures (add, delete, get, list). A new \KVIndexStore\ implementation provides the core key-value logic, now supporting a configurable \collection\_suffix\ to customize storage namespaces. Additionally, a \SimpleIndexStore\ is provided for in-memory or file-persisted use cases, exposing convenience methods like \from\_persist\_dir\ and \persist\. Utility functions for serializing and deserializing index structures to JSON are also included.
_llama-index-core/llama\_index/core/storage/index\store · high confidence
Pandas query engine moved to experimental
The PandasQueryEngine and PandasInstructionParser have been moved from llama-index-core to the llama-index-experimental package. Attempting to import these classes from their previous locations now raises a DeprecationWarning, directing users to install llama-index-experimental and import from llama\_index.experimental.query\_engine.pandas instead. This change reflects the security risks associated with the arbitrary code execution capabilities of these tools, which require careful sandboxing.
_llama-index-core/llama\_index/core/query\engine/pandas · high confidence
Query engine module reorganization and deprecations
The \llama\_index.core.query\engine\ package has been reorganized to expose a consolidated set of query engines (including \CitationQueryEngine\, \RouterQueryEngine\, \RetrieverQueryEngine\, and \ComposableGraphQueryEngine\) via a new \\\init\\_.py\. As part of this restructuring, \JSONalyzeQueryEngine\ has been deprecated in favor of the experimental package, and \SimpleMultiModalQueryEngine\ has been deprecated in favor of using the \multimodal\ flag on standard engines like \RetrieverQueryEngine\ and \CitationQueryEngine\. Additionally, the \QASummaryQueryEngineBuilder\ has been moved to the \composability\ module to support building unified QA and summarization graphs.
_llama-index-core/llama\_index/core/query\engine · high confidence
ReAct agent output parser and prompt structure refactored
The ReAct agent's output parsing logic has been restructured to improve robustness against malformed LLM responses. The new \ReActOutputParser\ explicitly prioritizes 'Action' keywords over 'Answer' keywords to prevent the agent from getting stuck when 'Answer:' appears within an action input. It also handles cases where the LLM omits the 'Thought:' prefix or outputs a direct answer without the standard format. Additionally, the system prompts are now loaded from an external Markdown template file (\system\_header\_template.md\) with explicit UTF-8 encoding, supporting better maintainability and multilingual contexts.
_llama-index-core/llama\index/core/agent/react · high confidence
Refactored VectorIndexRetriever with optimized embedding generation and instrumentation
The VectorIndexRetriever has been refactored to optimize query performance and improve observability. Embeddings are now generated only for query modes that require them (excluding TEXT\_SEARCH and SPARSE modes), preventing unnecessary computation. The retriever now supports hybrid search parameters (sparse\_top\_k, hybrid\_top\_k) and integrates with the new instrumentation module for better tracing. Additionally, the logic for fetching nodes from the docstore has been refined to handle cases where the vector store does not store text, ensuring correct node retrieval and replacement.
_llama-index-core/llama\_index/core/indices/vector\store/retrievers · high confidence
Refactored document storage with new KV-based architecture and serialization fixes
The document store implementation has been refactored to use a key-value store backend (KVDocumentStore) with configurable collection suffixes for nodes, reference documents, and metadata, replacing the previous structure. This change includes a new registry for document store types and updated serialization utilities that distinguish between modern and legacy JSON formats to prevent data loss during deserialization. Users benefit from more flexible storage configurations and improved reliability when loading previously stored documents, particularly those with reference document relationships.
_llama-index-core/llama\index/core/storage/docstore · high confidence
Reorganized index module structure and added PropertyGraphIndex
The \llama\_index.core.indices\ package has been restructured into a modular layout with dedicated subpackages for base classes, loading, registries, and common utilities, while explicitly exporting the new \PropertyGraphIndex\ alongside existing index types like \VectorStoreIndex\ and \KnowledgeGraphIndex\. This change introduces a central registry (\INDEX\_STRUCT\_TYPE\_TO\_INDEX\_CLASS\) to map index types to their implementations and adds a \load\_graph\_from\_storage\ function to reconstruct composable graphs from persisted storage, enabling users to load complex multi-index structures more easily.
_llama-index-core/llama\index/core/indices · high confidence
Repository migration to uv and updated developer tooling
The project has migrated its build and dependency management from Poetry to uv, introducing a \uv.lock\ file and updating the \Makefile\ to use \uv run\ for linting and testing. The \CONTRIBUTING.md\ guide has been rewritten to reflect this new workflow, and the \.pre-commit-config.yaml\ has been updated with new hook versions (e.g., ruff v0.11.8) and exclusions for the new lock file. Additionally, the \.gitignore\ has been updated to exclude \uv.lock\ and other build artifacts.
(repo-wide) · high confidence
Restructured Structured Store API with SQL Context Builder and JSON Query Engine
The \llama\_index.core.indices.struct\_store\ module has been reorganized to provide a cleaner public API and new capabilities for structured data. A new \SQLContextContainerBuilder\ class has been added to help construct SQL context containers from database schemas or documents, which can then be used to derive indices or query for context. The \JSONQueryEngine\ is now fully integrated into the core module, allowing users to convert natural language queries into JSON path expressions and synthesize responses from JSON data. Additionally, the legacy \PandasIndex\ has been deprecated in favor of \PandasQueryEngine\ in the experimental package, and the SQL query engines (\NLStructStoreQueryEngine\, \SQLStructStoreQueryEngine\) have been updated to support new prompt templates and improved error handling for SQL generation.
_llama-index-core/llama\_index/core/indices/struct\store · high confidence
SimpleDirectoryReader now supports fsspec and handles metadata robustly
The SimpleDirectoryReader in llama-index-core now supports reading from any filesystem backed by fsspec (e.g., S3, GCS) rather than just the local disk. It also includes improved metadata handling, such as filtering out null values in default file metadata and consistent timezone-aware timestamp formatting. Additionally, the reader now supports Path objects, excludes empty files by default, and uses fsspec for file content reading to prevent excessive memory consumption.
_llama-index-core/llama\index/core/readers/file · high confidence
Span handlers are re-exported from the core instrumentation package
The \llama\_index.core.instrumentation.span\_handlers\ module now acts as a re-export layer for span handler classes (\BaseSpanHandler\, \NullSpanHandler\, \SimpleSpanHandler\) originally defined in the external \llama\_index\_instrumentation\ package. This change allows users to import these handlers directly from the core library namespace, simplifying access to instrumentation components without needing to reference the separate instrumentation package.
_llama-index-core/llama\_index/core/instrumentation/span\handlers · high confidence
Test coverage
Added comprehensive test coverage for text splitters; Added mock utilities for testing; Added placeholder test directories for llama-index integrations, packs, and utils; Added test coverage for TreeIndex and its retrievers; Added test coverage for embeddings module; Added test coverage for evaluation metrics, batch runner, and dataset generation; Added test coverage for indices loading, prompt helper, and utilities; Added test coverage for node parsers; Added test coverage for node postprocessors; Added test coverage for program utilities and LLM programs; Added test coverage for prompt templates and utilities; Added test coverage for property graph index components; Added test coverage for schema components and multimodal support; Added test runner infrastructure for llama-dev; Added test suite and test infrastructure for llama-index-core; Added test suite for llama-dev CLI and utility functions; Added tests for BaseToolSpec and LoadAndSearchToolSpec; Added tests for ChatMessage block-based storage and multimodal content types; Added tests for DocumentContextExtractor and BaseExtractor resilience; Added tests for DocumentSummaryIndex and its retrievers; Added tests for LLM core behaviors and mock implementations; Added tests for LlamaDebugHandler and TokenCountingHandler; Added tests for MockSparseEmbedding; Added tests for Object Index and Node Mapping components; Added tests for ReAct agent prompt customization and output parsing; Added tests for SQL wrapper schema prefixing and CTE handling; Added tests for SimpleChatStore and SQLAlchemyChatStore; Added tests for SimpleDocumentStore and legacy JSON deserialization; Added tests for SimpleKVStore and MutableMappingKVStore; Added tests for SimplePropertyGraphStore persistence and operations; Added tests for SummaryIndex and its retrievers; Added tests for VectorMemory and SimpleComposableMemory; Added tests for VectorStoreQueryOutputParser validation; Added tests for agent utility functions; Added tests for chat engine history persistence and multi-modal support; Added tests for empty index, playground, and LLM selectors; Added tests for fact, static, and vector memory blocks; Added tests for ingestion pipeline components; Added tests for llama-dev CLI package commands; Added tests for memory system components and behaviors; Added tests for multimodal and router query engines; Added tests for query index composition and MMR embedding diversity; Added tests for release changelog and check commands; Added tests for response synthesizers; Added tests for response synthesizers and tree summarization; Added tests for retriever metadata preservation, deduplication, and async query generation; Added tests for storage serialization and fixtures; Added tests for struct store indices and query engines; Added tests for the LangChain bridge module's optional dependency handling; Added tests for the llama-dev test CLI command; Added tests for vector store index retrieval and deletion behavior; Added tests for vector store metadata filtering and node serialization; Added tests for voice agent event serialization and subclassing; Added unit tests for JSON, StringIterable, and reader loading functionality; Added unit tests for KnowledgeGraphIndex and KGTableRetriever; Added unit tests for SimpleDirectoryReader; Added unit tests for multi-modal LLM metadata and utility functions; Initial test suite for core tool abstractions.
Dependencies
2612 commits updating dependencies (723 manifests)
A dependency / build maintenance change in (dependencies) — 2612 commits (691 fixs), 723 files: “Add HuggingFace datasets reader integration (\#20468)”; “v0.12.28 (\#18340)”; “Support AG UI protocol (\#18898)”; and 2609 more (refs \#20468, \#18340, \#18898, \#10233, \#19954, …).
(dependencies) · high confidence · unverified
Written by cca.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.
How this codebase got here
This is the PUBLIC form of this artifact. Findings are listed in full, but the details of SECURITY findings — which rule fired, in which file, on which line, and how to fix it — are deliberately withheld, and any secret-scanner results are excluded entirely. Where detail is absent here it was REMOVED FOR PUBLICATION; it is not missing from the analysis. The complete artifact is available from the repository owner.
Score
- CAI 56 → 58 (+2.2)
- Rubric changed (rubric-2026.10.5 → rubric-2026.11.4) — scores are not directly comparable.
Lenses
- Code Health 60 → 56 (-3.5)
- Architecture 98 → 98 (+0.1)
- Maturity 62 → 67 (+5.5)
- Readiness 47 → 52 (+5.0)
- Security 67 → 74 (+6.6)
- Domain Modelling 59 → 59 (+0.0)
- Performance 100 → 100 (+0.0)
Resolved (79)
- Documentation: no project overview (llama-index-integrations/callbacks/llama-index-callbacks-wandb/README.md)
- Duplicated block (10 lines × 2) (llama-index-integrations/embeddings/llama-index-embeddings-nvidia/llama_index/embeddings/nvidia/utils.py)
- Duplicated block (11 lines × 2) (llama-index-integrations/graph_stores/llama-index-graph-stores-nebula/llama_index/graph_stores/nebula/utils.py)
- Duplicated block (11–14 lines × 2) (llama-index-integrations/llms/llama-index-llms-upstage/llama_index/llms/upstage/base.py)
- Duplicated block (12 lines × 2) (llama-index-core/llama_index/core/evaluation/batch_runner.py)
- Duplicated block (12–14 lines × 2) (llama-index-integrations/retrievers/llama-index-retrievers-mongodb-atlas-bm25-retriever/llama_index/retrievers/mongodb_atlas_bm25_retriever/base.py)
- Duplicated block (14 lines × 2) (llama-index-integrations/tools/llama-index-tools-aws-bedrock-agentcore/llama_index/tools/aws_bedrock_agentcore/browser/base.py)
- Duplicated block (19 lines × 2) (llama-index-integrations/embeddings/llama-index-embeddings-bedrock/llama_index/embeddings/bedrock/base.py)
- Duplicated block (20 lines × 2) (llama-index-integrations/tools/llama-index-tools-parallel-web-systems/llama_index/tools/parallel_web_systems/base.py)
- Duplicated block (6 lines × 2) (llama-index-integrations/tools/llama-index-tools-google/llama_index/tools/google/calendar/base.py)
- Duplicated block (7 lines × 2) (llama-index-integrations/vector_stores/llama-index-vector-stores-faiss/llama_index/vector_stores/faiss/base.py)
- Duplicated block (8 lines × 2) (llama-index-integrations/postprocessor/llama-index-postprocessor-nvidia-rerank/llama_index/postprocessor/nvidia_rerank/base.py)
- Duplicated block (9 lines × 2) (llama-index-integrations/readers/llama-index-readers-database/llama_index/readers/database/base.py)
- Duplicated block (9–11 lines × 2) (llama-index-integrations/vector_stores/llama-index-vector-stores-moorcheh/llama_index/vector_stores/moorcheh/base.py)
- Edited copy of a member (15 corresponding lines) (llama-index-integrations/tools/llama-index-tools-aws-bedrock-agentcore/llama_index/tools/aws_bedrock_agentcore/browser/base.py)
- Inverted test pyramid
- Leaked secret: hardcoded-credential (llama-index-integrations/embeddings/llama-index-embeddings-gaudi/examples/graphrag.py)
- Medium: security finding (details withheld)
- Members sharing a duplicated core (4 members, 50+ identical tokens) (llama-index-integrations/graph_stores/llama-index-graph-stores-falkordb/llama_index/graph_stores/falkordb/falkordb_property_graph.py)
- Members sharing a duplicated core (5 members, 50+ identical tokens) (llama-index-integrations/graph_stores/llama-index-graph-stores-falkordb/llama_index/graph_stores/falkordb/falkordb_property_graph.py)
- …and 59 more
New (2781)
- Critical CVE: [GHSA redacted] (llama-index-integrations/output_parsers/llama-index-output-parsers-guardrails/uv.lock)
- Critical CVE: [GHSA redacted] (llama-index-integrations/readers/llama-index-readers-lilac/uv.lock)
- Critical CVE: [GHSA redacted] (llama-index-integrations/readers/llama-index-readers-microsoft-outlook-emails/uv.lock)
- Critical CVE: [GHSA redacted] (llama-index-integrations/readers/llama-index-readers-pdf-marker/uv.lock)
- Critical CVE: [GHSA redacted] (llama-index-integrations/response_synthesizers/llama-index-response-synthesizers-google/uv.lock)
- Critical CVE: [GHSA redacted] (llama-index-integrations/retrievers/llama-index-retrievers-duckdb-retriever/uv.lock)
- Critical CVE: [GHSA redacted] (llama-index-integrations/retrievers/llama-index-retrievers-mongodb-atlas-bm25-retriever/uv.lock)
- Critical CVE: [GHSA redacted] (llama-index-integrations/sparse_embeddings/llama-index-sparse-embeddings-fastembed/uv.lock)
- Critical CVE: [GHSA redacted] (llama-index-integrations/storage/chat_store/llama-index-storage-chat-store-azure/uv.lock)
- Critical CVE: [GHSA redacted] (llama-index-integrations/storage/chat_store/llama-index-storage-chat-store-azurecosmosmongovcore/uv.lock)
- Critical CVE: [GHSA redacted] (llama-index-integrations/storage/chat_store/llama-index-storage-chat-store-azurecosmosnosql/uv.lock)
- Critical CVE: [GHSA redacted] (llama-index-integrations/storage/chat_store/llama-index-storage-chat-store-dynamodb/uv.lock)
- Critical CVE: [GHSA redacted] (llama-index-integrations/storage/chat_store/llama-index-storage-chat-store-gel/uv.lock)
- Critical CVE: [GHSA redacted] (llama-index-integrations/storage/chat_store/llama-index-storage-chat-store-mongo/uv.lock)
- Critical CVE: [GHSA redacted] (llama-index-integrations/storage/chat_store/llama-index-storage-chat-store-postgres/uv.lock)
- Critical CVE: [GHSA redacted] (llama-index-integrations/storage/chat_store/llama-index-storage-chat-store-sqlite/uv.lock)
- Critical CVE: [GHSA redacted] (llama-index-integrations/storage/chat_store/llama-index-storage-chat-store-tablestore/uv.lock)
- Critical CVE: [GHSA redacted] (llama-index-integrations/storage/chat_store/llama-index-storage-chat-store-upstash/uv.lock)
- Critical CVE: [GHSA redacted] (llama-index-integrations/storage/chat_store/llama-index-storage-chat-store-yugabytedb/uv.lock)
- Critical CVE: [GHSA redacted] (llama-index-integrations/storage/docstore/llama-index-storage-docstore-azure/uv.lock)
- …and 2761 more
Changes since last survey
- 7 commits — 5 feature/other, 2 fixes
By area
- llama-index-integrations/readers — 3 commits
- llama-index-integrations/llms — 2 commits
- (root) — 1 commit
- docs/src — 1 commit
Notable commits
- fix: BREAKING: raise requires-python for security fixes across integrations (stacked on #24039) (#24040)
- fix: fix(llms-openai): allow openai 3.x (#24138)
- change: Make fake credentials obviously fake (place-/holder- prefix) (#23923)
- change: Python security relock: all integration uv.lock files (lock files only, no Python floor changes) (#24039)
- change: Release 0.14.26 (#24189)
- change: deps: allow llama-index-llms-openai 0.8.x in the 15 integration packages still pinning <0.8 (#24125)
- change: docs: update Arize agent evaluation guide URL (#24017)
Architecture
- Unchanged — 0 containers · 1 contexts · 0 edges
Written by cca.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.
Survey your own repository
run-llama/llama_index was measured the same way every project in this corpus was: the same rubric, at a pinned commit, with the result published in full. Point a surveyor at a repository you know and see whether you agree with it.
About this page
- The score is its most recent published measurement, taken on 14 November 2026 at a pinned commit. It is not a live figure and does not change until the project is measured again.
- Measured at commit 3aaab93a2e823b9637ce3c6f0611e5e18a4a7a5b — the exact code this score is about.
- Scored under rubric-2026.11.4 — the same rubric and the same method as every other entry in this index.
- Measured by Canine Code Analyzer (cca.canine.dev) using cca measure-2026.11.65+preprod-36d4cd83012f.