LlamaIndex v0.14.23 adds multimodal synthesis and fixes key bugs

LlamaIndex v0.14.23, released June 24, 2026, brings significant enhancements to its core framework for building retrieval-augmented generation (RAG) applications. The headline feature is Multimodal synthesis part 2 (feat(core)), which expands the framework's ability to process and generate across multiple modalities — text, images, audio, video. This is complemented by new Multimodal query engines (feat(core)), allowing developers to query across diverse data types seamlessly. Additional improvements include adding DocumentBlock and VideoBlock support to FunctionTool._parse_tool_output, enabling better handling of rich media outputs. The release also addresses critical stability issues: a deep copy of initial_state in workflows to prevent mutation leaks across runs, preservation of TreeSelectLeafRetriever source nodes, and fixing a RecursionError in TokenTextSplitter & SentenceSplitter when units exceed chunk_size.

On the performance front, ingestion now uses a set instead of a list for within-batch deduplication, reducing overhead. Other fixes include handling missing metadata for NE (not equal) and NIN (not in) filters, and simplifying serialized payloads to instrumentation. Across the ecosystem, over 35 dependency directories received updates (uv and pip groups) to ensure security and compatibility with the latest libraries. Adapter, embedding, and callback packages also saw version bumps with minor fixes (e.g., explicit UTF-8 encoding for file I/O). This release solidifies LlamaIndex as a robust choice for multimodal RAG workflows, especially for enterprise applications requiring reliable, scalable data processing.