●Change8

v0.14.23

📦 llamaindexView on GitHub →
✨ 3 features🐛 12 fixes🔧 6 symbols

Summary

This release introduces significant multimodal capabilities, including synthesis and query engines, alongside numerous bug fixes in core functionality like file I/O encoding and splitter recursion errors. Numerous dependency updates were also applied across various packages.

Migration Steps

  1. Ensure that file I/O operations in llama-index-core explicitly use encoding='utf-8' if relying on default behavior that might have been inconsistent.

✨ New Features

  • Multimodal synthesis part 2 implemented in llama-index-core.
  • Multimodal query engines added.
  • Tool calling mock LLM added.

🐛 Bug Fixes

  • Added DocumentBlock and VideoBlock to FunctionTool._parse_tool_output.
  • Guarded against ZeroDivisionError on empty input sequences in prompt_helper.
  • Preserved URL-backed video and document memory blocks.
  • Added explicit encoding='utf-8' to llama-index-core text-mode file I/O.
  • Fixed refresh_ref_docs kwargs reuse.
  • Used a set instead of a list for within-batch dedup in Ingestion pipeline.
  • Used running loop in ingestion pipeline.
  • Preserved IndexNode obj during model dump.
  • Deep copied initial_state in workflow to prevent mutation leaks across runs.
  • Matched missing metadata for NE and NIN filters.
  • Preserved TreeSelectLeafRetriever source nodes.
  • Fixed RecursionError in TokenTextSplitter & SentenceSplitter for units larger than chunk_size.

Affected Symbols

FunctionTool._parse_tool_outputprompt_helperTokenTextSplitterSentenceSplitterIndexNodeTreeSelectLeafRetriever