← Back
LlamaIndex 0.14.15 adds multimodal support and recursive LLM types
· releasefeatureapibugfixintegration · github.com ↗

Multimodal Type Support

LlamaIndex core 0.14.15 significantly expands multimodal capabilities with support for basic operations on multimodal types. This includes new multimodal template variable formatting and multimodal prompt templates, enabling developers to work with non-text content more seamlessly.

Recursive LLM Types

The release introduces support for recursive LLM type definitions, allowing for more complex type hierarchies and nested structures in LLM interactions.

Enhanced Prompt Templating

Three new features improve prompt handling:

  • Multimodal template variable formatting for flexible content insertion
  • Multimodal prompt templates for structured multimodal interactions
  • Multimodal chat prompt helper to simplify building multimodal conversations

Robustness Improvements

  • BaseExtractor retry and error handling (#20693) adds resilience to data extraction operations
  • Memory fix ensures at least one message/content block is returned, preventing empty response issues
  • TextNode cleanup removes redundant metadata_separator field

Integration Updates

Beyond core changes, this release includes updates across the ecosystem:

  • Anthropic LLM: Added Claude Sonnet 4.6 support
  • Bedrock Converse: Added Sonnet 4.6 and fixed thinking_delta population
  • Mistral AI: Updated to latest Azure SDK
  • OCI DataScience: New streaming endpoint support
  • OpenTelemetry: Improved data serialization with custom span processor support