agentjido / req_llm / 01962a6c2718d975781974570b45bc695b762489-PR-147

53%

main: 49%

Build Type

Pull #147

github

Committed by neilberkman

Commit Message

Add Mistral AI model support and improve Bedrock formatter architecture

Introduces full support for 5 Mistral AI models on AWS Bedrock (7B, 8x7B,
Small, Large, and Pixtral Large) with comprehensive test coverage.

Generic improvements:
- Adds callback system for model-specific capabilities in Bedrock formatters:
  * requires_converse_api?/0 - formatters declare Converse API requirement
  * supports_converse_tool_choice?/0 - formatters declare toolChoice support
  * preserve_inference_profile?/1 - formatters control ID normalization
- Fixes formatter selection logic when using Converse API to properly delegate
  to family-specific formatters only when they wrap Converse (e.g., Mistral)
- Resolves incorrect formatter selection for Anthropic models with use_converse: true

Mistral-specific:
- Implements ReqLLM.Providers.AmazonBedrock.Mistral formatter wrapping Converse API
- Handles system message incompatibility by prepending to first user message
- Preserves inference profile prefix for Pixtral (region-specific invocation required)
- Sets tool_call: false for models lacking toolChoice support in Converse API
- Adds comprehensive streaming fixtures for all 5 models

All models pass comprehensive test suite at 100% coverage.

Pull Request Pull Request #147: Fix Bedrock streaming, tool round-trip, and usage metadata for 1.0

Run Details

52 of 66 new or added lines in 5 files covered. (78.79%)

253 existing lines in 6 files now uncovered.

3725 of 7077 relevant lines covered (52.64%)

220.07 hits per line