Mistral YAML Configuration - Official API Alignment

Overview

The mistral.yaml configuration file has been updated to align with the official Mistral AI API specification (OpenAPI 3.1.0). This document explains how our configuration maps to the official API capabilities and model specifications.

Official API Capabilities Mapping

Model Capabilities (from ModelCapabilities schema)

Our YAML configuration now includes comprehensive capability flags that map directly to the official API:

supports_streaming: true          # Maps to completion_chat streaming
supports_function_calling: true   # Maps to function_calling capability
supports_fim: true               # Maps to completion_fim capability
supports_vision: true            # Maps to vision capability
supports_classification: false   # Maps to classification capability
supports_fine_tuning: true       # Maps to fine_tuning capability

Model Types

Based on the official API, we now distinguish between:

Base Models (model_type: base): Standard Mistral models

Fine-tuned Models (model_type: fine-tuned): Custom fine-tuned models

Embedding Models: Specialized for vector embeddings

Official Model Registry

Our configuration includes all models from the official Mistral AI API:

Chat Completion Models

mistral-large-latest - Latest large model

mistral-small-latest - Latest small model

mistral-medium-latest - Latest medium model

mistral-nemo-latest - Latest Nemo model

Code Generation Models (FIM)

codestral-latest - Latest Codestral model

codestral-22b-latest - 22B parameter Codestral

codestral-2405 - Specific Codestral version

Vision Models

pixtral-12b-latest - 12B parameter vision model

Embedding Models

mistral-embed - Standard embedding model

mistral-embed-v2 - Version 2 embedding model

Fine-tunable Models

ministral-3b-latest - 3B parameter fine-tunable

ministral-8b-latest - 8B parameter fine-tunable

open-mistral-7b - Open source 7B model

open-mistral-nemo - Open source Nemo model

API Endpoint Mapping

Chat Completions (/v1/chat/completions)

assessed_capabilities:
  - chat_completion
  - streaming
  - function_calling

FIM Completions (/v1/fim/completions)

assessed_capabilities:
  - fim_completion
  - fill_in_middle

Embeddings (/v1/embeddings)

assessed_capabilities:
  - embedding
  - text_embedding
  - vector_search

Agents API (/v1/agents/*)

assessed_capabilities:
  - chat_completion
  - function_calling
  - streaming

OCR API (/v1/ocr)

assessed_capabilities:
  - vision
  - image_analysis
  - ocr

Enhanced Model Information

Comprehensive Capability Flags

Each model now includes detailed capability information:

mistral/mistral-large-latest:
  # Core capabilities
  supports_streaming: true
  supports_function_calling: true
  supports_fim: false
  supports_vision: false
  supports_classification: false
  supports_fine_tuning: false
  
  # Model metadata
  model_type: base
  owned_by: mistralai
  api_name: mistral-large-latest
  
  # Comprehensive capabilities
  assessed_capabilities:
    - text
    - reasoning
    - code_generation
    - multilingual
    - chat_completion
    - function_calling
    - streaming

Task-Specific Scoring

Our task scoring system aligns with common use cases:

reasoning: Complex problem-solving tasks

code_generation: Programming and development

writing: Content creation and editing

simple_chat: Basic conversational AI

data_analysis: Data processing and analysis

speed_sensitive: Low-latency requirements

Official API Integration Benefits

1. Accurate Model Selection

The enhanced configuration enables precise model selection based on:

Required capabilities (vision, FIM, function calling)

Performance requirements (reasoning scores)

Cost optimization (pricing information)

Context length requirements

2. API Compatibility

Direct mapping to official API endpoints ensures:

Correct model names for API calls

Proper capability validation

Accurate parameter handling

3. Future-Proof Design

The configuration structure supports:

New model additions

Capability updates

API version changes

Fine-tuned model integration

Usage Examples

Model Selection by Capability

# Select best model for vision tasks
vision_models = [model for model in models 
                if model.get('supports_vision', False)]

Select best model for code generation
code_models = [model for model in models 
              if 'fim_completion' in model.get('assessed_capabilities', [])]

Select fine-tunable models
fine_tunable = [model for model in models 
               if model.get('supports_fine_tuning', False)]

Task-Based Selection

# Get best model for reasoning tasks
best_reasoning = max(models, key=lambda m: m['task_scores']['reasoning'])

Get fastest model for speed-sensitive tasks
fastest = max(models, key=lambda m: m['task_scores']['speed_sensitive'])

Configuration Validation

The updated configuration includes validation against the official API:

Model Name Validation: All model names match official API

Capability Validation: Capabilities align with API specifications

Pricing Validation: Costs reflect official pricing

Context Length Validation: Max context lengths are accurate

Migration from Previous Version

Added Fields

supports_fim: Fill-in-the-middle capability

supports_vision: Vision/image processing capability

supports_classification: Content classification capability

supports_fine_tuning: Fine-tuning capability

model_type: Base or fine-tuned model type

owned_by: Model ownership information

Enhanced Capabilities

More comprehensive assessed_capabilities lists

Better task scoring granularity

Official API endpoint mapping

Backward Compatibility

All existing fields maintained

Existing functionality preserved

Enhanced with additional information

Conclusion

The updated mistral.yaml configuration provides comprehensive alignment with the official Mistral AI API specification, enabling:

Accurate Model Selection: Based on official capabilities

API Compatibility: Direct mapping to official endpoints

Future-Proof Design: Supports new models and capabilities

Enhanced Integration: Better model selection and routing

This alignment ensures that the MindX system can fully leverage Mistral AI's capabilities while maintaining compatibility with the official API specification.

---