Added #

• New examples and restructured the example directory for better clarity.
• Comprehensive Test Suite: Added extensive test coverage for core functionality
• UTF-8 Stream Decoder: Robust handling of multi-byte characters in streaming responses
  • Utf8StreamDecoder class for intelligent buffering of incomplete UTF-8 byte sequences
  • Prevents FormatException: Unfinished UTF-8 octet sequence errors in streaming

Fixed #

• UTF-8 Streaming Issues: Complete resolution of multi-byte character encoding problems
  • Fixed garbled text output (e.g., ä½ æ → 你好) in streaming responses
  • Updated all provider clients (OpenAI, Anthropic, DeepSeek, Groq, Google, xAI, Ollama) to use UTF-8 stream decoder
  • Proper handling of Chinese, Japanese, Korean, Arabic, and emoji characters in streams
  • Eliminated FormatException: Unfinished UTF-8 octet sequence errors when multi-byte characters are split across network chunks
• BaseProviderFactory: Improved validation logic in validateConfigWithDetails method

Added #

• Unified Web Search API: Provider-agnostic web search functionality across multiple LLM providers

  • WebSearchConfig - Universal configuration class supporting all provider-specific parameters
  • WebSearchLocation - Geographic location configuration for localized search results
  • WebSearchType - Search type enumeration (web, news, academic, combined)
  • WebSearchStrategy - Implementation strategy control (native, tool, plugin, parameter, auto)
  • WebSearchContextSize - Context size control for providers that support it
  • Provider Support:
    • xAI Grok: Live Search with search_parameters (mode, sources, date filtering, result limits)
    • Anthropic Claude: Web Search Tool with domain filtering and location-based search
    • OpenAI: Web Search with context size control for gpt-4o-search-preview models
    • OpenRouter: Plugin-based search with custom prompts and :online model shortcuts
    • Perplexity: Native search capabilities with context size control
  • Ergonomic Builder Methods:
    • enableWebSearch() - Simple web search activation
    • quickWebSearch() - Fast configuration with common settings
    • webSearch() - Advanced configuration with full parameter control
    • newsSearch() - News-specific search configuration
    • searchLocation() - Geographic context configuration
    • Provider-specific methods: openaiWebSearch(), openRouterWebSearch(), perplexityWebSearch()
    • advancedWebSearch() - Full control over all search parameters
  • Automatic Provider Adaptation: Same API automatically translates to provider-specific formats
  • Rich Configuration Options: Domain filtering, result limits, date ranges, geographic context
  • Type Safety: Strong typing with enums and configuration classes
  • Comprehensive Examples: Complete demonstration in 02_core_features/web_search.dart

• Enhanced LLM Capability System: Extended capability detection for web search

  • LLMCapability.liveSearch - New capability for real-time web search functionality
  • Updated provider capability declarations to include search support
  • Enhanced capability factory methods to support web search providers

• Capability Factory Methods: Type-safe provider building with compile-time capability checking

  • buildAudio() → AudioCapability - Build providers with audio capabilities
  • buildImageGeneration() → ImageGenerationCapability - Build providers with image generation
  • buildEmbedding() → EmbeddingCapability - Build providers with embedding capabilities
  • buildFileManagement() → FileManagementCapability - Build providers with file management
  • buildModeration() → ModerationCapability - Build providers with moderation capabilities
  • buildAssistant() → AssistantCapability - Build providers with assistant capabilities
  • buildModelListing() → ModelListingCapability - Build providers with model listing
  • Eliminates runtime type casting and provides compile-time type safety
  • Clear error messages with UnsupportedCapabilityError when capabilities are not supported
  • Better IDE support and autocomplete for capability-specific methods

• Enhanced Error Handling: New error types and improved error handling examples

  • UnsupportedCapabilityError - Thrown when building providers with unsupported capabilities
  • Updated error_handling.dart example to demonstrate capability factory method errors
  • Clear error messages listing supported providers for each capability

• Core Features Examples: New capability detection and model listing examples

  • capability_detection.dart - Demonstrates provider capability discovery and comparison
  • model_listing.dart - Shows how to explore and filter available models from providers
  • capability_factory_methods.dart - Comprehensive demonstration of type-safe capability building

Changed #

• xAI Provider: Enhanced with comprehensive Live Search support

  • Updated XAIConfig to include liveSearch boolean parameter for simple activation
  • Enhanced SearchParameters class with factory methods and better documentation
  • Improved _buildSearchParameters() method with automatic default configuration
  • Added isLiveSearchEnabled getter for configuration validation
  • Updated provider capabilities to include LLMCapability.liveSearch
  • Enhanced factory methods and convenience functions to support Live Search

• LLM Builder: Significantly improved web search ergonomics

  • Replaced provider-specific search methods with unified, ergonomic API
  • Added comprehensive web search configuration methods with clear documentation
  • Improved method naming for better developer experience
  • Enhanced parameter validation and type safety
  • Better integration with provider-specific implementations

• Core Features Examples: Enhanced with web search demonstration

  • Added web_search.dart to core features examples
  • Updated 02_core_features/README.md with web search documentation
  • Comprehensive examples showing unified API across all supported providers
  • Step-by-step learning path including web search integration

• Provider Examples: Updated to use new capability factory methods

  • elevenlabs/audio_capabilities.dart - Now uses buildAudio() for type-safe audio provider building
  • openai/image_generation.dart - Now uses buildImageGeneration() for type-safe image provider building
  • Demonstrates migration from runtime type casting to compile-time type safety

Fixed #

• Provider Capabilities: Standardized capability interface implementations across all providers
• Legacy Methods Cleanup: Removed outdated capability checking methods
  • Removed supportedCapabilitiesLegacy (string-based capability lists)
  • Removed supportsCapability(String) methods across all providers
  • Unified capability checking through ProviderCapabilities interface

Changed #

• Web Platform Support: Removed dart:io dependency to enable Web and WASM compatibility
  • Removed uploadFileFromPath() method from OpenAI Files API
  • Removed filePath property from ImageInput class for cross-platform compatibility
  • Enables WASM compatibility for future Dart Web applications
  • Use uploadFile() or uploadFileFromBytes() methods instead for cross-platform file uploads
  • Use ImageInput.fromBytes() or ImageInput.fromUrl() instead of file path-based image inputs

Added #

• Anthropic Files API: Complete file management support

  • Upload, list, retrieve, and delete files
  • Support for MIME types and downloadable flags
  • Cursor-based pagination with beforeId, afterId, and limit
  • Convenience methods: uploadFileFromBytes(), fileExists(), getFileContentAsString()
  • Batch operations: deleteFiles(), getTotalStorageUsed()
  • Beta API header support (anthropic-beta: files-api-2025-04-14)

• File Management API: Cross-provider file operations

  • FileManagementCapability interface for consistent file operations
  • Universal FileObject, FileUploadRequest, FileListResponse models
  • Support for both OpenAI and Anthropic file formats with automatic conversion
  • Unified FilePurpose and FileStatus enums
  • Provider-agnostic file operations with format adaptation

• Provider Configuration Centralization: Extracted default configurations

  • New ProviderDefaults class with all provider endpoints and models
  • OpenAICompatibleDefaults for OpenAI-compatible provider configurations
  • Centralized capability definitions for all providers
  • Eliminated configuration duplication across factory classes

• Unified Audio Capability Interface: Revolutionary audio processing design

  • Single AudioCapability interface for all audio operations (TTS, STT, translation)
  • Feature discovery system with supportedFeatures property for runtime capability detection
  • BaseAudioCapability class providing default implementations for convenience methods
  • Support for streaming TTS, real-time audio sessions, and advanced audio features
  • Enhanced audio models with character-level timing, speaker diarization, and audio events
  • Graceful degradation with UnsupportedError for unsupported features
  • Cross-provider audio functionality comparison and benchmarking support

• DALL-E Image Generation Support: Complete OpenAI image API implementation

  • Image generation, editing, and variations with DALL-E 2/3
  • Enhanced ImageGenerationCapability interface with new methods
  • Support for multiple formats, sizes, and quality options

Changed #

• All Providers: Refactored to modular architecture

  • OpenAI, Anthropic, DeepSeek, Groq, xAI, Phind, ElevenLabs providers now use modular design
  • Consistent file structure across all providers
  • Improved separation of concerns (config, client, capabilities)
  • Better error handling and logging

• Ollama Provider: Enhanced implementation with full API compliance

  • Fixed temperature parameter handling (was incorrectly excluded)
  • Added complete Ollama-specific parameter support (numCtx, numGpu, numThread, numa, numBatch, keepAlive, raw)
  • Enhanced multimodal support with automatic base64 image conversion
  • Improved message type handling and tool calling support
  • Added LLMBuilder convenience methods for Ollama-specific parameters
  • Full compatibility with official Ollama API specification

• Streaming Configuration: Removed stream parameters from configs

  • Stream behavior now controlled at method call time (chat() vs chatStream())
  • Simplified configuration classes by removing redundant stream parameters
  • More intuitive API design for streaming operations

• Factory Classes: Enhanced with full configuration support

  • All factory methods now support complete provider configurations
  • Removed redundant helper methods in favor of base factory functionality
  • Better parameter validation and error handling

• Audio Capabilities: Completely redesigned audio processing architecture

  • Replaced separate TextToSpeechCapability and SpeechToTextCapability interfaces
  • Unified all audio operations under single AudioCapability interface
  • Enhanced audio models with advanced features (timing, diarization, events)
  • Improved OpenAI audio support with translation capabilities
  • Enhanced ElevenLabs audio support with streaming and real-time features
  • Better error handling and feature detection for audio operations

• Image Generation: Enhanced OpenAI provider with complete DALL-E support

  • Extended ImageGenerationCapability interface with editing and variation methods
  • Updated multi-modal examples with real image generation implementations

Removed #

• OpenAI Compatible Provider: Removed OpenAICompatibleProvider class

  • Functionality replaced by convenience functions in modular implementation
  • createDeepSeekProvider(), createGroqProvider(), etc. provide same functionality
  • Simplified architecture with better performance

• Legacy Provider Files: Cleaned up deprecated implementations

  • Removed old provider files after modular refactoring
  • Eliminated duplicate code and inconsistent implementations
  • Streamlined provider architecture

• Stream Configuration Parameters: Removed from all provider configs

  • stream parameter removed from configuration classes
  • Streaming behavior now determined by method choice
  • Cleaner separation between configuration and runtime behavior

Migration Guide #

• Streaming: Update streaming usage pattern

  • Replace config.stream = true with direct method calls
  • Use provider.chatStream() instead of provider.chat() with stream config
  • More explicit and intuitive streaming control

• File Management: Use unified file API for cross-provider compatibility

  • OpenAI file operations remain unchanged
  • Anthropic now supports file operations through FileManagementCapability
  • Use universal file models for provider-agnostic code

• Audio Capabilities: Migrate to unified audio interface

  • Replace TextToSpeechCapability and SpeechToTextCapability checks with AudioCapability
  • Use provider.supportedFeatures.contains(AudioFeature.textToSpeech) for feature detection
  • Audio translation now available through translateAudio() method (OpenAI only)
  • Enhanced audio models support advanced features like character timing and speaker diarization
  • Convenience methods (speech(), transcribe(), translate()) automatically available

• Image Generation: Enhanced capabilities with new features

  • New editImage() and createVariation() methods available for DALL-E 2
  • Use provider.supportsImageEditing and provider.supportsImageVariations for feature detection

Fixed #

• Package Configuration: Updated pubspec.yaml repository URLs
  • Fixed homepage, repository, and issue_tracker URLs to point to correct standalone repository
  • Ensures proper package metadata on pub.flutter-io.cn for the independent library

Changed #

• Documentation: Updated README.md to reflect standalone library status
  • Added pub.flutter-io.cn package badge and link (https://pub.flutter-io.cn/packages/llm_dart)
  • Simplified installation instructions (removed unnecessary dio dependency mention)

Added #

• Initial release of LLM Dart library
• Multi-provider support for AI interactions
• Unified interface for OpenAI, Anthropic, Google, DeepSeek, Ollama, xAI, Phind, Groq, and ElevenLabs
• Builder pattern API for easy configuration
• Streaming support for real-time responses
• Tool calling capabilities for function execution
• Structured output with JSON schema support
• Comprehensive error handling with specific error types
• Provider registry system for extensibility
• Capability-based design for type safety

Supported Providers

• OpenAI: GPT models with reasoning support
• Anthropic: Claude models with thinking capabilities
• Google: Gemini models
• DeepSeek: DeepSeek reasoning models
• Ollama: Local model support
• xAI: Grok models
• Phind: Phind models
• Groq: Fast inference
• ElevenLabs: Text-to-Speech and Speech-to-Text

Features

• Chat completion with multiple providers
• Real-time streaming responses
• Function/tool calling
• Structured JSON output
• Provider-specific extensions
• Custom provider registration
• Comprehensive examples and documentation

Examples

• Basic usage examples for all providers
• Streaming examples
• Tool calling examples
• Custom provider implementation
• Advanced configuration examples
• ElevenLabs TTS/STT examples

Technical Details #

• Built with Dart 3.0+ support
• Flutter 3.8+ compatibility
• Uses Dio for HTTP requests
• Comprehensive error handling
• Type-safe interfaces
• Modular architecture

Documentation #

• Complete API documentation
• Extensive examples directory
• Setup and configuration guides
• Best practices documentation
• Provider-specific guides