openai/harmony

Features

• Conversation Tokenizers - Converts structured conversations and tool calls into the specific token sequences required for LLM inference and training.
• Tool Definitions - Provides systems for defining the schema and logic of functions that AI models can invoke.
• LLM Tooling Integrations - Defines callable functions and namespaces to manage how a model interacts with external tools and APIs.
• Model Output Transformers - Transforms raw completion tokens into structured message objects using recovery logic to handle malformed text output.
• LLM Token Stream Parsers - Extracts roles and content incrementally from a live token stream as the model produces them.
• Response Token Formatters - Structures model outputs and reasoning chains into consistent token sequences for open-weight model series.
• Tool Calling Integration Frameworks - Organizes tool namespaces and outputs to manage how a model interacts with external functions and hierarchies.
• Tool-Calling Schemas - Provides a framework for defining callable functions and namespaces to organize how a model interacts with external operations.
• Model Response Layouts - Structures conversations and reasoning outputs into consistent layouts to ensure models operate without losing critical tokens.
• LLM Response Parsers - Transforms raw completion tokens and live streams back into structured message objects with recovery logic for malformed text.
• Reasoning Layout Formatting - Structures reasoning chains and multi-channel outputs into consistent layouts to ensure models operate without losing critical tokens.
• Token-Message Converters - Transforms raw completion tokens into structured objects using recovery logic to handle malformed model output.
• Response Stream Parsing - Provides real-time parsing of live token streams to extract roles and content as model responses are generated.
• Token Sequence Formatters - Converts structured conversation objects into specific token sequences required for model inference and training.
• Reasoning Configuration - Sets system constraints like identity and reasoning effort to control how a model communicates and reasons.
• Conversation State Serializers - Organizes messages by author and role into formal sequences for consistent model interaction.
• Namespaced Tool Hierarchies - Organizes callable functions into hierarchies to specify the exact operations a model can execute.
• Malformed Output Recovery - Parses raw completion tokens into structured objects using logic to handle and fix malformed text output.
• Model Encoding Configurations - Fetches predefined encoding sets by name to manage how tokens are processed for different model series.
• Model Encoding Management - Fetches predefined encoding configurations by name to manage how tokens are processed and rendered for specific model series.
• Multi-Channel Stream Renderers - Generates reasoning chains, tool preambles, and standard responses simultaneously within a single output stream.
• Conversational Data Structures - Organizes messages between users and assistants into formal data structures to facilitate consistent interaction with a model.
• Token-Level Incremental Processing - Processes incoming tokens one by one to identify roles and content in real-time from a live stream.
• Artificial Intelligence - Renderer for the harmony response format.

Istoric stele