Skip to content

Latest commit

 

History

History
181 lines (141 loc) · 6.87 KB

python.md

File metadata and controls

181 lines (141 loc) · 6.87 KB

Python API Reference

The openai-harmony package exposes the Harmony renderer to Python via thin bindings generated with PyO3. It installs a module named openai_harmony which re‑exports a set of dataclasses mirroring the structures from the Rust crate together with helper classes for encoding and parsing.

Installation

Install the package from PyPI:

pip install openai-harmony

Typical imports look like:

from openai_harmony import Message, Conversation, load_harmony_encoding

Enumerations

Role

Represents the author of a message. Possible values are USER, ASSISTANT, SYSTEM, DEVELOPER and TOOL.

ReasoningEffort

Defines how much reasoning the assistant should apply. Values: LOW, MEDIUM, HIGH.

StreamState

State of an incremental parser: EXPECT_START, HEADER or CONTENT.

Dataclasses

Author

Author(role: Role, name: Optional[str] = None)

Helper for specifying the author of a message. Use Author.new(role, name) to create a named author.

TextContent

TextContent(text: str)

Simple text payload implementing Content.

ToolDescription

ToolDescription(name: str, description: str, parameters: Optional[dict] = None)

Describes an individual callable tool.

ToolNamespaceConfig

ToolNamespaceConfig(name: str, description: Optional[str], tools: List[ToolDescription])

Namespace for grouping tools. Convenience constructors browser() and python() return the built‑in configurations.

ChannelConfig

ChannelConfig(valid_channels: List[str], channel_required: bool)

Configuration for valid message channels. Use ChannelConfig.require_channels(channels) to demand that a specific set of channels is present.

SystemContent

SystemContent(
    model_identity: Optional[str] = "You are ChatGPT, a large language model trained by OpenAI.",
    reasoning_effort: Optional[ReasoningEffort] = ReasoningEffort.MEDIUM,
    conversation_start_date: Optional[str] = None,
    knowledge_cutoff: Optional[str] = "2024-06",
    channel_config: Optional[ChannelConfig] = ChannelConfig.require_channels(["analysis", "commentary", "final"]),
    tools: Optional[dict[str, ToolNamespaceConfig]] = None,
)

Represents a system message. Provides fluent helpers like with_model_identity(), with_reasoning_effort(), with_required_channels(), with_browser_tool() and with_python_tool().

DeveloperContent

DeveloperContent(instructions: Optional[str] = None, tools: Optional[dict[str, ToolNamespaceConfig]] = None)

Content of a developer message. Helper methods include with_instructions(), with_function_tools() and the tool helpers mirroring those from SystemContent.

Message

Message(author: Author, content: List[Content], channel: Optional[str] = None, recipient: Optional[str] = None, content_type: Optional[str] = None)

A single chat message. Convenience constructors from_role_and_content() and from_role_and_contents() mirror the Rust API. Builder methods allow adding content or setting channel/recipient information. to_dict() and from_dict() serialise to/from the canonical JSON format.

Conversation

Conversation(messages: List[Message])

Sequence of messages. Create a conversation using Conversation.from_messages(); serialisation helpers to_json() and from_json() mirror the Rust crate.

RenderConversationConfig

RenderConversationConfig(auto_drop_analysis: bool = True)

Optional configuration when rendering a conversation.

Encoding helpers

HarmonyEncoding

Wrapper around the low level bindings. Obtain an instance with load_harmony_encoding().

Methods:

  • name – name of the encoding.
  • render_conversation_for_completion(conversation, next_turn_role, config=None) – render a conversation into tokens.
  • render_conversation_for_training(conversation, config=None) – render a conversation for training.
  • render_conversation(conversation, config=None) – render a conversation without appending a new role.
  • render(message) – render a single message into tokens.
  • parse_messages_from_completion_tokens(tokens, role=None, strict=True) – parse tokens back into Message objects (set strict=False to enable permissive parsing).
  • decode_utf8(tokens) – decode tokens with the underlying tokenizer.
  • stop_tokens() / stop_tokens_for_assistant_actions() – lists of stop tokens.

Use strict=False when you need the parser to recover from malformed model output that omits markers such as <|message|>.

StreamableParser

Incremental parser built on top of an encoding. Construct with StreamableParser(encoding, role) and feed tokens via process(token). Inspect state via properties like current_content, current_role, tokens and state. Pass strict=False to enable permissive parsing (mirrors ParseOptions { strict: false } on the Rust side).

load_harmony_encoding(name)

Return a HarmonyEncoding by name. Accepts either the string name or a value from the HarmonyEncodingName enum (HARMONY_GPT_OSS).

Exports

The package re‑exports the above classes through __all__ so they are available via:

from openai_harmony import *

Usage Examples

Create a conversation with a system and user message, render it to tokens and parse them back again:

from openai_harmony import (
    Role,
    Message,
    Conversation,
    SystemContent,
    load_harmony_encoding,
    HarmonyEncodingName,
)

# Build messages
system = Message.from_role_and_content(Role.SYSTEM, SystemContent.new())
user = Message.from_role_and_content(Role.USER, "What is 2 + 2?")

# Assemble a conversation
convo = Conversation.from_messages([system, user])

# Render to tokens using the OSS encoding
enc = load_harmony_encoding(HarmonyEncodingName.HARMONY_GPT_OSS)
tokens = enc.render_conversation_for_completion(convo, Role.ASSISTANT)
print(tokens)

# Decode and roundtrip
print(enc.decode_utf8(tokens))
parsed = enc.parse_messages_from_completion_tokens(tokens, role=Role.ASSISTANT)
for m in parsed:
    print(m)

Exceptions

The bindings raise plain Python exceptions. The most common ones are:

  • RuntimeError – returned for rendering or parsing failures (for example if a token sequence is malformed or decoding fails).
  • ValueError – raised when an argument is invalid, e.g. an unknown Role is provided to load_harmony_encoding or StreamableParser.
  • ModuleNotFoundError – accessing the package without building the compiled extension results in this error.

In typical code you would wrap encoding operations in a try/except block:

try:
    tokens = enc.render_conversation_for_completion(convo, Role.ASSISTANT)
    parsed = enc.parse_messages_from_completion_tokens(tokens, Role.ASSISTANT)
except RuntimeError as err:
    print(f"Failed to handle conversation: {err}")