OpenAI GPT OSS

1
import datetime
2
from gpt_oss.tools.simple_browser import SimpleBrowserTool
3
from gpt_oss.tools.simple_browser.backend import YouComBackend
4
from openai_harmony import (
5
    SystemContent,
6
    Message,
7
    Conversation,
8
    Role,
9
    load_harmony_encoding,
10
    HarmonyEncodingName,
11
)
12
13
encoding = load_harmony_encoding(HarmonyEncodingName.HARMONY_GPT_OSS)
14
15
backend = YouComBackend(source="web")
16
browser_tool = SimpleBrowserTool(backend=backend)
17
18
system_content = (
19
    SystemContent.new()
20
    .with_conversation_start_date(datetime.datetime.now().strftime("%Y-%m-%d"))
21
    .with_tools(browser_tool.tool_config)
22
)
23
24
system_message = Message.from_role_and_content(Role.SYSTEM, system_content)
25
user_message = Message.from_role_and_content(
26
    Role.USER, "What's the latest news on AI regulation?"
27
)
28
29
conversation = Conversation.from_messages([system_message, user_message])
30
token_ids = encoding.render_conversation_for_completion(conversation, Role.ASSISTANT)
31
32
# Run inference with your chosen backend (triton, vLLM, etc.)
33
output_tokens = your_inference_engine(token_ids)
34
35
# Parse the model's response
36
messages = encoding.parse_messages_from_completion_tokens(output_tokens, Role.ASSISTANT)
37
last_message = messages[-1]
38
39
# If the model invokes the browser tool, process the call and re-run inference
40
if last_message.recipient.startswith("browser"):
41
    response_messages = await browser_tool.process(last_message)
42
    messages.extend(response_messages)