WebModel¶

WebModel is the core class of pydantic-ai-web-models. It implements the pydantic-ai Model interface and routes every inference request through a Temporal workflow to a web-based LLM.

Constructor¶

class WebModel:
    def __init__(
        self,
        provider: str,
        model_name: str,
        *,
        temporal_config: TemporalConfig | None = None,
    ) -> None: ...

Parameters¶

Raises¶

• ValueError — if provider is not a recognised provider string.
• ValueError — if model_name is not a valid value for the provider's enum (GptModelName or GeminiModelName).

Properties¶

Usage¶

from pydantic_ai import Agent
from pydantic_ai_web_models import WebModel, TemporalConfig

# Construct a WebModel explicitly instead of using a model string
model = WebModel(
    provider="google-web",
    model_name="gemini-3.1-pro",
    temporal_config=TemporalConfig(
        task_queue="gpu-workers",
        timeout_seconds=900,
    ),
)

agent = Agent(model=model)
result = agent.run_sync("Summarise the history of the internet.")
print(result.data)

Temporal Client Lifecycle¶

The Temporal client is created lazily on the first request and is cached for the lifetime of the WebModel instance. Subsequent requests from the same WebModel reuse the same client without reconnecting.

Per-run model_settings extensions¶

Pydantic AI merges model_settings from the agent constructor and each run() / run_sync() call and passes the result to WebModel.request(). Besides standard keys (temperature, max_tokens, …), this package recognises:

When the workflow completes successfully with a result like your worker’s LLMInvokeResult (response, thread_id, error empty), WebModel sets ModelResponse.metadata to {"thread_id": "<value>"} on the assistant message. Read it with result.response.metadata["thread_id"] (AgentRunResult.response). If error is non-empty, WebModel raises WorkflowExecutionError before any assistant response is produced. The same ModelResponse is also listed in result.new_messages() / all_messages() when you need the full step.

See Conversations: Server-side thread and Architecture: workflow I/O.

GptModelName¶

class GptModelName(StrEnum):
    GPT_5_5_INSTANT = "gpt-5-5-instant"
    GPT_5_5_THINKING = "gpt-5-5-thinking"

Enum of supported model names for the openai-web provider. Use the string values directly in Agent(model="openai-web:<value>") or pass them as model_name when constructing WebModel.

GeminiModelName¶

class GeminiModelName(StrEnum):
    GEMINI_3_5_FLASH = "gemini-3-5-flash"
    GEMINI_3_5_THINKING = "gemini-3-5-thinking"
    GEMINI_3_1_PRO = "gemini-3.1-pro"

Enum of supported model names for the google-web provider.