Support `o3-deep-research` & `o4-mini-deep-research`

[nathan-gage]

Description

:) !!!!!!

References

https://platform.openai.com/docs/guides/deep-research

[DouweM]

To use deep research, use the Responses API with the model set to o3-deep-research or o4-mini-deep-research. You must include at least one data source: web search and/or remote MCP servers.

I think this will already work with the example in https://ai.pydantic.dev/models/openai/#openai-responses-api if you just swap out gpt-4o with the new model names! Can you give that a try?

More complete support would require:

• Background mode
• Parsing more of the outputs (or at least putting them on vendor_details)

[nathan-gage]

It does; but the MCP'd deep research is incompatible -- the types for OpenAIResponsesModel do not allow the private datasource MCPs.

Also, streaming is not supported: (i assume this is what you meant by outputs)

UserWarning: Handling of this event type is not yet implemented. Please report on our GitHub: ResponseWebSearchCallSearchingEvent(item_id='<redacted>', output_index=83, sequence_number=293, type='response.web_search_call.searching')

UserWarning: Handling of this event type is not yet implemented. Please report on our GitHub: ResponseWebSearchCallCompletedEvent(item_id='<redacted>', output_index=83, sequence_number=294, type='response.web_search_call.completed')

UserWarning: Handling of this item type is not yet implemented. Please report on our GitHub: ResponseOutputItemAddedEvent(item=ResponseFunctionWebSearch(id='<redated>', status='in_progress', type='web_search_call'), output_index=85, sequence_number=298, type='response.output_item.added')

---

I supposed what I ultimately meant was the background mode

[vikigenius]

@DouweM how are the MCP servers supposed to be included/declared with the responses API? It seems like OpenaI expects them in a very specific way?

    { # ADD MCP TOOL SUPPORT
      "type": "mcp",
      "server_label": "internal_file_lookup",
      "server_url": "http://0.0.0.0:8000/sse/", # Update to the location of *your* MCP server
      "require_approval": "never"
    }

But this does not seem to be the way pydantic seems to be handling mcp servers. I don't see a type: mcp anywhere

[DouweM]

@vikigenius Can you try passing that dict in openai_builtin_tools=[...], as in the second example here: https://ai.pydantic.dev/models/openai/#openai-responses-api? It'll complain about typing, but the dict should still be sent along to the responses API.

Officially supporting that would be part of #1722, which @Kludex is working on -- although he's out this week.

[vikigenius]

@DouweM Yep, that does seem to work (at least tries to hit the MCP server), I still have to setup the Auth.

The only annoyances now seem to be the background mode thing. I have to set the http timeout to something like 20 mins and pass a custom http client to be safe.

[DouweM]

@vikigenius Yeah background mode would be nice to support.

The OpenAI docs have this example:

from openai import OpenAI
from time import sleep

client = OpenAI()

resp = client.responses.create(
  model="o3",
  input="Write a very long novel about otters in space.",
  background=True,
)

while resp.status in {"queued", "in_progress"}:
  print(f"Current status: {resp.status}")
  sleep(2)
  resp = client.responses.retrieve(resp.id)

print(f"Final status: {resp.status}\nOutput:\n{resp.output_text}")

We could implement that inside the OpenAIModel. That means your agent.run would still be sitting there running for a long time, but at least it wouldn't use the same HTTP connection that could time out. Would that be acceptable?

[vikigenius]

Yeah that does look acceptable for now. But I wonder if in the future it might be possible to treat long running Agents the same way.

Maybe something like resp = agent.start which just works in the background and you just keep doing other tasks, have similar API like resp.status etc.

In this case agent.run would indicate just waiting on the agent to finish but agent.start would be equivalent to backround mode

[DouweM]

@vikigenius I agree that would be a nice feature. Would you mind filing a new issue specifically on long-running agents with some more thoughts on API and behavior?

[DouweM]

@vikigenius Regarding the naive check-the-result-every-2-seconds solution, would you be up for submitting a PR? I think all the work would be in OpenAIResponsesModel._responses_create.