Deep Research ignores configured engine; hardcodes Ollama

[chip902]

Description

Deep Research is hard coded to use Ollama. If your main inference engine is your own trained model being served from LM Studio, vLLM or LlamaServer, Research will still try to be done using Ollama

Steps to Reproduce

1. Run a Deep Research prompt without Ollama gemma4:31b running locally
2. Observe the 404 failure response

Expected Behavior

I would expect the model that I'm already using for Chat to also be used for Deep Resarch as selected in the upper left

Actual Behavior

Fails with a 404 stating the model cannot be found.

Operating System

macOS

Python Version

3.12

Hardware

Apple Silicon

Engine

Other

Logs / Traceback

Research failed: RuntimeError: Ollama returned 404: {"error":"model 'gemma4:31b' not found"}

[gilbert-barajas]

Took a look at this. The hardcode is in server/research_router.py:

• _stream_research builds engine = OllamaEngine() directly (~line 323), so the configured engine is never consulted.
• The planner runs on DEFAULT_PLANNER_MODEL = "gemma4:31b" regardless of req.model (the route intentionally ignores it, per the ResearchRequest comment).

Pointing the engine at app.state.engine like the chat routes is the easy part. But gemma4:31b is an Ollama-specific tag, so it'll still 404 on LM Studio / vLLM / llama-server. The engine swap alone doesn't fix it.

The separate planner model looks deliberate (the comment says the chat selector shouldn't override the planner's function-calling requirement). So before I open a PR, what should the planner resolve to on a non-Ollama engine?

1. the configured engine's default model
2. a separate configurable planner-model setting
3. something else

The engine fix is small. Happy to PR it once the intended resolution is clear.

[ElliotSlusky]

Should be fixed now! Thanks!