Change to local Ollama but still require API_KEY

[wkholy]

I changed the setting from openai to ollama, and run the ollama serve. I also set an environmental variable of OPENAI_API_KEY='12434', for the checking part. When I run the command following the instruction, it showed an error below.

raise self._make_status_error_from_response(err.response) from None
openai.AuthenticationError: Error code: 401 - {'error': {'message': 'Incorrect API key provided: 12434. You can find your API key at https://platform.openai.com/account/api-keys.', 'type': 'invalid_request_error', 'param': None, 'code': 'invalid_api_key'}}
Error running command: Command '['python', 'database/script/faiss_command_help.py', '--database_path=./database']' returned non-zero exit status 1.

I would like to ask if there will still be a need for the API key for other purpose apart from the deep learning. I will appreciate if there is a tutorial for nozzle profile optimization or simulation with multiphase settings.

[wkholy]

As an update, I change the embedding method using HuggingFaceEmbeddings.

It gets a new error: ollama._types.ResponseError: registry.ollama.ai/library/deepseek-r1:14b does not support tools (status code: 400)
During task with name 'architect' and id 'b9c11b36-0194-cf59-7276-07f343790729'
Error running command: Command '['python', 'src/main.py', '--prompt_path=./user_requirement.txt', '--output_dir=./output']' returned non-zero exit status 1.

[LeoYML]

In the current version of Foam-Agent, we need OpenAI key to generate the embedding for the RAG part, but I agree your suggestion, this is not necessary, because we can upload the pre-calculated embedding. In the following days, we will fix this issues.

[LeoYML]

For your second question, I recommend you to change model to which supports “tool”, this will help us to get the structured output. We will also try to support the models which didn’t support the tools, if you are interested, welcome to join the development and raise an merge request.

[wkholy]

Dear Leo,
Thanks for the quick response. I have changed the pretrained embedding and qwen LLM for the "tool" support. It runs ok at the first round.
However, it shows a new error,
Workflow failed with error: 'NoneType' object has no attribute 'subtasks'
Traceback (most recent call last):
File "/home/wkholy/Foam-Agent/src/main.py", line 161, in
main(user_requirement, config, args.custom_mesh_path)
File "/home/wkholy/Foam-Agent/src/main.py", line 113, in main
result = app.invoke(initial_state)
^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/wkholy/anaconda3/envs/openfoamAgent/lib/python3.12/site-packages/langgraph/pregel/init.py", line 2069, in invoke
for chunk in self.stream(
^^^^^^^^^^^^
File "/home/wkholy/anaconda3/envs/openfoamAgent/lib/python3.12/site-packages/langgraph/pregel/init.py", line 1724, in stream
for _ in runner.tick(
^^^^^^^^^^^^
File "/home/wkholy/anaconda3/envs/openfoamAgent/lib/python3.12/site-packages/langgraph/pregel/runner.py", line 230, in tick
run_with_retry(
File "/home/wkholy/anaconda3/envs/openfoamAgent/lib/python3.12/site-packages/langgraph/pregel/retry.py", line 40, in run_with_retry
return task.proc.invoke(task.input, config)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/wkholy/anaconda3/envs/openfoamAgent/lib/python3.12/site-packages/langgraph/utils/runnable.py", line 506, in invoke
input = step.invoke(input, config, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/wkholy/anaconda3/envs/openfoamAgent/lib/python3.12/site-packages/langgraph/utils/runnable.py", line 270, in invoke
ret = context.run(self.func, *args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/wkholy/Foam-Agent/src/nodes/architect_node.py", line 82, in architect_node
subtasks = decompose_to_subtasks(user_requirement, dir_structure, dir_counts_str, state["llm_service"])
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/wkholy/Foam-Agent/src/services/architect.py", line 102, in decompose_to_subtasks
return [{"file_name": s.file_name, "folder_name": s.folder_name} for s in res.subtasks]
^^^^^^^^^^^^
AttributeError: 'NoneType' object has no attribute 'subtasks'
During task with name 'architect' and id 'ddd9b082-92c8-c102-652a-47da0b85309d'
Error running command: Command '['python', 'src/main.py', '--prompt_path=./user_requirement.txt', '--output_dir=./output']' returned non-zero exit status 1.

The qwen model is a 8b model with my 16GB 5080 laptop. I wonder the error is caused by the insufficient memory of the graphic card. I will further test it with my workstation with 48GB 3090*2. If it goes well, I will raise a merge request for the change of the embedding method. Cheers

[LeoYML]

I am not sure, but I guess it is due to the model being too small, so there are some weird issues. My recommendation is to start with a better model. When your case runs successfully, then you can switch to a smaller model to see what happens. But this case is good to show that our framework is not robust for some smaller models. Another suggestion is that you can have a look at the paper "CFDLLMBench"; you will find that the small model is very hard for this task, which is the reason why I suggest you start with a better, larger model. If your laptop doesn't have enough memory, you can try to use the API.

[LeoYML]

"CFDLLMBench: A Benchmark Suite for Evaluating Large Language Models in Computational Fluid Dynamics", https://arxiv.org/abs/2509.20374