fix: wire AgentGraph to YAML config

[bhavyakeerthi3]

Closes #108
Closes #148

Description

This PR addresses the long-standing issue where AgentGraph hardcoded the LLM and embedding models, ignoring any settings in config.yml.
First reported in #108, this completes the YAML configurability work by extending it to the most critical parameters.

Changes

• Added llm and embedding fields to Config in src/util/config_yml/__init__.py
• Updated .config.schema.yaml and config_default.yml with safe defaults
• Updated AgentGraph constructor to accept model configs instead of hardcoded strings
• Updated chat-chainlit.py to pass YAML-loaded configs to the agent

Tests Added

def test_config_loads_llm_and_embedding(mock_config_file):
    config = Config.from_yaml(mock_config_file)
    assert config.llm == "gpt-4o-mini"
    assert config.embedding == "text-embedding-3-large"

Relation to Other PRs

Completes configurability work started in #112 by AaryanCode69, extending it to cover LLM and embedding model parameters which were not included in that PR.

[adamjohnwright]

Thank you @bhavyakeerthi3 — and sorry for the long silence on this one.

The gap you identified was real: AgentGraph hardcoded its models and ignored config.yml entirely. It is fixed in #201, which closes #108 and #148 as you intended.

Two contributors independently reported the same problem within weeks of each other (yours and #112), which was strong evidence it was worth fixing properly rather than patching.

Where the implementation differs from yours

I went with #112's nested shape — llm: {provider, model, base_url} — rather than your flat llm: "openai/gpt-4o-mini". Your version is genuinely tidier to read, and if the only thing being configured were the model I would have taken it. The deciding factor was base_url: Plant Reactome serves its model from a self-hosted OpenAI-compatible endpoint, and a "provider/model" string has nowhere to put that.

Your tests/conftest.py turned out to be unnecessary — pyproject.toml already sets pythonpath = ["src", "bin"] under [tool.pytest.ini_options], which does the same job without a file. Worth knowing for next time; it was not obvious.

What I did not take: the embedding setting

The same part I rejected from #112, for the same reason, and it is a real trap rather than a preference.

A query embedded with a different model than built the stored vectors does not error — it returns confident nonsense. So the embedding model is read from the bundle path by resolve_embedding_model(). Your default of "openai/text-embedding-3-large" applied to Plant Reactome, which uses bge-m3, would silently retrieve wrong documents for every question with nothing reporting a fault.

config.yml now deliberately cannot name one: an unknown key stops the server at startup rather than being ignored.

The full reasoning is in specs/003-model-configuration/spec.md. Closing as implemented — thank you for finding it and for the clear write-up linking the two issues.