Installed Chat Eval Commands¶
These commands use datasets packaged with GeneralManager. Replace myproject.settings and provider paths with importable modules from the Django project where the command runs.
Smoke-test the installed wheel¶
python -m general_manager.chat.evals \
--settings myproject.settings \
--provider general_manager.chat.providers.OllamaProvider \
--fixture toy \
--dataset basic_queries \
--tier 0
This is the smallest directly usable installed-package check: it configures Django, registers the toy managers expected by basic_queries, loads the YAML dataset from package data, constructs the provider, and returns status 1 if any case fails.
Use project data and the configured provider¶
DJANGO_SETTINGS_MODULE=myproject.settings \
python -m general_manager.chat.evals \
--dataset follow_ups \
--tag grounding \
--verbose \
--trace-jsonl /tmp/follow-ups.jsonl
No fixture is registered in this version, so the dataset must match managers and data exposed by myproject.settings. Because --provider is absent, the command imports the provider from GENERAL_MANAGER["CHAT"].
Compare two provider classes¶
python -m general_manager.chat.evals \
--settings myproject.settings \
--fixture toy \
--dataset basic_queries \
--compare general_manager.chat.providers.OllamaProvider,myproject.chat.CandidateProvider
The comparison succeeds only if every result from both providers passes. See the task guide for setup and the CLI reference for every option and exit status.