Skip to content

operational_workflows.md

Latest commit

 

History

History
41 lines (30 loc) · 1.31 KB

operational_workflows.md

File metadata and controls

41 lines (30 loc) · 1.31 KB
Raw

Operational Workflows

AlignGPT now exposes three runnable vertical flows.

1. Alignment Evaluation

python - <<'PY'
import sys
sys.path.insert(0, "src")
from aligngpt.alignment_pipeline import AlignmentEvaluationPipeline
from aligngpt.schemas import AlignmentRequest

pipeline = AlignmentEvaluationPipeline()
result = pipeline.run(AlignmentRequest(prompt="Explain safe reward modeling with retrieval."))
print(result.reward_score)
print(result.route.backend_name)
PY

This path validates the prompt, applies safety rules, retrieves context, selects a backend with the GPU-aware router, generates a deterministic response, scores reward/grounding, emits metrics, and returns trace events.

2. Benchmark + Reproducibility

python benchmarking/reproducibility/run_operational_benchmark.py \
  --output-dir benchmarking/results/2026-05-20

The run writes JSON, CSV, Markdown, and comparison artifacts. The same artifact contract can be uploaded by CI or scheduled jobs to object storage.

3. Deployment + Observability

python -m pip install -e ".[api]"
uvicorn backend.api_gateway.app:app --reload --port 8000
curl http://localhost:8000/ready
curl http://localhost:8000/metrics

The API exposes health, readiness, Prometheus metrics, event streaming, alignment inference, and evaluation endpoints.