MLFlow eval

Author: EoghanOConnor

skills/rca-annotator/scripts/mlflow_eval.py

@@ -0,0 +1,208 @@
+"""
+MLflow evaluation functions for RCA-Annotator.
+
+Provides functions to download annotation files from jumpbox and log
+evaluation results to MLflow.
+"""
+
+import argparse
+import json
+import os
+import sys
+from pathlib import Path
+from typing import Any
+
+import mlflow  # type: ignore
+from jumpbox_io import download_from_jumpbox
+
+
+def load_annotation(job_id: str) -> dict[str, Any] | None:
+    """
+    Load annotation_draft.json for a given job_id.
+
+    Args:
+        job_id: Job ID to load annotation for
+
+    Returns:
+        Annotation dict if successful, None on failure
+    """
+    annotation_file = Path(".analysis") / job_id / "annotation_draft.json"
+
+    if not annotation_file.exists():
+        print(f"  Error: Annotation file not found: {annotation_file}")
+        return None
+
+    try:
+        with open(annotation_file) as f:
+            annotation = json.load(f)
+        return annotation
+    except json.JSONDecodeError as e:
+        print(f"  Error: Invalid JSON in {annotation_file}: {e}")
+        return None
+    except Exception as e:
+        print(f"  Error: Failed to read {annotation_file}: {e}")
+        return None
+
+
+def download_annotations_for_eval(
+    job_ids: list[str], jumpbox_uri: str | None = None
+) -> dict[str, dict[str, Any]]:
+    """
+    Download and load annotations for multiple jobs.
+
+    Args:
+        job_ids: List of job IDs to download
+        jumpbox_uri: JUMPBOX_URI connection string (defaults to env var)
+
+    Returns:
+        Dict mapping job_id to annotation data
+    """
+    annotations = {}
+
+    print(f"Downloading annotations for {len(job_ids)} jobs...")
+
+    for i, job_id in enumerate(job_ids, 1):
+        print(f"\n[{i}/{len(job_ids)}] Job {job_id}")
+
+        if download_from_jumpbox(job_id, jumpbox_uri):
+            annotation = load_annotation(job_id)
+            if annotation:
+                annotations[job_id] = annotation
+                print("  ✓ Loaded annotation")
+            else:
+                print("  ✗ Failed to load annotation")
+        else:
+            print("  ✗ Download failed")
+
+    print(f"\n{'=' * 60}")
+    print(f"Downloaded {len(annotations)}/{len(job_ids)} annotations")
+    print(f"{'=' * 60}")
+
+    return annotations
+
+
+def log_annotation_feedback(trace_id: str, annotations: dict[str, Any]) -> None:
+    """Log annotation details as MLflow feedback."""
+    root_cause = annotations.get("root_cause", {})
+    if root_cause:
+        mlflow.log_feedback(
+            trace_id=trace_id,
+            name="Root Cause",
+            value=f"Category: {root_cause.get('category')} \n Confidence: {root_cause.get('confidence')}",
+            rationale=root_cause.get("summary"),
+        )
+
+    for evidence_item in annotations.get("evidence", []):
+        mlflow.log_feedback(
+            trace_id=trace_id,
+            name="Evidence",
+            value=f"{evidence_item.get('source')}: {evidence_item.get('message')} \n Confidence {evidence_item.get('confidence')}",
+        )
+
+    for recommendation in annotations.get("recommendations", []):
+        mlflow.log_feedback(
+            trace_id=trace_id,
+            name="Recommendation",
+            value=f"Priority: {recommendation.get('priority')} \n Action: {recommendation.get('action')}",
+            rationale=f"File: {recommendation.get('file')}",
+        )
+
+    for alt_diagnosis in annotations.get("alternative_diagnoses", []):
+        mlflow.log_feedback(
+            trace_id=trace_id,
+            name="Alternative Diagnosis",
+            value=f"Category: {alt_diagnosis.get('category')} \n Summary: {alt_diagnosis.get('summary')}",
+            rationale=alt_diagnosis.get("why_wrong"),
+        )
+
+    for factor in annotations.get("contributing_factors", []):
+        mlflow.log_feedback(trace_id=trace_id, name="Contributing Factor", value=factor)
+
+    for key, value in annotations.get("consistency_check", {}).items():
+        mlflow.log_feedback(
+            trace_id=trace_id, name=f"Consistency Check: {key}", value=f"{key}: {value}"
+        )
+
+
+def log_expectation(trace_id: str, annotations: dict[str, Any]) -> None:
+    """Log expectations for the given annotations."""
+    human_review = annotations.get("human_review", {})
+    if human_review:
+        # Category review
+        mlflow.log_expectation(
+            trace_id=trace_id,
+            name="Human Review",
+            value=f"Summary accurate: {human_review.get('summary_accurate')} \n Summary comment: {human_review.get('summary_comment')}",
+        )
+
+        # Summary review
+        mlflow.log_expectation(
+            trace_id=trace_id,
+            name="Summary (Human Review)",
+            value=f"Summary accurate: {human_review.get('summary_accurate')}",
+        )
+
+        # Evidence review
+        mlflow.log_expectation(
+            trace_id=trace_id, name="Evidence ", value=human_review.get("evidence_feedback")
+        )
+
+        # Difficulty review
+        mlflow.log_expectation(
+            trace_id=trace_id,
+            name="Difficulty",
+            value=f"Difficulty appropriate: {human_review.get('difficulty_appropriate')}",
+        )
+
+        # Alternative diagnoses added by human reviewer
+        for alt_diagnosis in human_review.get("alternative_diagnoses_added", []):
+            mlflow.log_expectation(
+                trace_id=trace_id,
+                name="Alternative Diagnosis Added",
+                value=f"Category: {alt_diagnosis.get('category')} | Plausibility: {alt_diagnosis.get('plausibility')} \nSummary: {alt_diagnosis.get('summary')}",
+            )
+
+
+def evaluate_jobs(job_ids: list[str]) -> None:
+    """Run evaluation for the given job IDs."""
+    tracking_uri = os.environ.get("MLFLOW_TRACKING_URI", "http://localhost:5000")
+    mlflow.set_tracking_uri(tracking_uri)
+    experiment_name = os.environ.get("MLFLOW_EXPERIMENT_NAME", "Default")
+    mlflow.set_experiment(experiment_name)
+
+    with mlflow.start_run(run_name="ANNOTATOR_EVALUATION"):
+        # Create a traced span within the run to ensure linkage
+        with mlflow.start_span(name="download_annotations") as span:
+            data = download_annotations_for_eval(job_ids)
+            trace_id = span.request_id
+
+            for job_id in job_ids:
+                if job_id not in data:
+                    print(f"Warning: No data for job {job_id}")
+                    continue
+
+                annotations = data[job_id]
+
+                # Log feedback for annotation quality metrics
+                log_annotation_feedback(trace_id, annotations)
+
+                # Log ground truth annotation as expectation
+                log_expectation(trace_id, annotations)
+
+            # Log run params
+            mlflow.log_param("job_ids", job_ids)
+            if job_ids and job_ids[0] in data:
+                mlflow.log_param("annotator", data[job_ids[0]].get("annotator"))
+
+        print(f"Traces are {trace_id}")
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Evaluate RCA annotations using MLflow.")
+    parser.add_argument("job_ids", nargs="+", help="List of job IDs to evaluate")
+    args = parser.parse_args()
+    evaluate_jobs(args.job_ids)
+
+
+if __name__ == "__main__":
+    sys.exit(main())