compressor_induction_version1203

.gitignore

@@ -217,3 +217,6 @@ workspace/
 # Evaluation outputs
 eval_outputs/
 builds/
+
+# API keys setup script (contains sensitive credentials)
+setup_api_keys.sh

CAWM/debug_stage1.py

@@ -0,0 +1,178 @@
+#!/usr/bin/env python3
+"""
+Debug script for Stage 1: Trajectory Segmentation
+
+This runs ONLY Stage 1 on a single trajectory to help debug.
+"""
+
+import json
+from pathlib import Path
+
+from extractor_openai import WorkflowExtractor
+
+
+def debug_stage1():
+    """Debug Stage 1 segmentation on a single trajectory"""
+
+    print("=" * 70)
+    print("DEBUG: Stage 1 - Trajectory Segmentation")
+    print("=" * 70)
+    print()
+
+    # Configuration
+    base_url = "https://ai-gateway.andrew.cmu.edu/"
+    model = "gpt-5"  # Your key has access to: gpt-5, gpt-5-mini, claude-sonnet-4
+
+    # Load the 8th trajectory from resolved_trajectories.jsonl
+    trajectory_file = Path("./trajectories/resolved_trajectories.jsonl")
+    trajectory_index = 7  # 8th trajectory (0-indexed)
+
+    print(f"Loading trajectory #{trajectory_index + 1} from {trajectory_file}")
+
+    # Read the specific line from JSONL
+    with open(trajectory_file, "r") as f:
+        for i, line in enumerate(f):
+            if i == trajectory_index:
+                trajectory_data = json.loads(line)
+                break
+        else:
+            print(f"❌ Trajectory index {trajectory_index} not found in file!")
+            return 1
+
+    instance_id = trajectory_data["instance_id"]
+    events = trajectory_data.get("history", [])
+
+    print("Configuration:")
+    print(f"  Base URL: {base_url}")
+    print(f"  Model: {model}")
+    print(f"  Instance ID: {instance_id}")
+    print(f"  Events in trajectory: {len(events)}")
+    print()
+
+    # Create extractor
+    print("Creating extractor...")
+    extractor = WorkflowExtractor(
+        model=model, output_dir="./debug_output", base_url=base_url
+    )
+    print("✅ Extractor created")
+    print()
+
+    # Load and inspect trajectory
+    print("=" * 70)
+    print("STEP 1: Load trajectory data")
+    print("=" * 70)
+    print(f"✅ Loaded {len(events)} events")
+    print()
+
+    # Show first 3 events
+    print("First 3 events:")
+    for i, event in enumerate(events[:3]):
+        print(f"\nEvent {i}:")
+        for key in ["timestamp", "role", "action"]:
+            if key in event:
+                value = str(event[key])[:100]
+                print(f"  {key}: {value}")
+    print()
+
+    # Format for LLM
+    print("=" * 70)
+    print("STEP 2: Format trajectory for LLM")
+    print("=" * 70)
+    formatted = extractor.format_trajectory_for_segmentation(events)
+    print(f"✅ Formatted trajectory: {len(formatted)} chars")
+    print()
+    print("First 500 chars of formatted trajectory:")
+    print("-" * 70)
+    print(formatted[:500])
+    print("...")
+    print()
+
+    # Create prompt
+    print("=" * 70)
+    print("STEP 3: Create segmentation prompt")
+    print("=" * 70)
+    prompt = extractor.create_segmentation_prompt(instance_id, formatted, len(events))
+    print(f"✅ Created prompt: {len(prompt)} chars")
+    print()
+    print("First 500 chars of prompt:")
+    print("-" * 70)
+    print(prompt[:500])
+    print("...")
+    print()
+
+    # Call LLM
+    print("=" * 70)
+    print("STEP 4: Call LLM")
+    print("=" * 70)
+    print(f"Calling {model} via {base_url}")
+    print("This may take 30-60 seconds...")
+    print()
+
+    try:
+        raw_response = extractor.call_llm(prompt)
+        print(f"✅ Received response: {len(raw_response)} chars")
+        print()
+        print("Response (first 1000 chars):")
+        print("-" * 70)
+        print(raw_response[:1000])
+        print("...")
+        print()
+
+        # Parse JSON
+        print("=" * 70)
+        print("STEP 5: Parse JSON response")
+        print("=" * 70)
+
+        json_start = raw_response.find("{")
+        json_end = raw_response.rfind("}") + 1
+
+        if json_start == -1 or json_end == 0:
+            print("❌ No JSON found in response!")
+            print()
+            print("Full response:")
+            print(raw_response)
+            return 1
+
+        json_str = raw_response[json_start:json_end]
+        print(f"Extracted JSON: {len(json_str)} chars")
+        print()
+
+        result_data = json.loads(json_str)
+        segments = result_data.get("segments", [])
+
+        print(f"✅ Parsed {len(segments)} segments")
+        print()
+
+        # Show segments
+        print("=" * 70)
+        print("STEP 6: Examine segments")
+        print("=" * 70)
+
+        for seg in segments:
+            print(f"\nSegment {seg['segment_id']}:")
+            print(f"  Events: {seg['start_event']}-{seg['end_event']}")
+            print(f"  Purpose: {seg['purpose']}")
+            print(f"  Goal: {seg['goal'][:80]}...")
+            print(f"  Outcome: {seg['outcome']}")
+            print(f"  Boundary: {seg['boundary_reason']}")
+
+        print()
+        print("=" * 70)
+        print("✅ Stage 1 Debug Complete!")
+        print("=" * 70)
+        print()
+        print(f"Total segments extracted: {len(segments)}")
+        print("Output saved to: ./debug_output/stage1_segments/")
+
+        return 0
+
+    except Exception as e:
+        print(f"❌ Error: {e}")
+        import traceback
+
+        traceback.print_exc()
+        return 1
+
+
+if __name__ == "__main__":
+    exit(debug_stage1())