{
  "pipeline_name": "Qwen 3 + Qwen 2.5 Legal Discovery Pipeline",
  "version": "1.0",
  "architecture": "Object-Oriented",
  "configuration": {
    "primary_model": "Qwen 3 235B Instruct",
    "secondary_model": "Qwen 2.5 72B Instruct",
    "total_gpus": "6 \u00d7 A100 80GB",
    "cost_per_hour": "$3.84",
    "total_cost": "$515-968 (including attorney)"
  },
  "pipeline_steps": [
    {
      "step": 1,
      "name": "Load Data",
      "script": "pipeline/steps/step1_load_data.py",
      "class": "DataLoader",
      "description": "Load and preprocess Signal CSV messages"
    },
    {
      "step": 2,
      "name": "Create Chunks",
      "script": "pipeline/steps/step2_create_chunks.py",
      "class": "ChunkCreator",
      "description": "Create overlapping 20-message chunks"
    },
    {
      "step": 3,
      "name": "Keyword Filter",
      "script": "pipeline/steps/step3_keyword_filter.py",
      "class": "KeywordFilter",
      "description": "Filter by case-specific keywords"
    },
    {
      "step": 4,
      "name": "Semantic Filter",
      "script": "pipeline/steps/step4_semantic_filter.py",
      "class": "SemanticFilter",
      "description": "Dual-model semantic filtering"
    },
    {
      "step": 5,
      "name": "Random Sampling",
      "script": "pipeline/steps/step5_random_sampling.py",
      "class": "RandomSampler",
      "description": "Stratified random sampling for attorney"
    },
    {
      "step": 6,
      "name": "Labeling Template",
      "script": "pipeline/steps/step6_labeling_template.py",
      "class": "LabelingTemplateGenerator",
      "description": "Generate attorney labeling template"
    },
    {
      "step": 7,
      "name": "Inference Prep",
      "script": "pipeline/steps/step7_inference_prep.py",
      "class": "InferencePreparation",
      "description": "Prepare dual Qwen inference requests"
    },
    {
      "step": 8,
      "name": "Merge Results",
      "script": "pipeline/steps/step8_merge_results.py",
      "class": "ResultsMerger",
      "description": "Merge dual-model results with confidence"
    }
  ],
  "utilities": [
    {
      "name": "Text Utils",
      "script": "pipeline/utils/text_utils.py",
      "functions": [
        "normalize_text",
        "extract_keywords",
        "calculate_keyword_score"
      ]
    },
    {
      "name": "Deployment Helper",
      "script": "pipeline/utils/deployment_helper.py",
      "class": "ModelDeployer",
      "description": "Helper for deploying Qwen models on Vast.ai"
    },
    {
      "name": "Inference Runner",
      "script": "pipeline/utils/inference_runner.py",
      "class": "InferenceRunner",
      "description": "Run inference on dual Qwen models"
    }
  ],
  "core_modules": [
    {
      "name": "Common Definitions",
      "script": "pipeline/common_defs.py",
      "contains": [
        "Case criteria",
        "Model configs",
        "Data classes",
        "Constants"
      ]
    },
    {
      "name": "Base Classes",
      "script": "pipeline/models/base.py",
      "contains": [
        "PipelineStep abstract class",
        "Logging setup",
        "File I/O"
      ]
    },
    {
      "name": "Main Pipeline",
      "script": "pipeline/main_pipeline.py",
      "class": "DiscoveryPipeline",
      "description": "Main orchestrator for running all steps"
    }
  ],
  "expected_performance": {
    "recall": "88-97%",
    "precision": "65-85%",
    "high_confidence_cases": "60-70%",
    "medium_confidence_cases": "25-35%",
    "low_confidence_cases": "5-10%"
  },
  "file_structure": {
    "total_files": 17,
    "total_size_kb": 59.4,
    "python_modules": 14,
    "documentation": 1,
    "config_files": 1,
    "scripts": 1
  }
}