File size: 11,216 Bytes
28f1212 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 | # [Track C: Iterative Refinement]
"""
Track C — Run iterative refinement experiments against MedQA.
Usage:
cd src/backend
python -m tracks.iterative.run_iterative # all configs
python -m tracks.iterative.run_iterative --config C1_3rounds # single config
python -m tracks.iterative.run_iterative --max-cases 10 # quick test
Each config runs the full baseline pipeline first, then feeds the initial
reasoning through N self-critique iterations. Results include per-iteration
accuracy AND cost, enabling cost/benefit charts.
"""
from __future__ import annotations
import asyncio
import json
import logging
import sys
import time
from datetime import datetime, timezone
from pathlib import Path
from typing import List, Optional
BACKEND_DIR = Path(__file__).resolve().parent.parent.parent
if str(BACKEND_DIR) not in sys.path:
sys.path.insert(0, str(BACKEND_DIR))
from app.agent.orchestrator import Orchestrator
from app.models.schemas import (
CaseSubmission,
CDSReport,
AgentStepStatus,
ClinicalReasoningResult,
)
from tracks.iterative.config import CONFIGS, IterativeConfig
from tracks.iterative.refiner import IterativeRefiner
from tracks.shared.cost_tracker import CostLedger
from validation.base import (
ValidationCase,
ValidationResult,
ValidationSummary,
diagnosis_in_differential,
run_cds_pipeline,
save_results,
print_summary,
)
logger = logging.getLogger(__name__)
RESULTS_DIR = Path(__file__).resolve().parent / "results"
MEDQA_PATH = BACKEND_DIR / "validation" / "data" / "medqa_test.jsonl"
# ──────────────────────────────────────────────
# Per-case runner
# ──────────────────────────────────────────────
async def run_case_iterative(
case: ValidationCase,
config: IterativeConfig,
ledger: CostLedger,
) -> ValidationResult:
"""
Run one case through:
1. The baseline pipeline (Track A) to get initial reasoning
2. The iterative refinement loop (Track C)
3. Re-synthesize the final report with the refined differential
"""
t0 = time.monotonic()
# ── Step 1: Run baseline pipeline ──
state, report, error = await run_cds_pipeline(
patient_text=case.input_text,
include_drug_check=True,
include_guidelines=True,
)
if error or not state or not state.clinical_reasoning or not state.patient_profile:
return ValidationResult(
case_id=case.case_id,
source_dataset=f"trackC_{config.config_id}",
success=False,
scores={},
pipeline_time_ms=int((time.monotonic() - t0) * 1000),
error=error or "Baseline pipeline failed to produce reasoning",
)
# ── Step 2: Iterative refinement ──
refiner = IterativeRefiner(config, ledger)
refined_reasoning, history = await refiner.refine(
profile=state.patient_profile,
initial_reasoning=state.clinical_reasoning,
)
# ── Step 3: Re-synthesize with the refined differential ──
# Inject the refined reasoning back into the orchestrator state and
# re-run just the synthesis step
from app.tools.synthesis import SynthesisTool
synth = SynthesisTool()
try:
refined_report = await synth.run(
patient_profile=state.patient_profile,
clinical_reasoning=refined_reasoning,
drug_interactions=state.drug_interactions,
guideline_retrieval=state.guideline_retrieval,
conflict_detection=state.conflict_detection,
)
except Exception as e:
refined_report = report # Fall back to baseline report
logger.warning(f"Re-synthesis failed, using baseline report: {e}")
elapsed_ms = int((time.monotonic() - t0) * 1000)
# ── Score — compare baseline vs. refined ──
scores: dict = {}
details: dict = {"iterations": len(history) - 1} # subtract the initial
if "answer" in case.ground_truth:
gt = case.ground_truth["answer"]
# Score the baseline
if report:
b_found, b_rank, b_loc = diagnosis_in_differential(gt, report)
scores["baseline_top1"] = 1.0 if (b_found and b_rank == 0) else 0.0
scores["baseline_mentioned"] = 1.0 if b_found else 0.0
# Score the refined report
target_report = refined_report or report
if target_report:
r_found, r_rank, r_loc = diagnosis_in_differential(gt, target_report)
scores["top1_accuracy"] = 1.0 if (r_found and r_rank == 0) else 0.0
scores["top3_accuracy"] = 1.0 if (r_found and r_rank < 3) else 0.0
scores["mentioned"] = 1.0 if r_found else 0.0
details["rank"] = r_rank
details["match_location"] = r_loc
details["improved"] = scores.get("top1_accuracy", 0) > scores.get("baseline_top1", 0)
# Per-iteration differential snapshots (for cost/benefit charts)
details["per_iteration_top_dx"] = [
h.differential_diagnosis[0].diagnosis if h.differential_diagnosis else "?"
for h in history
]
details["cost_ledger"] = ledger.to_dict()
return ValidationResult(
case_id=case.case_id,
source_dataset=f"trackC_{config.config_id}",
success=True,
scores=scores,
pipeline_time_ms=elapsed_ms,
report_summary=(refined_report or report).patient_summary[:200] if (refined_report or report) else None,
details=details,
)
# ──────────────────────────────────────────────
# Experiment runner
# ──────────────────────────────────────────────
async def run_config(
config: IterativeConfig,
cases: List[ValidationCase],
) -> ValidationSummary:
"""Run all cases through the iterative config."""
results: List[ValidationResult] = []
start = time.monotonic()
for i, case in enumerate(cases, 1):
logger.info(f" [{config.config_id}] case {i}/{len(cases)}: {case.case_id}")
ledger = CostLedger(track_id=f"C_{config.config_id}")
vr = await run_case_iterative(case, config, ledger)
results.append(vr)
elapsed = time.monotonic() - start
successful = [r for r in results if r.success]
metrics = {}
for key in ("top1_accuracy", "top3_accuracy", "mentioned", "baseline_top1", "baseline_mentioned"):
vals = [r.scores[key] for r in successful if key in r.scores]
metrics[key] = sum(vals) / len(vals) if vals else 0.0
metrics["pipeline_success"] = len(successful) / len(results) if results else 0.0
# Average iterations used
iters = [r.details.get("iterations", 0) for r in successful]
metrics["avg_iterations"] = sum(iters) / len(iters) if iters else 0.0
# Improvement rate
improved = [r for r in successful if r.details.get("improved")]
metrics["improvement_rate"] = len(improved) / len(successful) if successful else 0.0
return ValidationSummary(
dataset=f"trackC_{config.config_id}",
total_cases=len(results),
successful_cases=len(successful),
failed_cases=len(results) - len(successful),
metrics=metrics,
per_case=results,
run_duration_sec=round(elapsed, 1),
)
# ──────────────────────────────────────────────
# Data loading (reuse from validation)
# ──────────────────────────────────────────────
def load_medqa_cases(max_cases: Optional[int] = None) -> List[ValidationCase]:
if not MEDQA_PATH.exists():
logger.error(f"MedQA data not found at {MEDQA_PATH}")
return []
cases: List[ValidationCase] = []
with open(MEDQA_PATH, "r", encoding="utf-8") as f:
for ln, line in enumerate(f, 1):
if max_cases and len(cases) >= max_cases:
break
if not line.strip():
continue
data = json.loads(line)
cases.append(ValidationCase(
case_id=data.get("id", f"medqa_{ln}"),
source_dataset="medqa",
input_text=data.get("question", data.get("input", "")),
ground_truth={"answer": data.get("answer", data.get("target", ""))},
metadata=data.get("metadata", {}),
))
return cases
# ──────────────────────────────────────────────
# CLI
# ──────────────────────────────────────────────
async def main():
import argparse
parser = argparse.ArgumentParser(description="Track C: Iterative refinement experiments")
parser.add_argument("--config", type=str, default=None, help="Run a single config by ID")
parser.add_argument("--max-cases", type=int, default=None, help="Limit cases per config")
parser.add_argument("--quiet", action="store_true")
args = parser.parse_args()
logging.basicConfig(
level=logging.WARNING if args.quiet else logging.INFO,
format="%(asctime)s %(levelname)s %(name)s: %(message)s",
)
# Wait for endpoint to be online (handles scale-to-zero)
from tracks.shared.endpoint_check import wait_for_endpoint
if not await wait_for_endpoint(quiet=args.quiet):
print("ABORT: MedGemma endpoint is not reachable. Resume it and try again.")
sys.exit(1)
configs = CONFIGS
if args.config:
configs = [c for c in CONFIGS if c.config_id == args.config]
if not configs:
print(f"Unknown config: {args.config}")
print(f"Available: {[c.config_id for c in CONFIGS]}")
sys.exit(1)
cases = load_medqa_cases(args.max_cases)
if not cases:
print("No MedQA cases loaded")
sys.exit(1)
print(f"Loaded {len(cases)} MedQA cases\n")
RESULTS_DIR.mkdir(parents=True, exist_ok=True)
for cfg in configs:
print(f"\n{'='*60}")
print(f" Running config: {cfg.config_id}")
print(f" {cfg.description}")
print(f" Max iterations: {cfg.max_iterations}, convergence: {cfg.convergence_threshold}")
print(f"{'='*60}")
summary = await run_config(cfg, cases)
ts = datetime.now(timezone.utc).strftime("%Y%m%d_%H%M%S")
fname = f"trackC_{cfg.config_id}_{ts}.json"
path = RESULTS_DIR / fname
# Use validation save then move
save_path = save_results(summary, filename=fname)
if save_path != path:
import shutil
shutil.move(str(save_path), str(path))
if not args.quiet:
print_summary(summary)
if __name__ == "__main__":
asyncio.run(main())
|