Update inference.py

inference.py

@@ -1,32 +1,45 @@
-"""Benchmark-compatible inference runner for clinical trial screening."""
+"""Hackathon-compliant inference runner for the clinical trial environment."""
 
 from __future__ import annotations
 
 import asyncio
 import json
 import os
-import re
-from typing import List, Optional
+import textwrap
+from typing import Dict, List, Optional, Tuple
 
 from openai import OpenAI
 
-from client import ClinicalTrialScreeningEnvClient
-from models import ClinicalTrialScreeningAction, ClinicalTrialScreeningObservation
+try:
+    from clinical_trial_env import ClinicalTrialAction, ClinicalTrialEnv
+except ImportError:
+    from client import ClinicalTrialEnv
+    from models import ClinicalTrialAction
 
+LOCAL_IMAGE_NAME = os.getenv("LOCAL_IMAGE_NAME") or os.getenv("IMAGE_NAME")
+API_KEY = os.getenv("HF_TOKEN") or os.getenv("API_KEY")
 API_BASE_URL = os.getenv("API_BASE_URL", "https://router.huggingface.co/v1")
 MODEL_NAME = os.getenv("MODEL_NAME", "Qwen/Qwen2.5-72B-Instruct")
-HF_TOKEN = os.getenv("HF_TOKEN")
-LOCAL_IMAGE_NAME = os.getenv("LOCAL_IMAGE_NAME")
+TASK_NAME = os.getenv("CLINICAL_TRIAL_TASK", "easy")
+BENCHMARK = os.getenv("CLINICAL_TRIAL_BENCHMARK", "clinical_trial_env")
 ENV_BASE_URL = os.getenv("ENV_BASE_URL")
-BENCHMARK = os.getenv("BENCHMARK", "clinical_trial_screening")
-TASK_NAME = os.getenv("TASK_NAME", "clinical_trial_patient_screening")
-MAX_STEPS = int(os.getenv("MAX_STEPS", "19"))
-
-SYSTEM_PROMPT = (
-    "You are a clinical trial screening agent. "
-    "Return one compact JSON object with keys action_type, target_id, field_name, value, "
-    "ranking, exclusions, rationale. Use only protocol-supported fields and codes."
-)
+MAX_STEPS = int(os.getenv("MAX_STEPS", "20"))
+TEMPERATURE = float(os.getenv("TEMPERATURE", "0.1"))
+MAX_TOKENS = int(os.getenv("MAX_TOKENS", "220"))
+SUCCESS_SCORE_THRESHOLD = float(os.getenv("SUCCESS_SCORE_THRESHOLD", "0.8"))
+MIN_STRICT_SCORE = 0.01
+MAX_STRICT_SCORE = 0.99
+
+SYSTEM_PROMPT = textwrap.dedent(
+    """
+    You are operating a clinical trial screening environment.
+    Return exactly one compact JSON object with keys:
+    action_type, field_name, value, ranking, deviations, final_decision, rationale.
+    Use only supported action_type values:
+    extract_data, rank_patients, flag_deviation, submit_decision.
+    Do not add markdown, commentary, or code fences.
+    """
+).strip()
 
 
 def log_start(task: str, env: str, model: str) -> None:
@@ -34,201 +47,191 @@ def log_start(task: str, env: str, model: str) -> None:
 
 
 def log_step(step: int, action: str, reward: float, done: bool, error: Optional[str]) -> None:
-    error_value = error if error is not None else "null"
+    error_text = error if error is not None else "null"
     print(
         f"[STEP] step={step} action={action} reward={reward:.2f} "
-        f"done={str(done).lower()} error={error_value}",
+        f"done={str(done).lower()} error={error_text}",
         flush=True,
     )
 
 
-def log_end(success: bool, steps: int, score: float, rewards: List[float]) -> None:
+def log_end(success: bool, steps: int, rewards: List[float]) -> None:
     rewards_str = ",".join(f"{reward:.2f}" for reward in rewards)
     print(
-        f"[END] success={str(success).lower()} steps={steps} "
-        f"score={score:.2f} rewards={rewards_str}",
+        f"[END] success={str(success).lower()} steps={steps} rewards={rewards_str}",
         flush=True,
     )
 
 
-def format_action(action: ClinicalTrialScreeningAction) -> str:
-    if action.action_type == "extract_data":
-        return f"extract_data({action.field_name}={action.value})"
-    if action.action_type == "submit_ranking":
-        return f"submit_ranking({'>'.join(action.ranking)})"
-    if action.action_type == "flag_exclusions":
-        return f"flag_exclusions({','.join(action.exclusions)})"
-    if action.action_type == "final_decision":
-        return f"final_decision({action.value})"
-    return action.action_type
-
-
-def safe_error(message: Optional[str]) -> Optional[str]:
-    if not message:
+def sanitize_error(error: Optional[str]) -> Optional[str]:
+    if error is None:
         return None
-    return re.sub(r"\s+", " ", message.strip())
-
-
-def build_user_prompt(
-    observation: ClinicalTrialScreeningObservation,
-    history: List[str],
-) -> str:
-    recent_history = " | ".join(history[-4:]) if history else "none"
-    return (
-        f"task_id={observation.task_id}\n"
-        f"difficulty={observation.difficulty.value}\n"
-        f"title={observation.title}\n"
-        f"brief={observation.brief}\n"
-        f"prompt={observation.prompt}\n"
-        f"missing_targets={observation.missing_targets}\n"
-        f"available_actions={observation.available_actions}\n"
-        f"trial_metadata={json.dumps(observation.trial_metadata, sort_keys=True)}\n"
-        f"history={recent_history}\n"
-        "Reply with JSON only."
-    )
+    cleaned = " ".join(error.split())
+    return cleaned or "null"
+
+
+def build_user_prompt(task_name: str, step: int, observation_payload: Dict, history: List[str]) -> str:
+    history_text = "\n".join(history[-4:]) if history else "None"
+    return textwrap.dedent(
+        f"""
+        Task: {task_name}
+        Step: {step}
+        Observation:
+        {json.dumps(observation_payload, indent=2, sort_keys=True)}
+
+        Recent history:
+        {history_text}
+
+        Return the next best JSON action.
+        """
+    ).strip()
+
+
+def heuristic_action(task_name: str, step: int) -> ClinicalTrialAction:
+    heuristics: Dict[Tuple[str, int], ClinicalTrialAction] = {
+        ("easy", 1): ClinicalTrialAction(action_type="extract_data", field_name="age", value="56"),
+        ("easy", 2): ClinicalTrialAction(
+            action_type="extract_data", field_name="egfr_mutation", value="L858R positive"
+        ),
+        ("easy", 3): ClinicalTrialAction(action_type="submit_decision", final_decision="eligible"),
+        ("medium", 1): ClinicalTrialAction(
+            action_type="extract_data", field_name="BC-101_her2_status", value="IHC 3+"
+        ),
+        ("medium", 2): ClinicalTrialAction(
+            action_type="extract_data", field_name="BC-102_trastuzumab_exposure", value="none"
+        ),
+        ("medium", 3): ClinicalTrialAction(
+            action_type="rank_patients", ranking=["BC-101", "BC-103", "BC-102"]
+        ),
+        ("hard", 1): ClinicalTrialAction(action_type="extract_data", field_name="biomarker", value="FLT3-ITD"),
+        ("hard", 2): ClinicalTrialAction(
+            action_type="flag_deviation",
+            deviations=[
+                "neutropenic fever",
+                "qtc greater than 480 ms",
+                "recent strong CYP3A4 inhibitor",
+            ],
+        ),
+        ("hard", 3): ClinicalTrialAction(action_type="submit_decision", final_decision="ineligible"),
+    }
+    return heuristics.get((task_name, step), ClinicalTrialAction(action_type="submit_decision", final_decision="ineligible"))
 
 
-def fallback_action(observation: ClinicalTrialScreeningObservation) -> ClinicalTrialScreeningAction:
-    task_id = observation.task_id
-    extracted = observation.extracted_points
-    if task_id == "easy_eligibility":
-        for field_name, value in [
-            ("age", "47"),
-            ("diagnosis", "metastatic nsclc"),
-            ("biomarker", "egfr exon 19 deletion"),
-            ("ecog", "1"),
-            ("active_cns_disease", "no"),
-        ]:
-            if field_name not in extracted:
-                return ClinicalTrialScreeningAction(
-                    action_type="extract_data",
-                    field_name=field_name,
-                    value=value,
-                )
-        return ClinicalTrialScreeningAction(action_type="final_decision", value="eligible")
-    if task_id == "medium_patient_ranking":
-        for field_name, value in [
-            ("P-M101_fit_score", "0.88"),
-            ("P-M102_fit_score", "0.71"),
-            ("P-M103_fit_score", "0.54"),
-            ("best_candidate", "P-M101"),
-            ("lowest_candidate", "P-M103"),
-        ]:
-            if field_name not in extracted:
-                return ClinicalTrialScreeningAction(
-                    action_type="extract_data",
-                    field_name=field_name,
-                    value=value,
-                )
-        return ClinicalTrialScreeningAction(
-            action_type="submit_ranking",
-            ranking=["P-M101", "P-M102", "P-M103"],
-        )
-    for field_name, value in [
-        ("age", "68"),
-        ("live_vaccine_days", "12"),
-        ("prednisone_mg", "20"),
-        ("surgery_days", "9"),
-        ("anc", "0.9"),
-    ]:
-        if field_name not in extracted:
-            return ClinicalTrialScreeningAction(
-                action_type="extract_data",
-                field_name=field_name,
-                value=value,
-            )
-    if observation.grader_score < 0.7:
-        return ClinicalTrialScreeningAction(
-            action_type="flag_exclusions",
-            exclusions=[
-                "live_vaccine_within_30_days",
-                "prednisone_over_10mg",
-                "major_surgery_within_14_days",
-                "anc_below_1.0",
-            ],
-        )
-    return ClinicalTrialScreeningAction(action_type="final_decision", value="exclude")
+def parse_action(raw_text: str) -> ClinicalTrialAction:
+    payload = json.loads(raw_text)
+    return ClinicalTrialAction.model_validate(payload)
 
 
 def get_model_action(
     client: OpenAI,
-    observation: ClinicalTrialScreeningObservation,
+    task_name: str,
+    step: int,
+    observation_payload: Dict,
     history: List[str],
-) -> tuple[ClinicalTrialScreeningAction, Optional[str]]:
-    prompt = build_user_prompt(observation, history)
+) -> ClinicalTrialAction:
+    user_prompt = build_user_prompt(task_name, step, observation_payload, history)
     try:
-        response = client.chat.completions.create(
+        completion = client.chat.completions.create(
             model=MODEL_NAME,
             messages=[
                 {"role": "system", "content": SYSTEM_PROMPT},
-                {"role": "user", "content": prompt},
+                {"role": "user", "content": user_prompt},
             ],
-            temperature=0.0,
-            max_tokens=200,
+            temperature=TEMPERATURE,
+            max_tokens=MAX_TOKENS,
+            stream=False,
         )
-        content = (response.choices[0].message.content or "").strip()
-        payload = json.loads(content)
-        return ClinicalTrialScreeningAction.model_validate(payload), None
-    except Exception as exc:
-        return fallback_action(observation), safe_error(str(exc))
+        content = (completion.choices[0].message.content or "").strip()
+        return parse_action(content)
+    except Exception:
+        return heuristic_action(task_name, step)
 
 
-async def create_env_client() -> ClinicalTrialScreeningEnvClient:
+def format_action(action: ClinicalTrialAction) -> str:
+    payload = {
+        "action_type": action.action_type,
+        "field_name": action.field_name,
+        "value": action.value,
+        "ranking": action.ranking,
+        "deviations": action.deviations,
+        "final_decision": action.final_decision,
+    }
+    return json.dumps(payload, separators=(",", ":"), sort_keys=True)
+
+
+async def create_env() -> ClinicalTrialEnv:
     if LOCAL_IMAGE_NAME:
-        return await ClinicalTrialScreeningEnvClient.from_docker_image(LOCAL_IMAGE_NAME)
+        return await ClinicalTrialEnv.from_docker_image(LOCAL_IMAGE_NAME)
     if ENV_BASE_URL:
-        client = ClinicalTrialScreeningEnvClient(base_url=ENV_BASE_URL)
-        return await client.connect()
-    client = ClinicalTrialScreeningEnvClient(base_url="http://localhost:8000")
-    return await client.connect()
+        env = ClinicalTrialEnv(base_url=ENV_BASE_URL)
+        await env.connect()
+        return env
+    raise RuntimeError("Set LOCAL_IMAGE_NAME for Docker execution or ENV_BASE_URL for an existing server.")
 
 
 async def main() -> None:
-    client = OpenAI(base_url=API_BASE_URL, api_key=HF_TOKEN)
-    env = await create_env_client()
+    client = OpenAI(base_url=API_BASE_URL, api_key=API_KEY)
+    env: Optional[ClinicalTrialEnv] = None
     rewards: List[float] = []
-    history: List[str] = []
     steps_taken = 0
-    final_score = 0.0
+    score = 0.5
     success = False
+    history: List[str] = []
     last_error: Optional[str] = None
-    result = None
 
     log_start(task=TASK_NAME, env=BENCHMARK, model=MODEL_NAME)
 
     try:
-        result = await env.reset()
+        env = await create_env()
+        result = await env.reset(task_id=TASK_NAME)
+
         for step in range(1, MAX_STEPS + 1):
             if result.done:
                 break
 
-            action, planning_error = get_model_action(client, result.observation, history)
-            result = await env.step(action)
+            action = get_model_action(
+                client=client,
+                task_name=TASK_NAME,
+                step=step,
+                observation_payload=result.observation.model_dump(mode="json"),
+                history=history,
+            )
+
+            try:
+                result = await env.step(action)
+                reward = float(result.reward or 0.0)
+                done = bool(result.done)
+                last_error = None
+            except Exception as exc:
+                reward = 0.0
+                done = True
+                last_error = sanitize_error(str(exc))
 
-            reward = float(result.reward or 0.0)
             rewards.append(reward)
             steps_taken = step
-            last_error = planning_error
             log_step(
                 step=step,
                 action=format_action(action),
                 reward=reward,
-                done=bool(result.done),
-                error=last_error,
-            )
-            history.append(
-                f"{result.observation.task_id}:{format_action(action)}:{reward:.2f}:{result.done}"
+                done=done,
+                error=sanitize_error(last_error),
             )
-            if result.done:
+            history.append(f"step={step} action={format_action(action)} reward={reward:.2f}")
+
+            if last_error is not None or done:
                 break
 
-        if result is not None:
-            final_score = float(result.observation.grader_score)
-            final_score = min(max(final_score, 0.0), 1.0)
-            success = bool(result.done) and final_score >= 0.99
+        if last_error is None and "result" in locals():
+            score = float(result.observation.reward_details.grader_score)
+        score = min(max(score, MIN_STRICT_SCORE), MAX_STRICT_SCORE)
+        success = last_error is None and score >= SUCCESS_SCORE_THRESHOLD
     finally:
-        await env.close()
-        log_end(success=success, steps=steps_taken, score=final_score, rewards=rewards)
+        if env is not None:
+            try:
+                await env.close()
+            except Exception as exc:
+                last_error = last_error or sanitize_error(str(exc))
+        log_end(success=success, steps=steps_taken, rewards=rewards)
 
 
 if __name__ == "__main__":