File size: 8,905 Bytes
2043afa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Groq-powered action suggester for PolypharmacyEnv."""

from __future__ import annotations

import json
import os
from typing import Any

from openai import OpenAI

from ..models import PolypharmacyAction, PolypharmacyObservation

DEFAULT_MODEL = "llama-3.1-8b-instant"
FALLBACK_MODELS = [
    "llama-3.1-8b-instant",
    "llama-3.3-70b-versatile",
    "gemma2-9b-it",
]
CRITICAL_DRUG_IDS = {"DRUG_WARFARIN", "DRUG_INSULIN_GLARGINE", "DRUG_DIGOXIN"}

SYSTEM_PROMPT = """You are a clinical medication safety assistant.
Return exactly one JSON object describing the next action.
Allowed output schema:
{
  "action_type": "query_ddi" | "propose_intervention" | "finish_review",
  "drug_id_1": "optional",
  "drug_id_2": "optional",
  "target_drug_id": "optional",
  "intervention_type": "stop|dose_reduce|substitute|add_monitoring|none",
  "proposed_new_drug_id": "optional",
  "rationale": "optional"
}
No markdown fences. No extra text.
Do NOT use finish_review early. First, gather evidence with query_ddi and/or
perform at least one meaningful intervention when needed.
"""


def _obs_to_prompt(obs: PolypharmacyObservation) -> str:
    meds = ", ".join(m.drug_id for m in obs.current_medications)
    conds = ", ".join(obs.conditions)
    return (
        f"Task: {obs.task_id}\n"
        f"Age: {obs.age}, sex: {obs.sex}\n"
        f"Conditions: {conds}\n"
        f"Medications: {meds}\n"
        f"Query budget: {obs.remaining_query_budget}\n"
        f"Intervention budget: {obs.remaining_intervention_budget}\n"
        f"Step index: {obs.step_index}\n"
        "Choose the single safest, most useful next action."
    )


def _parse_action(text: str) -> PolypharmacyAction:
    raw = text.strip()
    if raw.startswith("```"):
        raw = raw.split("\n", 1)[-1]
    if raw.endswith("```"):
        raw = raw.rsplit("```", 1)[0]
    raw = raw.strip()
    payload: dict[str, Any] = json.loads(raw)
    return PolypharmacyAction.model_validate(payload)


def _fallback_query_action(obs: PolypharmacyObservation) -> PolypharmacyAction:
    meds = [m.drug_id for m in obs.current_medications]
    if len(meds) >= 2 and obs.remaining_query_budget > 0:
        return PolypharmacyAction(
            action_type="query_ddi",
            drug_id_1=meds[0],
            drug_id_2=meds[1],
        )
    return PolypharmacyAction(action_type="finish_review")


def _norm_pair(a: str, b: str) -> tuple[str, str]:
    return (a, b) if a < b else (b, a)


def _pick_unseen_query_pair(obs: PolypharmacyObservation) -> tuple[str, str] | None:
    meds = [m.drug_id for m in obs.current_medications]
    if len(meds) < 2 or obs.remaining_query_budget <= 0:
        return None

    seen = {
        _norm_pair(q.drug_id_1, q.drug_id_2)
        for q in obs.interaction_queries
    }
    # Prioritize pairs containing high-risk drugs.
    high_risk = [m.drug_id for m in obs.current_medications if m.is_high_risk_elderly]
    ordered = high_risk + [m for m in meds if m not in set(high_risk)]

    for i in range(len(ordered)):
        for j in range(i + 1, len(ordered)):
            p = _norm_pair(ordered[i], ordered[j])
            if p not in seen:
                return p
    return None


def _pick_intervention_target(obs: PolypharmacyObservation) -> str | None:
    if obs.remaining_intervention_budget <= 0:
        return None
    med_set = {m.drug_id for m in obs.current_medications}

    # Use latest discovered severe/moderate query as intervention target.
    for q in reversed(obs.interaction_queries):
        if q.severity in ("severe", "moderate"):
            m1 = next((m for m in obs.current_medications if m.drug_id == q.drug_id_1), None)
            m2 = next((m for m in obs.current_medications if m.drug_id == q.drug_id_2), None)
            candidates = [m for m in (m1, m2) if m is not None]
            if not candidates:
                continue
            # Prefer non-critical risky drugs first.
            candidates.sort(
                key=lambda m: (
                    m.drug_id in CRITICAL_DRUG_IDS,
                    0 if any("avoid" in f for f in m.beers_flags) else 1,
                    0 if m.is_high_risk_elderly else 1,
                )
            )
            return candidates[0].drug_id

    # Fallback: if no severe/moderate discovered, still intervene on obviously
    # risky medications (Beers/high-risk flags) when budgets permit.
    risky = sorted(
        obs.current_medications,
        key=lambda m: (
            0 if any("avoid" in f for f in m.beers_flags) else 1,
            0 if m.is_high_risk_elderly else 1,
            1 if m.drug_id in CRITICAL_DRUG_IDS else 0,
        ),
    )
    for med in risky:
        if any("avoid" in f for f in med.beers_flags) or med.is_high_risk_elderly:
            return med.drug_id
    return None


def _rule_based_action(obs: PolypharmacyObservation) -> PolypharmacyAction | None:
    # If we already discovered significant risk, intervene before more querying.
    target = _pick_intervention_target(obs)
    if target and (
        obs.step_index >= 1
        and (
            obs.remaining_query_budget <= 2
            or len(obs.interaction_queries) >= 4
            or any(q.severity in ("severe", "moderate") for q in obs.interaction_queries)
        )
    ):
        intervention = "stop"
        rationale = "Remove likely contributor to discovered interaction risk"
        if target in CRITICAL_DRUG_IDS:
            # Avoid blunt stop for critical meds.
            intervention = "dose_reduce"
            rationale = "Critical medication: prefer dose reduction over abrupt stop"
        return PolypharmacyAction(
            action_type="propose_intervention",
            target_drug_id=target,
            intervention_type=intervention,
            rationale=rationale,
        )

    pair = _pick_unseen_query_pair(obs)
    if pair:
        return PolypharmacyAction(
            action_type="query_ddi",
            drug_id_1=pair[0],
            drug_id_2=pair[1],
        )

    if obs.remaining_intervention_budget > 0:
        # Final fallback before finish: at least one safety action.
        target = _pick_intervention_target(obs)
        if target:
            return PolypharmacyAction(
                action_type="propose_intervention",
                target_drug_id=target,
                intervention_type="dose_reduce"
                if target in CRITICAL_DRUG_IDS
                else "stop",
                rationale="Fallback intervention when query options are exhausted",
            )

    if obs.step_index >= 3:
        return PolypharmacyAction(action_type="finish_review")
    return None


def _postprocess_action(
    obs: PolypharmacyObservation, action: PolypharmacyAction
) -> PolypharmacyAction:
    # First apply deterministic guardrails to avoid repetitive loops.
    ruled = _rule_based_action(obs)
    if ruled is not None:
        return ruled

    # Guardrail: prevent useless immediate finish actions.
    if action.action_type == "finish_review":
        if obs.step_index < 2 and obs.remaining_query_budget > 0:
            return _fallback_query_action(obs)
        if len(obs.interaction_queries) == 0 and obs.remaining_query_budget > 0:
            return _fallback_query_action(obs)
    return action


def suggest_action_from_observation(
    observation: PolypharmacyObservation,
    model_name: str | None = None,
) -> PolypharmacyAction:
    """Use Groq chat completions to suggest a valid action."""
    api_key = os.getenv("GROQ_API_KEY", "").strip()
    if not api_key:
        raise ValueError("GROQ_API_KEY is missing. Add it to your .env file.")

    base_url = os.getenv("GROQ_BASE_URL", "https://api.groq.com/openai/v1").strip()
    model = (model_name or os.getenv("GROQ_MODEL_NAME", DEFAULT_MODEL)).strip()
    client = OpenAI(api_key=api_key, base_url=base_url)

    user_prompt = _obs_to_prompt(observation)
    tried: list[tuple[str, str]] = []
    candidates: list[str] = [model] + [m for m in FALLBACK_MODELS if m != model]

    for candidate in candidates:
        try:
            resp = client.chat.completions.create(
                model=candidate,
                messages=[
                    {"role": "system", "content": SYSTEM_PROMPT},
                    {"role": "user", "content": user_prompt},
                ],
                temperature=0.2,
                max_tokens=220,
            )
            generated = (resp.choices[0].message.content or "").strip()
            parsed = _parse_action(generated)
            return _postprocess_action(observation, parsed)
        except Exception as exc:
            tried.append((candidate, str(exc)))

    tried_txt = " | ".join(f"{m}: {err}" for m, err in tried)
    raise ValueError(
        "No Groq model worked. Try one of: "
        "llama-3.3-70b-versatile, llama-3.1-8b-instant, gemma2-9b-it. "
        f"Errors: {tried_txt}"
    )