Update agent.py
Browse files
agent.py
CHANGED
@@ -1,250 +1,477 @@
|
|
1 |
-
"""
|
2 |
-
SynapseAI Clinical Decision Support System
|
3 |
-
Expert-Level Implementation with Safety-Centric Architecture
|
4 |
-
"""
|
5 |
-
|
6 |
import os
|
7 |
import re
|
8 |
import json
|
9 |
import logging
|
10 |
-
|
11 |
-
Callable, Sequence, Tuple, Union)
|
12 |
from functools import lru_cache
|
13 |
-
from
|
14 |
|
15 |
import requests
|
16 |
-
from pydantic import BaseModel, Field, ValidationError
|
17 |
from langchain_groq import ChatGroq
|
18 |
-
from
|
19 |
-
|
20 |
-
from langchain_core.
|
21 |
-
from
|
22 |
from langgraph.prebuilt import ToolExecutor
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
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 |
-
@tool("order_lab_test", args_schema=LabOrderInput)
|
68 |
-
def order_lab_test(test_name: str, rationale: str,
|
69 |
-
priority: ClinicalPriority) -> Dict[str, Any]:
|
70 |
-
"""Standardized lab ordering with clinical validation"""
|
71 |
-
# Implementation details...
|
72 |
-
return {"status": "ordered", "details": {...}}
|
73 |
-
|
74 |
-
class PrescriptionSafetyCheck(BaseModel):
|
75 |
-
medication: str
|
76 |
-
rxcui: Optional[str]
|
77 |
-
contraindications: List[str]
|
78 |
-
# Additional safety fields...
|
79 |
-
|
80 |
-
@classmethod
|
81 |
-
def validate_prescription(cls, rx_data: Dict) -> PrescriptionSafetyCheck:
|
82 |
-
"""Pharmaceutical safety validation pipeline"""
|
83 |
-
# Comprehensive validation logic...
|
84 |
-
return PrescriptionSafetyCheck(...)
|
85 |
-
|
86 |
-
# ββ State Management Engine βββββββββββββοΏ½οΏ½ββββββββββββββββββββββββββββ
|
87 |
-
class ClinicalStateManager:
|
88 |
-
@staticmethod
|
89 |
-
def initialize_state(patient_data: Dict) -> ClinicalState:
|
90 |
-
"""Create validated initial state with clinical context"""
|
91 |
-
return {
|
92 |
-
"messages": [
|
93 |
-
SystemMessage(content=ClinicalPrompts.SYSTEM_PROMPT),
|
94 |
-
HumanMessage(content="Initiate clinical consultation")
|
95 |
-
],
|
96 |
-
"patient_data": ClinicalValidator.sanitize_patient_data(patient_data),
|
97 |
-
"safety_warnings": [],
|
98 |
-
"workflow_metadata": {
|
99 |
-
"iterations": 0,
|
100 |
-
"active_alerts": 0,
|
101 |
-
"safety_override": False
|
102 |
-
},
|
103 |
-
"execution_log": []
|
104 |
-
}
|
105 |
-
|
106 |
-
@staticmethod
|
107 |
-
def propagate_state(previous: ClinicalState,
|
108 |
-
updates: Dict) -> ClinicalState:
|
109 |
-
"""State transition with clinical context preservation"""
|
110 |
-
preserved_fields = {
|
111 |
-
'patient_data': previous['patient_data'],
|
112 |
-
'workflow_metadata': {
|
113 |
-
**previous['workflow_metadata'],
|
114 |
-
**updates.get('workflow_metadata', {})
|
115 |
-
}
|
116 |
-
}
|
117 |
-
return ClinicalValidator.validate_state({
|
118 |
-
**preserved_fields,
|
119 |
-
**updates
|
120 |
-
})
|
121 |
|
122 |
-
# ββ
|
123 |
-
class
|
124 |
-
|
125 |
-
|
126 |
-
|
127 |
-
|
128 |
-
|
129 |
-
|
130 |
-
|
131 |
-
|
132 |
-
|
133 |
-
|
134 |
-
|
135 |
-
|
136 |
-
|
137 |
-
|
138 |
-
|
139 |
-
|
140 |
-
|
141 |
-
return new_state
|
142 |
-
except CriticalClinicalError as e:
|
143 |
-
return ClinicalErrorHandler.handle_critical_error(state, e)
|
144 |
-
|
145 |
-
@staticmethod
|
146 |
-
def tool_node(state: ClinicalState) -> ClinicalState:
|
147 |
-
"""HIPAA-compliant tool execution with safety audit"""
|
148 |
-
ClinicalSafetyEngine.pre_execution_checks(state)
|
149 |
-
|
150 |
-
tool_results = []
|
151 |
-
for tool_call in state["messages"][-1].tool_calls:
|
152 |
-
result = ClinicalToolExecutor.execute_with_audit(tool_call)
|
153 |
-
tool_results.append(result)
|
154 |
-
|
155 |
-
if result['category'] == "DRUG_ORDER":
|
156 |
-
ClinicalSafetyEngine.post_drug_order_checks(result)
|
157 |
-
|
158 |
-
return ClinicalStateManager.propagate_state(state, {
|
159 |
-
"messages": [ToolMessage(...)],
|
160 |
-
"safety_warnings": ClinicalSafetyEngine.aggregate_warnings(tool_results)
|
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 |
def __init__(self):
|
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 |
-
workflow.add_edge("tool_execution", "clinical_reasoning")
|
219 |
-
workflow.add_edge("safety_review", "clinical_reasoning")
|
220 |
-
|
221 |
-
return workflow.compile()
|
222 |
-
|
223 |
-
def execute_consultation(self, patient_data: Dict) -> ClinicalState:
|
224 |
-
"""Execute full clinical workflow with safety audits"""
|
225 |
-
initial_state = ClinicalStateManager.initialize_state(patient_data)
|
226 |
-
return self.workflow.invoke(
|
227 |
-
initial_state,
|
228 |
-
config={"recursion_limit": ClinicalConfig.MAX_ITERATIONS + ClinicalConfig.RECURSION_BUFFER}
|
229 |
-
)
|
230 |
-
|
231 |
-
# ββ Usage Example βββββββββββββββββββββββββββββββββββββββββββββββββββ
|
232 |
-
if __name__ == "__main__":
|
233 |
-
# Initialize clinical environment
|
234 |
-
ClinicalValidator.validate_environment()
|
235 |
-
|
236 |
-
# Sample patient scenario
|
237 |
-
complex_case = {
|
238 |
-
"demographics": {"age": 68, "sex": "F", "weight_kg": 82},
|
239 |
-
"presenting_complaint": "Chest pain radiating to left arm",
|
240 |
-
"medical_history": ["HTN", "Type 2 DM", "HLD"],
|
241 |
-
"current_meds": ["Atenolol 50mg daily", "Simvastatin 40mg HS"]
|
242 |
-
}
|
243 |
-
|
244 |
-
# Execute clinical workflow
|
245 |
-
workflow = ClinicalWorkflow()
|
246 |
-
result = workflow.execute_consultation(complex_case)
|
247 |
-
|
248 |
-
# Generate clinical summary
|
249 |
-
final_report = ClinicalDocumentation.generate_report(result)
|
250 |
-
print(json.dumps(final_report, indent=2))
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
import os
|
2 |
import re
|
3 |
import json
|
4 |
import logging
|
5 |
+
import traceback
|
|
|
6 |
from functools import lru_cache
|
7 |
+
from typing import List, Dict, Any, Optional, TypedDict
|
8 |
|
9 |
import requests
|
|
|
10 |
from langchain_groq import ChatGroq
|
11 |
+
from langchain_community.tools.tavily_search import TavilySearchResults
|
12 |
+
from langchain_core.messages import HumanMessage, SystemMessage, AIMessage, ToolMessage
|
13 |
+
from langchain_core.pydantic_v1 import BaseModel, Field
|
14 |
+
from langchain_core.tools import tool
|
15 |
from langgraph.prebuilt import ToolExecutor
|
16 |
+
from langgraph.graph import StateGraph, END
|
17 |
+
|
18 |
+
# ββ Logging Configuration ββββββββββββββββββββββββββββββββββββββββββββββ
|
19 |
+
logger = logging.getLogger(__name__)
|
20 |
+
logging.basicConfig(level=logging.INFO)
|
21 |
+
|
22 |
+
# ββ Environment Variables ββββββββββββββββββββββββββββββββββββββββββββββ
|
23 |
+
UMLS_API_KEY = os.getenv("UMLS_API_KEY")
|
24 |
+
GROQ_API_KEY = os.getenv("GROQ_API_KEY")
|
25 |
+
TAVILY_API_KEY = os.getenv("TAVILY_API_KEY")
|
26 |
+
|
27 |
+
if not all([UMLS_API_KEY, GROQ_API_KEY, TAVILY_API_KEY]):
|
28 |
+
logger.error("Missing one or more required API keys: UMLS_API_KEY, GROQ_API_KEY, TAVILY_API_KEY")
|
29 |
+
raise RuntimeError("Missing required API keys")
|
30 |
+
|
31 |
+
# ββ Agent Configuration ββββββββββββββββββββββββββββββββββββββββββββββ
|
32 |
+
AGENT_MODEL_NAME = "llama3-70b-8192"
|
33 |
+
AGENT_TEMPERATURE = 0.1
|
34 |
+
MAX_SEARCH_RESULTS = 3
|
35 |
+
|
36 |
+
class ClinicalPrompts:
|
37 |
+
SYSTEM_PROMPT = """
|
38 |
+
You are SynapseAI, an expert AI clinical assistant engaged in an interactive consultation...
|
39 |
+
[SYSTEM PROMPT CONTENT HERE]
|
40 |
+
"""
|
41 |
+
|
42 |
+
# ββ Helper Functions βββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
43 |
+
UMLS_AUTH_ENDPOINT = "https://utslogin.nlm.nih.gov/cas/v1/api-key"
|
44 |
+
RXNORM_API_BASE = "https://rxnav.nlm.nih.gov/REST"
|
45 |
+
OPENFDA_API_BASE = "https://api.fda.gov/drug/label.json"
|
46 |
+
|
47 |
+
@lru_cache(maxsize=256)
|
48 |
+
def get_rxcui(drug_name: str) -> Optional[str]:
|
49 |
+
"""Lookup RxNorm CUI for a given drug name."""
|
50 |
+
drug_name = (drug_name or "").strip()
|
51 |
+
if not drug_name:
|
52 |
+
return None
|
53 |
+
logger.info(f"Looking up RxCUI for '{drug_name}'")
|
54 |
+
try:
|
55 |
+
params = {"name": drug_name, "search": 1}
|
56 |
+
r = requests.get(f"{RXNORM_API_BASE}/rxcui.json", params=params, timeout=10)
|
57 |
+
r.raise_for_status()
|
58 |
+
ids = r.json().get("idGroup", {}).get("rxnormId")
|
59 |
+
if ids:
|
60 |
+
logger.info(f"Found RxCUI {ids[0]} for '{drug_name}'")
|
61 |
+
return ids[0]
|
62 |
+
r = requests.get(f"{RXNORM_API_BASE}/drugs.json", params={"name": drug_name}, timeout=10)
|
63 |
+
r.raise_for_status()
|
64 |
+
for grp in r.json().get("drugGroup", {}).get("conceptGroup", []):
|
65 |
+
props = grp.get("conceptProperties")
|
66 |
+
if props:
|
67 |
+
logger.info(f"Found RxCUI {props[0]['rxcui']} via /drugs for '{drug_name}'")
|
68 |
+
return props[0]["rxcui"]
|
69 |
+
except Exception:
|
70 |
+
logger.exception(f"Error fetching RxCUI for '{drug_name}'")
|
71 |
+
return None
|
72 |
|
73 |
+
@lru_cache(maxsize=128)
|
74 |
+
def get_openfda_label(rxcui: Optional[str] = None, drug_name: Optional[str] = None) -> Optional[Dict[str, Any]]:
|
75 |
+
"""Fetch the OpenFDA label for a drug by RxCUI or name."""
|
76 |
+
if not (rxcui or drug_name):
|
77 |
+
return None
|
78 |
+
terms = []
|
79 |
+
if rxcui:
|
80 |
+
terms.append(f'spl_rxnorm_code:"{rxcui}" OR openfda.rxcui:"{rxcui}"')
|
81 |
+
if drug_name:
|
82 |
+
dn = drug_name.lower()
|
83 |
+
terms.append(f'(openfda.brand_name:"{dn}" OR openfda.generic_name:"{dn}")')
|
84 |
+
query = " OR ".join(terms)
|
85 |
+
logger.info(f"Looking up OpenFDA label with query: {query}")
|
86 |
+
try:
|
87 |
+
r = requests.get(OPENFDA_API_BASE, params={"search": query, "limit": 1}, timeout=15)
|
88 |
+
r.raise_for_status()
|
89 |
+
results = r.json().get("results", [])
|
90 |
+
if results:
|
91 |
+
return results[0]
|
92 |
+
except Exception:
|
93 |
+
logger.exception("Error fetching OpenFDA label")
|
94 |
+
return None
|
95 |
+
|
96 |
+
def search_text_list(texts: List[str], terms: List[str]) -> List[str]:
|
97 |
+
"""Return highlighted snippets from a list of texts containing any of the search terms."""
|
98 |
+
snippets = []
|
99 |
+
lowers = [t.lower() for t in terms if t]
|
100 |
+
for text in texts or []:
|
101 |
+
tl = text.lower()
|
102 |
+
for term in lowers:
|
103 |
+
if term in tl:
|
104 |
+
i = tl.find(term)
|
105 |
+
start = max(0, i - 50)
|
106 |
+
end = min(len(text), i + len(term) + 100)
|
107 |
+
snippet = text[start:end]
|
108 |
+
snippet = re.sub(f"({re.escape(term)})", r"**\1**", snippet, flags=re.IGNORECASE)
|
109 |
+
snippets.append(f"...{snippet}...")
|
110 |
+
break
|
111 |
+
return snippets
|
112 |
+
|
113 |
+
def parse_bp(bp: str) -> Optional[tuple[int, int]]:
|
114 |
+
"""Parse 'SYS/DIA' blood pressure string into a (sys, dia) tuple."""
|
115 |
+
if m := re.match(r"(\d{1,3})\s*/\s*(\d{1,3})", (bp or "").strip()):
|
116 |
+
return int(m.group(1)), int(m.group(2))
|
117 |
+
return None
|
118 |
+
|
119 |
+
def check_red_flags(patient_data: Dict[str, Any]) -> List[str]:
|
120 |
+
"""Identify immediate red flags from patient_data."""
|
121 |
+
flags: List[str] = []
|
122 |
+
hpi = patient_data.get("hpi", {})
|
123 |
+
vitals = patient_data.get("vitals", {})
|
124 |
+
syms = [s.lower() for s in hpi.get("symptoms", []) if isinstance(s, str)]
|
125 |
+
mapping = {
|
126 |
+
"chest pain": "Chest pain reported",
|
127 |
+
"shortness of breath": "Shortness of breath reported",
|
128 |
+
"severe headache": "Severe headache reported",
|
129 |
+
"syncope": "Syncope reported",
|
130 |
+
"hemoptysis": "Hemoptysis reported"
|
131 |
}
|
132 |
+
for term, desc in mapping.items():
|
133 |
+
if term in syms:
|
134 |
+
flags.append(f"Red Flag: {desc}.")
|
135 |
+
temp = vitals.get("temp_c")
|
136 |
+
hr = vitals.get("hr_bpm")
|
137 |
+
rr = vitals.get("rr_rpm")
|
138 |
+
spo2 = vitals.get("spo2_percent")
|
139 |
+
bp = parse_bp(vitals.get("bp_mmhg", ""))
|
140 |
+
if temp is not None and temp >= 38.5:
|
141 |
+
flags.append(f"Red Flag: Fever ({temp}Β°C).")
|
142 |
+
if hr is not None:
|
143 |
+
if hr >= 120:
|
144 |
+
flags.append(f"Red Flag: Tachycardia ({hr} bpm).")
|
145 |
+
if hr <= 50:
|
146 |
+
flags.append(f"Red Flag: Bradycardia ({hr} bpm).")
|
147 |
+
if rr is not None and rr >= 24:
|
148 |
+
flags.append(f"Red Flag: Tachypnea ({rr} rpm).")
|
149 |
+
if spo2 is not None and spo2 <= 92:
|
150 |
+
flags.append(f"Red Flag: Hypoxia ({spo2}%).")
|
151 |
+
if bp:
|
152 |
+
sys, dia = bp
|
153 |
+
if sys >= 180 or dia >= 110:
|
154 |
+
flags.append(f"Red Flag: Hypertensive urgency/emergency ({sys}/{dia} mmHg).")
|
155 |
+
if sys <= 90 or dia <= 60:
|
156 |
+
flags.append(f"Red Flag: Hypotension ({sys}/{dia} mmHg).")
|
157 |
+
return list(dict.fromkeys(flags))
|
158 |
|
159 |
+
def format_patient_data_for_prompt(data: Dict[str, Any]) -> str:
|
160 |
+
"""Format patient_data dict into a markdown-like prompt section."""
|
161 |
+
if not data:
|
162 |
+
return "No patient data provided."
|
163 |
+
lines: List[str] = []
|
164 |
+
for section, value in data.items():
|
165 |
+
title = section.replace("_", " ").title()
|
166 |
+
if isinstance(value, dict) and any(value.values()):
|
167 |
+
lines.append(f"**{title}:**")
|
168 |
+
for k, v in value.items():
|
169 |
+
if v:
|
170 |
+
lines.append(f"- {k.replace('_',' ').title()}: {v}")
|
171 |
+
elif isinstance(value, list) and value:
|
172 |
+
lines.append(f"**{title}:** {', '.join(map(str, value))}")
|
173 |
+
elif value:
|
174 |
+
lines.append(f"**{title}:** {value}")
|
175 |
+
return "\n".join(lines)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
176 |
|
177 |
+
# ββ Tool Input Schemas βββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
178 |
+
class LabOrderInput(BaseModel):
|
179 |
+
test_name: str = Field(...)
|
180 |
+
reason: str = Field(...)
|
181 |
+
priority: str = Field("Routine")
|
182 |
+
|
183 |
+
class PrescriptionInput(BaseModel):
|
184 |
+
medication_name: str = Field(...)
|
185 |
+
dosage: str = Field(...)
|
186 |
+
route: str = Field(...)
|
187 |
+
frequency: str = Field(...)
|
188 |
+
duration: str = Field("As directed")
|
189 |
+
reason: str = Field(...)
|
190 |
+
|
191 |
+
class InteractionCheckInput(BaseModel):
|
192 |
+
potential_prescription: str
|
193 |
+
current_medications: Optional[List[str]] = Field(None)
|
194 |
+
allergies: Optional[List[str]] = Field(None)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
195 |
|
196 |
+
class FlagRiskInput(BaseModel):
|
197 |
+
risk_description: str = Field(...)
|
198 |
+
urgency: str = Field("High")
|
199 |
+
|
200 |
+
# ββ Tool Implementations βββββββββββββββββββββββββββββββββββββββββββββββββββ
|
201 |
+
@tool("order_lab_test", args_schema=LabOrderInput)
|
202 |
+
def order_lab_test(test_name: str, reason: str, priority: str = "Routine") -> str:
|
203 |
+
"""
|
204 |
+
Place an order for a laboratory test.
|
205 |
+
"""
|
206 |
+
logger.info(f"Ordering lab test: {test_name}, reason: {reason}, priority: {priority}")
|
207 |
+
return json.dumps({
|
208 |
+
"status": "success",
|
209 |
+
"message": f"Lab Ordered: {test_name} ({priority})",
|
210 |
+
"details": f"Reason: {reason}"
|
211 |
+
})
|
212 |
+
|
213 |
+
@tool("prescribe_medication", args_schema=PrescriptionInput)
|
214 |
+
def prescribe_medication(
|
215 |
+
medication_name: str,
|
216 |
+
dosage: str,
|
217 |
+
route: str,
|
218 |
+
frequency: str,
|
219 |
+
duration: str,
|
220 |
+
reason: str
|
221 |
+
) -> str:
|
222 |
+
"""
|
223 |
+
Prepare a medication prescription.
|
224 |
+
"""
|
225 |
+
logger.info(f"Preparing prescription: {medication_name} {dosage}, route: {route}, freq: {frequency}")
|
226 |
+
return json.dumps({
|
227 |
+
"status": "success",
|
228 |
+
"message": f"Prescription Prepared: {medication_name} {dosage} {route} {frequency}",
|
229 |
+
"details": f"Duration: {duration}. Reason: {reason}"
|
230 |
+
})
|
231 |
+
|
232 |
+
@tool("check_drug_interactions", args_schema=InteractionCheckInput)
|
233 |
+
def check_drug_interactions(
|
234 |
+
potential_prescription: str,
|
235 |
+
current_medications: Optional[List[str]] = None,
|
236 |
+
allergies: Optional[List[str]] = None
|
237 |
+
) -> str:
|
238 |
+
"""
|
239 |
+
Check for drugβdrug interactions and allergy risks.
|
240 |
+
"""
|
241 |
+
logger.info(f"Checking interactions for: {potential_prescription}")
|
242 |
+
warnings: List[str] = []
|
243 |
+
pm = [m.lower().strip() for m in (current_medications or []) if m]
|
244 |
+
al = [a.lower().strip() for a in (allergies or []) if a]
|
245 |
+
if potential_prescription.lower().strip() in al:
|
246 |
+
warnings.append(f"CRITICAL ALLERGY: Patient allergic to '{potential_prescription}'.")
|
247 |
+
rxcui = get_rxcui(potential_prescription)
|
248 |
+
label = get_openfda_label(rxcui=rxcui, drug_name=potential_prescription)
|
249 |
+
if not (rxcui or label):
|
250 |
+
warnings.append(f"INFO: Could not identify '{potential_prescription}'. Checks may be incomplete.")
|
251 |
+
for section in ("contraindications", "warnings_and_cautions", "warnings"):
|
252 |
+
items = label.get(section) if label else None
|
253 |
+
if isinstance(items, list):
|
254 |
+
snippets = search_text_list(items, al)
|
255 |
+
if snippets:
|
256 |
+
warnings.append(f"ALLERGY RISK ({section}): {'; '.join(snippets)}")
|
257 |
+
for med in pm:
|
258 |
+
mrxcui = get_rxcui(med)
|
259 |
+
mlabel = get_openfda_label(rxcui=mrxcui, drug_name=med)
|
260 |
+
for sec in ("drug_interactions",):
|
261 |
+
for src_label, src_name in ((label, potential_prescription), (mlabel, med)):
|
262 |
+
items = src_label.get(sec) if src_label else None
|
263 |
+
if isinstance(items, list):
|
264 |
+
snippets = search_text_list(items, [med if src_name == potential_prescription else potential_prescription])
|
265 |
+
if snippets:
|
266 |
+
warnings.append(f"Interaction ({src_name} label): {'; '.join(snippets)}")
|
267 |
+
status = "warning" if warnings else "clear"
|
268 |
+
message = (
|
269 |
+
f"{len(warnings)} issue(s) found for '{potential_prescription}'."
|
270 |
+
if warnings else
|
271 |
+
f"No major interactions or allergy issues identified for '{potential_prescription}'."
|
272 |
+
)
|
273 |
+
return json.dumps({"status": status, "message": message, "warnings": warnings})
|
274 |
+
|
275 |
+
@tool("flag_risk", args_schema=FlagRiskInput)
|
276 |
+
def flag_risk(risk_description: str, urgency: str = "High") -> str:
|
277 |
+
"""
|
278 |
+
Flag a clinical risk with given urgency.
|
279 |
+
"""
|
280 |
+
logger.info(f"Flagging risk: {risk_description} (urgency={urgency})")
|
281 |
+
return json.dumps({
|
282 |
+
"status": "flagged",
|
283 |
+
"message": f"Risk '{risk_description}' flagged with {urgency} urgency."
|
284 |
+
})
|
285 |
+
|
286 |
+
# Include the Tavily search tool
|
287 |
+
search_tool = TavilySearchResults(max_results=MAX_SEARCH_RESULTS, name="tavily_search_results")
|
288 |
+
all_tools = [order_lab_test, prescribe_medication, check_drug_interactions, flag_risk, search_tool]
|
289 |
+
|
290 |
+
# ββ LLM & Tool Executor βββββββββββββββββββββββββββββββββββββββββββββββββββ
|
291 |
+
llm = ChatGroq(temperature=AGENT_TEMPERATURE, model=AGENT_MODEL_NAME)
|
292 |
+
model_with_tools = llm.bind_tools(all_tools)
|
293 |
+
tool_executor = ToolExecutor(all_tools)
|
294 |
+
|
295 |
+
# ββ State Definition βββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
296 |
+
class AgentState(TypedDict):
|
297 |
+
messages: List[Any]
|
298 |
+
patient_data: Optional[Dict[str, Any]]
|
299 |
+
summary: Optional[str]
|
300 |
+
interaction_warnings: Optional[List[str]]
|
301 |
+
done: Optional[bool]
|
302 |
+
iterations: Optional[int]
|
303 |
+
|
304 |
+
# Helper to propagate state fields between nodes
|
305 |
+
def propagate_state(new: Dict[str, Any], old: Dict[str, Any]) -> Dict[str, Any]:
|
306 |
+
for key in ["iterations", "done", "patient_data", "summary", "interaction_warnings"]:
|
307 |
+
if key in old and key not in new:
|
308 |
+
new[key] = old[key]
|
309 |
+
return new
|
310 |
+
|
311 |
+
# ββ Graph Nodes βββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
312 |
+
def agent_node(state: AgentState) -> Dict[str, Any]:
|
313 |
+
if state.get("done", False):
|
314 |
+
return state
|
315 |
+
msgs = state.get("messages", [])
|
316 |
+
if not msgs or not isinstance(msgs[0], SystemMessage):
|
317 |
+
msgs = [SystemMessage(content=ClinicalPrompts.SYSTEM_PROMPT)] + msgs
|
318 |
+
logger.info(f"Invoking LLM with {len(msgs)} messages")
|
319 |
+
try:
|
320 |
+
response = model_with_tools.invoke(msgs)
|
321 |
+
new_state = {"messages": [response]}
|
322 |
+
return propagate_state(new_state, state)
|
323 |
+
except Exception as e:
|
324 |
+
logger.exception("Error in agent_node")
|
325 |
+
new_state = {"messages": [AIMessage(content=f"Error: {e}")]}
|
326 |
+
return propagate_state(new_state, state)
|
327 |
+
|
328 |
+
def tool_node(state: AgentState) -> Dict[str, Any]:
|
329 |
+
if state.get("done", False):
|
330 |
+
return state
|
331 |
+
last = state.get("messages", [])[-1]
|
332 |
+
if not isinstance(last, AIMessage) or not getattr(last, "tool_calls", None):
|
333 |
+
logger.warning("tool_node invoked without pending tool_calls")
|
334 |
+
new_state = {"messages": []}
|
335 |
+
return propagate_state(new_state, state)
|
336 |
+
calls = last.tool_calls
|
337 |
+
blocked_ids = set()
|
338 |
+
for call in calls:
|
339 |
+
if call["name"] == "prescribe_medication":
|
340 |
+
med = call["args"].get("medication_name", "").lower()
|
341 |
+
if not any(
|
342 |
+
c["name"] == "check_drug_interactions" and
|
343 |
+
c["args"].get("potential_prescription", "").lower() == med
|
344 |
+
for c in calls
|
345 |
+
):
|
346 |
+
logger.warning(f"Blocking prescribe_medication for '{med}' without interaction check")
|
347 |
+
blocked_ids.add(call["id"])
|
348 |
+
to_execute = [c for c in calls if c["id"] not in blocked_ids]
|
349 |
+
pd = state.get("patient_data", {})
|
350 |
+
for call in to_execute:
|
351 |
+
if call["name"] == "check_drug_interactions":
|
352 |
+
call["args"].setdefault("current_medications", pd.get("medications", {}).get("current", []))
|
353 |
+
call["args"].setdefault("allergies", pd.get("allergies", []))
|
354 |
+
messages: List[ToolMessage] = []
|
355 |
+
warnings: List[str] = []
|
356 |
+
try:
|
357 |
+
responses = tool_executor.batch(to_execute, return_exceptions=True)
|
358 |
+
for call, resp in zip(to_execute, responses):
|
359 |
+
if isinstance(resp, Exception):
|
360 |
+
logger.exception(f"Error executing tool {call['name']}")
|
361 |
+
content = json.dumps({"status": "error", "message": str(resp)})
|
362 |
+
else:
|
363 |
+
content = str(resp)
|
364 |
+
if call["name"] == "check_drug_interactions":
|
365 |
+
data = json.loads(content)
|
366 |
+
if data.get("status") == "warning":
|
367 |
+
warnings.extend(data.get("warnings", []))
|
368 |
+
messages.append(ToolMessage(content=content, tool_call_id=call["id"], name=call["name"]))
|
369 |
+
except Exception as e:
|
370 |
+
logger.exception("Critical error in tool_node")
|
371 |
+
for call in to_execute:
|
372 |
+
messages.append(ToolMessage(
|
373 |
+
content=json.dumps({"status": "error", "message": str(e)}),
|
374 |
+
tool_call_id=call["id"],
|
375 |
+
name=call["name"]
|
376 |
+
))
|
377 |
+
new_state = {"messages": messages, "interaction_warnings": warnings or None}
|
378 |
+
return propagate_state(new_state, state)
|
379 |
+
|
380 |
+
def reflection_node(state: AgentState) -> Dict[str, Any]:
|
381 |
+
if state.get("done", False):
|
382 |
+
return state
|
383 |
+
warns = state.get("interaction_warnings")
|
384 |
+
if not warns:
|
385 |
+
logger.warning("reflection_node called without warnings")
|
386 |
+
new_state = {"messages": []}
|
387 |
+
return propagate_state(new_state, state)
|
388 |
+
triggering = None
|
389 |
+
for msg in reversed(state.get("messages", [])):
|
390 |
+
if isinstance(msg, AIMessage) and getattr(msg, "tool_calls", None):
|
391 |
+
triggering = msg
|
392 |
+
break
|
393 |
+
if not triggering:
|
394 |
+
new_state = {"messages": [AIMessage(content="Internal Error: reflection context missing.")]}
|
395 |
+
return propagate_state(new_state, state)
|
396 |
+
prompt = (
|
397 |
+
"You are SynapseAI, performing a focused safety review of the following plan:\n\n"
|
398 |
+
f"{triggering.content}\n\n"
|
399 |
+
"Highlight any issues based on these warnings:\n" +
|
400 |
+
"\n".join(f"- {w}" for w in warns)
|
401 |
+
)
|
402 |
+
try:
|
403 |
+
resp = llm.invoke([SystemMessage(content="Safety reflection"), HumanMessage(content=prompt)])
|
404 |
+
new_state = {"messages": [AIMessage(content=resp.content)]}
|
405 |
+
return propagate_state(new_state, state)
|
406 |
+
except Exception as e:
|
407 |
+
logger.exception("Error during reflection")
|
408 |
+
new_state = {"messages": [AIMessage(content=f"Error during reflection: {e}")]}
|
409 |
+
return propagate_state(new_state, state)
|
410 |
+
|
411 |
+
# ββ Routing Functions ββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
412 |
+
def should_continue(state: AgentState) -> str:
|
413 |
+
state.setdefault("iterations", 0)
|
414 |
+
state["iterations"] += 1
|
415 |
+
logger.info(f"Iteration count: {state['iterations']}")
|
416 |
+
# When iterations exceed threshold, force final output and terminate.
|
417 |
+
if state["iterations"] >= 4:
|
418 |
+
state.setdefault("messages", []).append(AIMessage(content="Final output: consultation complete."))
|
419 |
+
state["done"] = True
|
420 |
+
return "end_conversation_turn"
|
421 |
+
if not state.get("messages"):
|
422 |
+
state["done"] = True
|
423 |
+
return "end_conversation_turn"
|
424 |
+
last = state["messages"][-1]
|
425 |
+
if not isinstance(last, AIMessage):
|
426 |
+
state["done"] = True
|
427 |
+
return "end_conversation_turn"
|
428 |
+
if getattr(last, "tool_calls", None):
|
429 |
+
return "continue_tools"
|
430 |
+
if "consultation complete" in last.content.lower():
|
431 |
+
state["done"] = True
|
432 |
+
return "end_conversation_turn"
|
433 |
+
state["done"] = False
|
434 |
+
return "agent"
|
435 |
+
|
436 |
+
def after_tools_router(state: AgentState) -> str:
|
437 |
+
# Instead of routing back to agent, route reflection to END to break the cycle.
|
438 |
+
if state.get("interaction_warnings"):
|
439 |
+
return "reflection"
|
440 |
+
return "end_conversation_turn"
|
441 |
+
|
442 |
+
# ββ ClinicalAgent βββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
443 |
+
class ClinicalAgent:
|
444 |
def __init__(self):
|
445 |
+
logger.info("Building ClinicalAgent workflow")
|
446 |
+
wf = StateGraph(AgentState)
|
447 |
+
wf.add_node("agent", agent_node)
|
448 |
+
wf.add_node("tools", tool_node)
|
449 |
+
wf.add_node("reflection", reflection_node)
|
450 |
+
wf.set_entry_point("agent")
|
451 |
+
wf.add_conditional_edges("agent", should_continue, {
|
452 |
+
"continue_tools": "tools",
|
453 |
+
"end_conversation_turn": END
|
454 |
+
})
|
455 |
+
wf.add_conditional_edges("tools", after_tools_router, {
|
456 |
+
"reflection": "reflection",
|
457 |
+
"end_conversation_turn": END
|
458 |
+
})
|
459 |
+
# Removed the edge from reflection back to agent to break the cycle.
|
460 |
+
self.graph_app = wf.compile()
|
461 |
+
logger.info("ClinicalAgent ready")
|
462 |
+
|
463 |
+
def invoke_turn(self, state: Dict[str, Any]) -> Dict[str, Any]:
|
464 |
+
try:
|
465 |
+
# Increase recursion limit if needed.
|
466 |
+
result = self.graph_app.invoke(state, {"recursion_limit": 100})
|
467 |
+
result.setdefault("summary", state.get("summary"))
|
468 |
+
result.setdefault("interaction_warnings", None)
|
469 |
+
return result
|
470 |
+
except Exception as e:
|
471 |
+
logger.exception("Error during graph invocation")
|
472 |
+
return {
|
473 |
+
"messages": state.get("messages", []) + [AIMessage(content=f"Error: {e}")],
|
474 |
+
"patient_data": state.get("patient_data"),
|
475 |
+
"summary": state.get("summary"),
|
476 |
+
"interaction_warnings": None
|
477 |
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|