SynapseAI / agent.py
mgbam's picture
Update agent.py
8e9de1e verified
raw
history blame
18 kB
import os
import re
import json
import requests
import traceback
import operator
from functools import lru_cache
from typing import Any, Dict, List, Optional, TypedDict, Annotated
from langchain_groq import ChatGroq
from langchain_community.tools.tavily_search import TavilySearchResults
from langchain_core.messages import HumanMessage, SystemMessage, AIMessage, ToolMessage
from langchain_core.pydantic_v1 import BaseModel, Field
from langchain_core.tools import tool
from langgraph.prebuilt import ToolExecutor
from langgraph.graph import StateGraph, END
# --- Configuration & Constants ---
UMLS_API_KEY = os.environ.get("UMLS_API_KEY")
GROQ_API_KEY = os.environ.get("GROQ_API_KEY")
TAVILY_API_KEY = os.environ.get("TAVILY_API_KEY")
AGENT_MODEL_NAME = "llama3-70b-8192"
AGENT_TEMPERATURE = 0.1
MAX_SEARCH_RESULTS = 3
UMLS_AUTH_ENDPOINT = "https://utslogin.nlm.nih.gov/cas/v1/api-key"
RXNORM_API_BASE = "https://rxnav.nlm.nih.gov/REST"
OPENFDA_API_BASE = "https://api.fda.gov/drug/label.json"
class ClinicalPrompts:
SYSTEM_PROMPT = (
"""
You are SynapseAI, an expert AI clinical assistant in an interactive consultation.
Analyze patient data, provide differential diagnoses, suggest management plans,
and identify risks according to current standards of care.
1. Process information sequentially; use full conversation history.
2. Ask for clarification if data is insufficient; do not guess.
3. When ready, output a complete JSON assessment as specified.
4. Before prescribing, run drug-interaction checks and report results.
5. Flag urgent red flags immediately.
6. Use tools logically; await results when needed.
7. Query clinical guidelines via tavily_search_results and cite them.
8. Be concise, accurate, and use standard terminology.
"""
)
# --- Helper Functions ---
@lru_cache(maxsize=256)
def get_rxcui(drug_name: str) -> Optional[str]:
"""Return RxNorm CUI for a given drug name."""
if not drug_name:
return None
name = drug_name.strip()
if not name:
return None
try:
# Primary lookup
params = {"name": name, "search": 1}
resp = requests.get(f"{RXNORM_API_BASE}/rxcui.json", params=params, timeout=10)
resp.raise_for_status()
data = resp.json()
ids = data.get("idGroup", {}).get("rxnormId", [])
if ids:
return ids[0]
# Fallback lookup
params = {"name": name}
resp = requests.get(f"{RXNORM_API_BASE}/drugs.json", params=params, timeout=10)
resp.raise_for_status()
data = resp.json()
groups = data.get("drugGroup", {}).get("conceptGroup", [])
for grp in groups:
if grp.get("tty") in ["SBD", "SCD", "GPCK", "BPCK", "IN", "MIN", "PIN"]:
props = grp.get("conceptProperties", [])
if props:
return props[0].get("rxcui")
except Exception:
traceback.print_exc()
return None
@lru_cache(maxsize=128)
def get_openfda_label(
rxcui: Optional[str] = None,
drug_name: Optional[str] = None
) -> Optional[dict]:
"""Fetch OpenFDA drug label by RxCUI or name."""
if not (rxcui or drug_name):
return None
terms = []
if rxcui:
terms.append(f'spl_rxnorm_code:"{rxcui}" OR openfda.rxcui:"{rxcui}"')
if drug_name:
name = drug_name.lower()
terms.append(f'(openfda.brand_name:"{name}" OR openfda.generic_name:"{name}")')
query = " OR ".join(terms)
params = {"search": query, "limit": 1}
try:
resp = requests.get(OPENFDA_API_BASE, params=params, timeout=15)
resp.raise_for_status()
data = resp.json()
results = data.get("results", [])
if results:
return results[0]
except Exception:
traceback.print_exc()
return None
def search_text_list(texts: List[str], terms: List[str]) -> List[str]:
"""Return snippets where any term appears in texts."""
snippets = []
lowers = [t.lower() for t in terms if t]
for txt in texts or []:
if not isinstance(txt, str):
continue
low_txt = txt.lower()
for term in lowers:
idx = low_txt.find(term)
if idx >= 0:
start = max(0, idx - 50)
end = min(len(txt), idx + len(term) + 100)
snippet = txt[start:end]
snippet = re.sub(
f"({re.escape(term)})",
r"**\1**",
snippet,
count=1,
flags=re.IGNORECASE,
)
snippets.append(f"...{snippet}...")
break
return snippets
def parse_bp(bp_str: str) -> Optional[tuple[int, int]]:
"""Parse blood pressure string 'systolic/diastolic'."""
match = re.match(r"(\d{1,3})\s*/\s*(\d{1,3})", bp_str or "")
if match:
return int(match.group(1)), int(match.group(2))
return None
def check_red_flags(patient_data: Dict) -> List[str]:
"""Identify critical red flags from patient data."""
flags = []
if not patient_data:
return flags
symptoms = [s.lower() for s in patient_data.get("hpi", {}).get("symptoms", [])]
vitals = patient_data.get("vitals", {})
history = patient_data.get("pmh", {}).get("conditions", "").lower()
# Symptom-based flags
mapping = {
"chest pain": "Chest Pain reported.",
"shortness of breath": "Shortness of Breath reported.",
"severe headache": "Severe Headache reported.",
"sudden vision loss": "Sudden Vision Loss reported.",
"weakness on one side": "Unilateral Weakness reported (potential stroke).",
"hemoptysis": "Hemoptysis (coughing up blood).",
"syncope": "Syncope (fainting).",
}
for term, desc in mapping.items():
if term in symptoms:
flags.append(f"Red Flag: {desc}")
# Vital sign flags
temp = vitals.get("temp_c")
hr = vitals.get("hr_bpm")
rr = vitals.get("rr_rpm")
spo2 = vitals.get("spo2_percent")
bp = parse_bp(vitals.get("bp_mmhg", ""))
if temp and temp >= 38.5:
flags.append(f"Red Flag: Fever ({temp}°C).")
if hr:
if hr >= 120:
flags.append(f"Red Flag: Tachycardia ({hr} bpm).")
if hr <= 50:
flags.append(f"Red Flag: Bradycardia ({hr} bpm).")
if rr and rr >= 24:
flags.append(f"Red Flag: Tachypnea ({rr} rpm).")
if spo2 and spo2 <= 92:
flags.append(f"Red Flag: Hypoxia ({spo2}%).")
if bp:
sys, dia = bp
if sys >= 180 or dia >= 110:
flags.append(f"Red Flag: Hypertensive Urgency/Emergency (BP: {sys}/{dia} mmHg).")
if sys <= 90 or dia <= 60:
flags.append(f"Red Flag: Hypotension (BP: {sys}/{dia} mmHg).")
# History-based flags
if "history of mi" in history and "chest pain" in symptoms:
flags.append("Red Flag: History of MI with current Chest Pain.")
if "history of dvt/pe" in history and "shortness of breath" in symptoms:
flags.append("Red Flag: History of DVT/PE with current Shortness of Breath.")
return list(set(flags))
def format_patient_data_for_prompt(data: Dict) -> str:
"""Convert patient data dict into a human-readable prompt section."""
if not data:
return "No patient data provided."
sections = []
for key, val in data.items():
title = key.replace("_", " ").title()
if isinstance(val, dict) and any(val.values()):
lines = [f"**{title}:**"]
for subk, subv in val.items():
if subv:
lines.append(f"- {subk.replace('_', ' ').title()}: {subv}")
sections.append("\n".join(lines))
elif isinstance(val, list) and val:
sections.append(f"**{title}:** {', '.join(map(str, val))}")
elif val:
sections.append(f"**{title}:** {val}")
return "\n\n".join(sections)
# --- Tool Schemas & Definitions ---
class LabOrderInput(BaseModel):
test_name: str = Field(...)
reason: str = Field(...)
priority: str = Field("Routine")
class PrescriptionInput(BaseModel):
medication_name: str = Field(...)
dosage: str = Field(...)
route: str = Field(...)
frequency: str = Field(...)
duration: str = Field("As directed")
reason: str = Field(...)
class InteractionCheckInput(BaseModel):
potential_prescription: str = Field(...)
current_medications: Optional[List[str]] = Field(None)
allergies: Optional[List[str]] = Field(None)
class FlagRiskInput(BaseModel):
risk_description: str = Field(...)
urgency: str = Field("High")
@tool("order_lab_test", args_schema=LabOrderInput)
def order_lab_test(test_name: str, reason: str, priority: str = "Routine") -> str:
result = {
"status": "success",
"message": f"Lab Ordered: {test_name} ({priority})",
"details": f"Reason: {reason}"
}
return json.dumps(result)
@tool("prescribe_medication", args_schema=PrescriptionInput)
def prescribe_medication(
medication_name: str,
dosage: str,
route: str,
frequency: str,
duration: str,
reason: str
) -> str:
result = {
"status": "success",
"message": f"Prescription Prepared: {medication_name} {dosage} {route} {frequency}",
"details": f"Duration: {duration}. Reason: {reason}"
}
return json.dumps(result)
@tool("check_drug_interactions", args_schema=InteractionCheckInput)
def check_drug_interactions(
potential_prescription: str,
current_medications: Optional[List[str]] = None,
allergies: Optional[List[str]] = None
) -> str:
warnings: List[str] = []
presc_lower = potential_prescription.lower().strip()
current = [m.lower().strip() for m in (current_medications or [])]
allergy_list = [a.lower().strip() for a in (allergies or [])]
# Normalize and lookup
rxcui = get_rxcui(potential_prescription)
label = get_openfda_label(rxcui=rxcui, drug_name=potential_prescription)
if not rxcui and not label:
warnings.append(f"INFO: Could not identify '{potential_prescription}'.")
# Allergy checks
for alg in allergy_list:
if alg == presc_lower:
warnings.append(f"CRITICAL ALLERGY: Patient allergic to '{alg}'.")
# Additional cross-allergy logic...
# Drug-drug interactions
if rxcui or label:
for med in current:
if med and med != presc_lower:
# interaction search on label sections
interactions = []
if label and label.get("drug_interactions"):
interactions = search_text_list(label["drug_interactions"], [med])
if interactions:
warnings.append(
f"Potential Interaction: '{potential_prescription}' & '{med}'. Snippets: {'; '.join(interactions)}"
)
else:
warnings.append(f"INFO: Skipped interaction check for '{potential_prescription}'.")
status = "warning" if warnings else "clear"
message = (
f"Interaction/Allergy check for '{potential_prescription}': {len(warnings)} issue(s)."
if warnings else
f"No major issues for '{potential_prescription}'."
)
return json.dumps({"status": status, "message": message, "warnings": warnings})
@tool("flag_risk", args_schema=FlagRiskInput)
def flag_risk(risk_description: str, urgency: str) -> str:
return json.dumps({
"status": "flagged",
"message": f"Risk '{risk_description}' flagged with {urgency} urgency."
})
# Initialize search tool and tool list
search_tool = TavilySearchResults(max_results=MAX_SEARCH_RESULTS, name="tavily_search_results")
all_tools = [order_lab_test, prescribe_medication, check_drug_interactions, flag_risk, search_tool]
# --- LangGraph Setup ---
class AgentState(TypedDict):
messages: Annotated[List[Any], operator.add]
patient_data: Optional[Dict]
summary: Optional[str]
interaction_warnings: Optional[List[str]]
# LLM and executor
llm = ChatGroq(temperature=AGENT_TEMPERATURE, model=AGENT_MODEL_NAME)
model_with_tools = llm.bind_tools(all_tools)
tool_executor = ToolExecutor(all_tools)
def agent_node(state: AgentState) -> Dict:
"""Invoke the LLM agent node."""
msgs = state['messages'][:]
if not msgs or not isinstance(msgs[0], SystemMessage):
msgs.insert(0, SystemMessage(content=ClinicalPrompts.SYSTEM_PROMPT))
try:
response = model_with_tools.invoke(msgs)
return {"messages": [response]}
except Exception as e:
traceback.print_exc()
err = AIMessage(content=f"Error: {e}")
return {"messages": [err]}
def tool_node(state: AgentState) -> Dict:
"""Execute any pending tool calls from the last AI message."""
last = state['messages'][-1]
if not isinstance(last, AIMessage) or not getattr(last, 'tool_calls', None):
return {"messages": [], "interaction_warnings": None}
calls = last.tool_calls
# Enforce safety: require interaction check before prescribing
blocked_ids = set()
for call in calls:
if call['name'] == 'prescribe_medication':
# block if no interaction check for this med
med = call['args'].get('medication_name', '').lower()
if not any(
c['name'] == 'check_drug_interactions' and
c['args'].get('potential_prescription', '').lower() == med
for c in calls
):
blocked_ids.add(call['id'])
valid_calls = [c for c in calls if c['id'] not in blocked_ids]
# Augment interaction checks with patient data
for c in valid_calls:
if c['name'] == 'check_drug_interactions':
c['args']['current_medications'] = state.get('patient_data', {}).get('medications', {}).get('current', [])
c['args']['allergies'] = state.get('patient_data', {}).get('allergies', [])
results = []
warnings: List[str] = []
try:
responses = tool_executor.batch(valid_calls, return_exceptions=True)
for call, resp in zip(valid_calls, responses):
if isinstance(resp, Exception):
traceback.print_exc()
content = json.dumps({"status": "error", "message": str(resp)})
else:
content = str(resp)
if call['name'] == 'check_drug_interactions':
data = json.loads(content)
if data.get('warnings'):
warnings.extend(data['warnings'])
results.append(ToolMessage(content=content, tool_call_id=call['id'], name=call['name']))
except Exception as e:
traceback.print_exc()
content = json.dumps({"status": "error", "message": str(e)})
for c in valid_calls:
results.append(ToolMessage(content=content, tool_call_id=c['id'], name=c['name']))
return {"messages": results, "interaction_warnings": warnings or None}
def reflection_node(state: AgentState) -> Dict:
"""Review interaction warnings and adjust plan if needed."""
warnings = state.get('interaction_warnings')
if not warnings:
return {"messages": [], "interaction_warnings": None}
# Find the AI message that triggered the warnings
trigger_id = None
for msg in reversed(state['messages']):
if isinstance(msg, ToolMessage) and msg.name == 'check_drug_interactions':
trigger_id = msg.tool_call_id
break
prompt = (
f"Interaction warnings:\n{json.dumps(warnings, indent=2)}\n"
"Provide a revised therapeutics plan addressing these issues."
)
msgs = [
SystemMessage(content="Safety reflection on drug interactions."),
HumanMessage(content=prompt)
]
try:
resp = llm.invoke(msgs)
return {"messages": [AIMessage(content=resp.content)], "interaction_warnings": None}
except Exception as e:
traceback.print_exc()
return {"messages": [AIMessage(content=f"Reflection error: {e}")], "interaction_warnings": None}
def should_continue(state: AgentState) -> str:
last = state['messages'][-1] if state['messages'] else None
if not isinstance(last, AIMessage):
return 'end_conversation_turn'
if getattr(last, 'tool_calls', None):
return 'continue_tools'
return 'end_conversation_turn'
def after_tools_router(state: AgentState) -> str:
if state.get('interaction_warnings'):
return 'reflect_on_warnings'
return 'continue_to_agent'
class ClinicalAgent:
def __init__(self):
graph = StateGraph(AgentState)
graph.add_node('agent', agent_node)
graph.add_node('tools', tool_node)
graph.add_node('reflection', reflection_node)
graph.set_entry_point('agent')
graph.add_conditional_edges(
'agent', should_continue,
{'continue_tools': 'tools', 'end_conversation_turn': END}
)
graph.add_conditional_edges(
'tools', after_tools_router,
{'reflect_on_warnings': 'reflection', 'continue_to_agent': 'agent'}
)
graph.add_edge('reflection', 'agent')
self.app = graph.compile()
def invoke_turn(self, state: Dict) -> Dict:
try:
result = self.app.invoke(state, {'recursion_limit': 15})
result.setdefault('summary', state.get('summary'))
result.setdefault('interaction_warnings', None)
return result
except Exception as e:
traceback.print_exc()
err = AIMessage(content=f"Critical error: {e}")
return {
'messages': state.get('messages', []) + [err],
'patient_data': state.get('patient_data'),
'summary': state.get('summary'),
'interaction_warnings': None
}