medchat / medical_chatbot.py
vihashini-18
i
0a5c991
"""
Medical Chatbot using Gemini 1.5 Flash with citation and confidence scoring
"""
import google.generativeai as genai
from google.generativeai import types
from typing import List, Dict, Any
from config import GOOGLE_API_KEY, LLM_MODEL, TOP_K, SIMILARITY_THRESHOLD
from embedding_service import EmbeddingService
class MedicalChatbot:
def __init__(self, embedding_service: EmbeddingService):
"""Initialize the medical chatbot"""
self.embedding_service = embedding_service
# Configure Gemini
genai.configure(api_key=GOOGLE_API_KEY)
# Try available model names
model_attempts = [
"models/gemini-2.5-flash", # Fast and efficient
"models/gemini-2.0-flash", # Alternative fast model
"models/gemini-2.5-pro", # More capable
"models/gemini-flash-latest",
"models/gemini-pro-latest",
]
self.model = None
for model_name in model_attempts:
try:
self.model = genai.GenerativeModel(model_name)
# Test if it actually works
test_response = self.model.generate_content("test")
print(f"✓ Successfully initialized model: {model_name}")
break
except Exception as e:
print(f"✗ Failed to initialize {model_name}: {str(e)[:80]}")
continue
if self.model is None:
raise Exception("Could not initialize any Gemini model. Please check your API key and model availability.")
# System prompt for medical chatbot
self.system_prompt = """You are a medical information assistant. Based ONLY on the provided medical context, answer the user's question accurately and concisely.
IMPORTANT RULES:
1. Answer ONLY using information from the provided context below
2. DO NOT make up or guess information
3. If the context doesn't contain enough information, say "Based on the available information..."
4. Be accurate and factual
5. Keep answers concise and clear
6. At the end, add a disclaimer: "⚠️ This is not medical advice. Consult healthcare professionals."
"""
def calculate_confidence_score(self, similarity_scores: List[float]) -> tuple:
"""Calculate confidence score based on similarity scores"""
if not similarity_scores:
return "Low", 0.0
avg_score = sum(similarity_scores) / len(similarity_scores)
max_score = max(similarity_scores)
# Confidence based on best match
if max_score >= 0.85:
return "High", max_score
elif max_score >= 0.65:
return "Medium", max_score
else:
return "Low", max_score
def format_context_with_citations(self, results: List[Dict[str, Any]]) -> str:
"""Format retrieved context with citations"""
context_parts = []
citation_map = {}
for idx, result in enumerate(results):
metadata = result.metadata
score = result.score
text = metadata.get('text', '')
citation_id = f"[Source {idx + 1}]"
citation_map[f"Source_{idx + 1}"] = {
'id': citation_id,
'text': text[:300] + "..." if len(text) > 300 else text,
'source': metadata.get('source', 'unknown'),
'similarity_score': round(score, 3),
'metadata': metadata
}
# Format the context more clearly
context_parts.append(f"{citation_id}\n{text}\n")
return "".join(context_parts), citation_map
def generate_response(self, user_query: str) -> Dict[str, Any]:
"""Generate response to user query with citations and confidence"""
# Check if query is medical-related
is_medical_query = self.is_medical_related(user_query)
if not is_medical_query:
return {
'response': "I'm a medical assistant. Please ask me medical or health-related questions only.",
'confidence': "N/A",
'confidence_score': 0.0,
'sources': [],
'citations': {}
}
# Search for relevant documents
results = self.embedding_service.search(user_query, top_k=TOP_K)
if not results.matches:
return {
'response': "I couldn't find relevant medical information for your query. Please consult with a healthcare professional for accurate medical advice.",
'confidence': "Low",
'confidence_score': 0.0,
'sources': [],
'citations': {}
}
# Filter results by similarity threshold
filtered_results = [
r for r in results.matches
if r.score >= SIMILARITY_THRESHOLD
]
if not filtered_results:
return {
'response': "I couldn't find enough reliable information for your query. Please consult with a healthcare professional.",
'confidence': "Low",
'confidence_score': 0.0,
'sources': [],
'citations': {}
}
# Format context with citations
context, citation_map = self.format_context_with_citations(filtered_results)
# Generate response using Gemini
prompt = f"""{self.system_prompt}
MEDICAL CONTEXT FROM DATABASE:
{context}
USER QUESTION: {user_query}
INSTRUCTIONS:
Based on the medical context above, provide a helpful answer to the user's question.
- Use information from the context when available
- If the context has relevant but not exact information, explain what you found
- Be clear and helpful
- End with: "⚠️ This is not medical advice. Consult healthcare professionals."
Answer the question:"""
try:
response = self.model.generate_content(
prompt,
generation_config={
"temperature": 0.3, # Lower temperature for more factual responses
"top_p": 0.8,
"top_k": 40,
"max_output_tokens": 500,
}
)
answer = response.text
except Exception as e:
answer = f"Error generating response: {str(e)}"
print(f"DEBUG: Model error: {e}")
print(f"DEBUG: Model object: {self.model}")
# Calculate confidence
similarity_scores = [r.score for r in filtered_results]
confidence_level, confidence_score = self.calculate_confidence_score(similarity_scores)
return {
'response': answer,
'confidence': confidence_level,
'confidence_score': confidence_score,
'sources': [r.metadata.get('source', 'unknown') for r in filtered_results],
'citations': citation_map
}
def is_medical_related(self, query: str) -> bool:
"""Check if query is medical-related - very permissive"""
query_lower = query.lower()
# Comprehensive medical keywords
medical_keywords = [
'health', 'medical', 'disease', 'symptom', 'treatment', 'diagnosis',
'medicine', 'patient', 'doctor', 'hospital', 'therapy', 'condition',
'illness', 'sick', 'pain', 'cure', 'medication', 'physician',
'nurse', 'clinical', 'healthcare', 'surgery', 'cure', 'heal',
'blood', 'heart', 'lung', 'brain', 'cancer', 'diabetes', 'covid',
'vaccine', 'pandemic', 'infection', 'fever', 'cough', 'ache',
'eye', 'vision', 'irritation', 'red', 'tear', 'dry', 'irritated',
'head', 'headache', 'stomach', 'nausea', 'dizzy', 'tired',
'chest', 'breathing', 'breath', 'wheeze', 'nose', 'runny',
'ear', 'throat', 'sore', 'inflam', 'swell', 'burn', 'itch',
'suffering', 'problem', 'issue', 'hurt', 'injury', 'wound'
]
# Accept any query that contains medical keywords or looks like a medical concern
has_medical_keyword = any(keyword in query_lower for keyword in medical_keywords)
# Also accept questions with medical-sounding patterns
medical_patterns = [
'i have', 'i am suffering', 'i feel', 'why do i', 'what should i',
'why is', 'how to', 'how can i', 'what causes'
]
has_medical_pattern = any(pattern in query_lower for pattern in medical_patterns)
# Be permissive - if it sounds like a medical concern, accept it
return has_medical_keyword or has_medical_pattern