CareCompass Architecture & Implementation

Multi-agent healthcare navigator providing evidence-based wellness insights

CareCompass is a production-ready FastAPI backend that uses advanced AI agents to process healthcare queries through a sophisticated LangGraph workflow. Our system combines cutting-edge AI technology with healthcare domain expertise to deliver reliable, evidence-based guidance.

Technology Stack

FastAPI High-performance web framework for building APIs with automatic documentation

LangGraph Agent workflow orchestration platform for complex AI agent coordination

MongoDB Atlas Cloud-native document database for scalable data persistence

OpenAI API Large language models for intelligent planning and classification

Tavily API Real-time search and content extraction from trusted sources

Multi-Agent Architecture

CareCompass employs a sophisticated multi-agent system where specialized AI agents work together to process healthcare queries through a coordinated workflow:

Intent Agent

Purpose: Classifies and filters healthcare-related queries from general requests

Technology: OpenAI classification models with intelligent keyword fallback mechanisms

Output: Healthcare classification confidence scores and intent categorization

Search Agent

Purpose: Discovers and retrieves relevant healthcare information from trusted sources

Technology: Tavily Search API for production environments, mock data for development

Output: Curated collection of healthcare source URLs and metadata

Extractor Agent

Purpose: Extracts comprehensive content from identified healthcare sources

Technology: Tavily Extract API with intelligent content caching and persistence

Output: Full-text content from verified healthcare authorities and institutions

Validator Agent

Purpose: Evaluates content quality, credibility, and healthcare relevance

Technology: Rule-based scoring algorithms with trusted domain validation

Output: Content credibility scores, quality metrics, and source reliability assessments

Planner Agent

Purpose: Synthesizes information into actionable healthcare guidance

Technology: Template-based deterministic planning or AI-powered dynamic planning

Output: Structured action plans with safety recommendations and clinical guidance

LangGraph Workflow

The workflow orchestrates agents in a sequential pipeline with error handling:

Step 1: Query intake and user validation

Step 2: Intent classification (healthcare vs. non-healthcare)

Step 3: Information search and source discovery

Step 4: Content extraction from trusted sources

Step 5: Content validation and credibility scoring

Step 6: Action plan generation with safety checks

Step 7: Response compilation and persistence

Database Schema

Users Collection

{
  "_id": "user123",
  "runs_count": 5,
  "created_at": "2025-01-15T10:00:00Z",
  "last_run": "2025-01-15T14:30:00Z"
}

Runs Collection

{
  "_id": "uuid-run-id",
  "user_id": "user123",
  "prompt": "What are symptoms of vitamin D deficiency?",
  "status": "completed|running|failed|queued",
  "result": { /* Final response object */ },
  "artifacts": { /* Agent execution details */ },
  "created_at": "2025-01-15T14:30:00Z",
  "updated_at": "2025-01-15T14:30:45Z"
}

Extracts Collection

{
  "_id": "url-hash",
  "url": "https://mayo.edu/article",
  "content": "Full extracted content...",
  "metadata": { /* Content metadata */ },
  "created_at": "2025-01-15T14:30:30Z"
}

Deployment Guide

Environment Variables

# Database Configuration
MONGODB_URI=mongodb+srv://...
MONGO_DB=carecompass

# API Keys
OPENAI_API_KEY=sk-...
TAVILY_API_KEY=tvly-...

# Feature Flags
USE_TAVILY=true
USE_LLM_PLANNER=true

# Optional Configuration
CC_MAX_EXTRACT_URLS=5
CC_PERSIST_EXTRACTS=true

Production Deployment

Platform: AWS Elastic Beanstalk (via Procfile)
Database: MongoDB Atlas (cloud-hosted)
Monitoring: Health checks at /health, /env-check, /db-check
Rate Limiting: One active query per user every 5 minutes
Error Handling: Comprehensive logging with structured context

Safety Features

High-Risk Detection: Automatic identification of medication, dosing, and emergency queries
Trusted Sources: Tier-based credibility scoring for healthcare domains
Clinical Consultation: Mandatory recommendations for high-risk scenarios
Conservative Approach: Safety-first guidance when evidence is uncertain
Content Validation: Multi-layer validation against trusted healthcare sources

Key Metrics

Query processing time: Target <10 seconds
Agent success rates and fallback usage
Evidence quality scores and citation accuracy
Clinical consultation recommendation rates
API error rates and system reliability

API Integration

The backend exposes a RESTful API for healthcare query processing:

Query Submission: POST /v1/query
Result Polling: GET /v1/query/{run_id}/result
History Management: GET/DELETE /v1/history
System Health: GET /health, /env-check, /db-check

CareCompass Backend v1.0.0
Evidence-based healthcare navigation through AI-powered multi-agent workflows
Built with FastAPI, LangGraph, MongoDB, OpenAI, and Tavily API