Skip to content

AI Agents

We've made a number of different agents for experimentation and will continue to make more and depricate others as they're obsolte or uninteresting. You can follow the progress in the Evaluations Learnings Log and watch them on the Leader Board

Currently we have the following:

🚀 Try All Agents Live →

Available Agents

ASHAI Agent

Main conversational agent with intelligent search routing

The primary agent for patient interactions, featuring:

  • Intelligent FAQ Routing - Automatically selects best sources based on query context
  • Pregnancy Context Detection - Identifies pregnancy-related queries for specialized routing
  • Medical Knowledge Search - Searches structured medical databases and research literature
  • LLM-Driven Search Sufficiency - Quality-based search continuation until sufficient information found

Routing Strategy: - Pregnancy queries → Googlesheet FAQ + PubMed - General medical → September Health Library + PubMed - Treatment queries → Always include PubMed for evidence - Complex conditions → PubMed + multiple sources

Strict-Referenced Agent

Evidence-only agent requiring full research sourcing

A specialized wrapper around ASHAI that enforces strict evidence-based requirements:

  • Research-Only Responses - Cannot use AI knowledge, everything must be sourced from research tools
  • Mandatory References - All claims must include complete citations
  • Conservative Approach - Refuses to answer if insufficient sources are found
  • Clinical Decision Support - Ideal for healthcare professionals requiring evidence-based information

Strict-Referenced-After Agent

Answer-first-then-verify agent with transparent validation

Uses an answer-first-then-verify approach for transparent medical guidance:

  • Two-Phase Process - Initial AI answer followed by research validation
  • Self-Correction - Revises responses when research contradicts initial answer
  • Educational Value - Shows validation methodology in thinking process
  • Balanced Approach - Combines AI speed with research accuracy

September Agent

Specialized agent using only the September Health Library

A focused version of ASHAI that exclusively uses the September Health Library for medical information:

  • High-Quality Content - Curated medical content from authoritative sources
  • Section-Based Search - Natural medical sections (symptoms, treatment, causes, etc.)
  • RAG-Based Retrieval - Semantic search over structured medical articles
  • Consistent Format - Standardized medical content presentation

Perplexity Agent

Just uses Perplexity directly to answer these questions (no openai)

Uses Perplexity's search capabilities for current medical information:

  • Real-Time Sources - Accesses current web-based medical information
  • Fallback Option - Useful when specialized databases lack information
  • Current Research - Can find very recent medical developments and studies

Agent Architecture

All agents implement a common interface for consistency:

from agents.ashai.ashai import AshaiAgent
from models import AshaiRequest, Message

# Initialize agent
agent = AshaiAgent()

# Process request
request = AshaiRequest(
    messages=[Message(role='user', content='What are symptoms of diabetes?')],
    profile='Adult patient, no known conditions'
)

response = agent.process_request(request)

Request/Response Models

AshaiRequest Structure: 

{
    "messages": [{"role": "user", "content": "..."}],
    "profile": "Patient profile information",  
    "model": "gpt-4o-mini",
    "with_evaluation": true,
    "tools": ["optional_tool_restrictions"]
}

AshaiResponse Structure: 

{
    "response": "AI agent's response based on search results",
    "searches": [
        {
            "question": "reformulated search query",
            "tool": "search_system_used", 
            "sources": [{"title": "...", "content": "...", "url": "..."}]
        }
    ],
    "evaluation": {"score": 0.95, "reasoning": "..."}  # if requested
}

Intelligent Routing

ASHAI agents use sophisticated routing logic to select the best knowledge sources:

graph TD
    A[User Query] --> B{Pregnancy Related?}
    B -->|Yes| C[Googlesheet FAQ + PubMed]
    B -->|No| D{Complex Medical?}
    D -->|Yes| E[PubMed + Multiple Sources]
    D -->|No| F{General Medical?}
    F -->|Yes| G[September Health + PubMed]
    F -->|No| H[PubMed Only]

    C --> I[Quality Evaluation]
    E --> I
    G --> I  
    H --> I

    I --> J[Evidence-Based Response]

Performance & Monitoring

All agents include comprehensive monitoring:

  • Request Processing Times - Track response latency
  • Search Iteration Counts - Monitor search efficiency
  • Source Retrieval Metrics - Success rates per knowledge source
  • Quality Scores - Response quality evaluation
  • Error Tracking - Graceful error handling and logging

Performance logs are available in ashai_performance.log.

Agent Guidelines

When developing or using agents:

  • Evidence-based responses only - No hallucination allowed
  • Source transparency - Always return sources used
  • Error handling - Graceful degradation when sources fail
  • Performance logging - Track timing and quality metrics
  • Cultural sensitivity - Consider patient demographics and context

API Integration

Agents are automatically integrated with the main FastAPI server:

Endpoint Agent Description
/agent/ashai ASHAI Agent Main conversational agent
/september September Agent September Health Library only

API Documentation

For detailed API documentation, see:

CORS Support

If you are integrating from a web client, use fetch requests with mode: 'cors' and Content-Type: application/json. No custom headers are required in most cases.

  • Default allowed origins include https://whatsapp.turn.io, https://ashai.hackathon.noorahealth.org, and http://localhost:8000.
  • You can override the list with the ALLOWED_CORS_ORIGINS environment variable (comma-separated). See utils/cors.py.

Search Systems Used

Agents leverage multiple search systems:

Evaluation Framework

Quality assessment is provided by: