🧠 Prompt Engineering System

Overview

The Fever Model uses a LangGraph-based architecture where prompt engineering is distributed across specialized nodes. Instead of a single monolithic prompt, we use distinct prompts for specific tasks: Extraction (Receptor), Inquiry, and Recommendation.

Architecture

The system is composed of specialized agents (nodes), each with its own prompt engineering strategy:

1. Receptor Node (Extraction)

Goal: Extract structured data from unstructured user messages into the State.

  • Strategy: One-shot extraction with complex JSON-like structure parsing.
  • Prompt:
    • Contains definitions for all state fields (e.g., patient_age_months, temperature, fever_duration_hours).
    • Instructs the LLM to output only the fields found in the user’s message.
    • Handles normalization (e.g., “dos años” -> “24” months).

2. Inquiry Node (Question Generation)

Goal: Ask the next most important question to complete the clinical checklist.

  • Strategy: Context-aware question generation based on missing fields.
  • Prompt:
    • Receives the list of missing critical and important fields.
    • Instructs the LLM to ask one clear, empathetic question to gather the missing info.
    • Prioritizes urgency (e.g., if age is missing, ask age first).

3. Recommendation Node (Advice)

Goal: Provide clinical advice based on the completed checklist.

  • Strategy: Chain-of-thought reasoning based on guidelines.
  • Prompt:
    • Receives the fully populated State.
    • Includes clinical guidelines (NICE/AAP) for fever management.
    • Generates a structured response: Assessment -> Home Care -> Warning Signs.

Key Prompt Features

State-Driven Context

Instead of relying on conversation history strings, we inject the Structured State into the prompt. This reduces hallucinations and ensures the LLM focuses on what is known vs. unknown.

Safety Rails

  • Urgency Detection: A deterministic Python function (assess_urgency) runs before the LLM to detect red flags (e.g., age < 3 months, temp > 40°C).
  • Red Flag Prompts: If red flags are detected, the system bypasses standard inquiry and triggers the UrgencyRecommendation node with a prompt focused on immediate action (ER/911).

Example: Receptor Prompt 

SYSTEM_PROMPT = """
You are an expert medical data extractor.
Your job is to update the patient state based on the user's message.

OUTPUT FORMAT:
field_name: value

FIELDS:
- patient_age_months: Age in months (convert years to months).
- temperature: Body temperature in Celsius.
- symptoms: Comma-separated list of symptoms.

RULES:
- Only output fields present in the message.
- Normalize values (e.g., "39.5 degrees" -> "39.5").
"""

Example: Inquiry Prompt 

SYSTEM_PROMPT = """
You are a pediatric triage assistant.
Your goal is to complete the clinical checklist.

MISSING DATA:
{missing_fields}

INSTRUCTION:
Ask ONE clear, polite question to get the most important missing field.
Do not give medical advice yet.
"""

Configuration

Prompts are located in:

  • src/fever_routing/nodes/receptor/prompt.py
  • src/fever_routing/nodes/inquiry/prompt.py
  • src/fever_routing/nodes/recommendation/prompt.py