Cami

A Counselor Agent for Motivational Interviewing through State Inference and Topic Exploration

Overview

Property	Value
Key	cami
Type	LLM-based
Focus	Motivational Interviewing (MI) + topics

Key Features

• Stage inference: Infers the client's readiness stage (Precontemplation, Contemplation, Preparation) each turn
• MI strategies: Selects up to 2 MI-aligned strategies per response
• Topic graph navigation: Explores counseling topics using a configurable topic graph and topic stack
• Multi-candidate generation: Generates several candidate responses and selects the best one

How It Works

1. Init: Loads prompts from data/prompts/therapist/cami.yaml and a topic graph from JSON.
2. Greeting: The first generate_response() call returns a greeting (from the greeting prompt key).
3. Each turn:
   • Infers client stage from conversation history
   • Selects MI strategies
   • Picks the next topic using the topic stack and graph
   • Generates candidate responses, selects and optionally refines one

Usage

CLI

uv run python -m examples.simulate therapist=cami

Python

from patienthub.therapists import get_therapist

therapist = get_therapist(agent_name="cami", lang='en')

response = therapist.generate_response("I don't really want to talk about this.")
print(response.content)

Configuration

Parameter	Type	Default	Description
agent_name	string	"cami"	Therapist identifier
prompt_path	string	data/prompts/therapist/cami.yaml	Path to prompt file
topic_graph	string	data/resources/CAMI/topic_graph.json	Topic graph JSON path
goal	string	"reducing drug use"	Target goal used in prompts
behavior	string	"drug use"	Target behavior used in prompts
model_type	string	"OPENAI"	Model provider key
model_name	string	"gpt-4o"	Model identifier
temperature	float	0.7	Sampling temperature
max_tokens	int	8192	Max response tokens
max_retries	int	3	API retry attempts
lang	string	"en"	Language key for prompt loading

Notes

• Cami relies on prompt templates and a topic graph; it does not load character/persona JSON files.
• The topic graph defines the counseling topics and their relationships.

Use Cases

• MI-style counseling simulations with topic guidance
• Training/evaluation where topic coverage and MI strategy adherence matter