Tutorial 2: Persona Generation

Deep dive into the persona system and customization.

Learning Objectives

• Understand persona dimensions and constraints

• Customize persona configurations

• Control diversity in sampling

Persona Dimensions

Personas are defined by dimensions in input/persona.yaml:

dimensions:
  age_band:
    name: age_band
    is_constraint: true        # Always included
    values:
      - u18
      - 18_24
      - 25_34
      - 35_44
      - 45_60
      - 60p

  role:
    name: role
    is_constraint: true
    values:
      - student
      - engineer
      - data
      - pm
      - designer
      # ... more roles

  communication_style:
    name: communication_style
    is_constraint: false       # Optionally included
    values:
      - casual
      - professional
      - technical
      - creative

Constraint vs Optional

Type	is_constraint	Behavior
Constraint	true	Always sampled
Optional	false	Sampled based on feature_availability_rate

Customizing Dimensions

Add a New Dimension

Edit input/persona.yaml:

dimensions:
  # Add custom dimension
  industry:
    name: industry
    is_constraint: false
    values:
      - technology
      - healthcare
      - finance
      - education
      - manufacturing

Modify Values

dimensions:
  role:
    name: role
    is_constraint: true
    values:
      - student
      - researcher        # Added
      - software_engineer # More specific
      - data_scientist    # Added
      - product_manager

Sampling Configuration

Configure sampling in input/sampling_config.yaml:

sampling:
  feature_availability_rate: 0.7   # 70% of optional dims
  min_features: 10
  max_features: 20
  required_dimensions:
    - query_length_pref            # Always include

diversity:
  enabled: true
  min_hamming_distance: 3          # Minimum difference
  max_retries: 100

Diversity Enforcement

Two personas with Hamming distance = 2:

persona_1 = {'age': '25_34', 'role': 'engineer', 'style': 'casual'}
persona_2 = {'age': '25_34', 'role': 'designer', 'style': 'formal'}
# Differs in: role, style (distance = 2)

If min_hamming_distance = 3, persona_2 would be rejected.

Programmatic Persona Generation

Using PersonaSampler

from src.persona_bank import PersonaBank
from src.sampling import PersonaSampler

# Load dimensions
bank = PersonaBank("input/persona.yaml")
print(f"Dimensions: {list(bank.get_all_dimensions().keys())}")

# Initialize sampler
sampler = PersonaSampler("input/sampling_config.yaml", bank)

# Sample personas
for i in range(5):
    features = sampler.sample_persona()
    print(f"Persona {i+1}: {features}")

Inspect Dimensions

# Get constraint dimensions
constraints = bank.get_constraint_dimensions()
print(f"Constraints: {constraints}")

# Get values for a dimension
roles = bank.get_dimension_values('role')
print(f"Roles: {roles}")

System Prompt Generation

Personas get personalized system prompts:

from src.llm_client import create_llm_client, LLMFormulator

client = create_llm_client(config['api'])
formulator = LLMFormulator(
    client,
    template_path="prompts/persona_to_system_prompt.txt"
)

# Generate prompt from features
features = {'role': 'engineer', 'style': 'casual', 'level': 'advanced'}
prompt = formulator.formulate(features)
print(prompt)

Output:

You are a helpful AI assistant. You are currently helping an advanced-level
engineer who prefers casual communication. Adapt your responses to be:
- Technical but approachable
- Concise and practical
- Friendly in tone

Customize Prompt Template

Edit prompts/persona_to_system_prompt.txt:

You are an AI assistant helping a user with the following profile:

{persona_features}

Guidelines:
- Match the user's communication style: {communication_style}
- Adjust technical depth to: {technical_level}
- Keep responses: {response_length}

Be helpful, accurate, and tailored to this user's needs.

Exercises

1. Add a “hobby” dimension with 5 values

2. Increase min_hamming_distance to 5 and observe sampling behavior

3. Create a custom prompt template for your use case

Next Steps

Continue to Tutorial 3: Query & Interaction to learn about generating conversations.