Paladin Configuration Guide

This guide covers how to configure Paladin agents for optimal performance, from basic setup to advanced tuning.

Table of Contents

• Basic Configuration
• System Prompt Best Practices
• Model Selection
• Temperature and Sampling
• Stop Words and Termination
• Timeout and Retry Settings
• Advanced Configuration

Basic Configuration

Minimal Setup

use paladin::prelude::*;

let paladin = PaladinBuilder::new(llm_adapter)
    .name("Assistant")
    .system_prompt("You are a helpful assistant.")
    .build()?;

Common Configuration

let paladin = PaladinBuilder::new(llm_adapter)
    .name("DataAnalyst")
    .system_prompt("You are an expert data analyst. Provide clear, data-driven insights.")
    .model("gpt-4")
    .temperature(0.7)
    .max_loops(5)
    .timeout_seconds(120)
    .build()?;

Full Configuration

let paladin = PaladinBuilder::new(llm_adapter)
    .name("ResearchAssistant")
    .system_prompt("You are a research assistant specializing in academic papers.")
    .user_name("Researcher")
    .model("gpt-4-turbo")
    .temperature(0.8)
    .max_loops(10)
    .add_stop_word("END").add_stop_word("STOP").add_stop_word("FINAL_ANSWER")
    .timeout_seconds(300)
    .retry_attempts(3)
    .with_garrison(garrison)
    .add_armament(search_tool)
    .add_armament(calculator_tool)
    .build()?;

System Prompt Best Practices

The system prompt defines your Paladin's behavior and capabilities. Follow these best practices:

1. Be Specific About Role

❌ Vague:

.system_prompt("You are helpful.")

✅ Specific:

.system_prompt("You are a senior software engineer specializing in Rust. \
                You provide code reviews focused on safety, performance, and idiomatic patterns.")

2. Define Output Format

.system_prompt("You are a JSON API. Always respond with valid JSON. \
                Structure: {\"status\": \"success|error\", \"data\": {...}, \"message\": \"...\"}  \
                Never include markdown code blocks or explanations outside the JSON.")

3. Set Boundaries

.system_prompt("You are a customer support agent for TechCorp. \
                - Only answer questions about our products and services \
                - Escalate billing questions to the finance team \
                - Do not provide medical, legal, or financial advice \
                - Be polite and professional at all times")

4. Include Examples (Few-Shot)

.system_prompt("You categorize customer feedback as: FEATURE_REQUEST, BUG_REPORT, or PRAISE. \
                \
                Examples: \
                Input: 'The app crashes when I upload large files' \
                Output: BUG_REPORT \
                \
                Input: 'It would be great to have dark mode' \
                Output: FEATURE_REQUEST \
                \
                Input: 'Love the new design!' \
                Output: PRAISE")

5. Specify Tone and Style

.system_prompt("You are a technical writer creating documentation for developers. \
                - Use clear, concise language \
                - Prefer active voice \
                - Include code examples \
                - Target audience: junior to mid-level developers \
                - Avoid jargon unless necessary")

Model Selection

Choose the right model for your use case:

OpenAI Models

// GPT-4 Turbo - Best for complex reasoning
.model("gpt-4-turbo")  // Latest turbo model
.model("gpt-4")        // Standard GPT-4

// GPT-3.5 - Fast and cost-effective
.model("gpt-3.5-turbo")  // Recommended for most tasks

When to use:

• GPT-4: Complex reasoning, code generation, detailed analysis
• GPT-3.5: Simple queries, classification, summarization

DeepSeek Models

// DeepSeek Chat - Strong coding capabilities
.model("deepseek-chat")

// DeepSeek Coder - Specialized for code
.model("deepseek-coder")

When to use:

• deepseek-chat: General purpose, good for multi-turn conversations
• deepseek-coder: Code generation, technical documentation

Anthropic Models

// Claude 3 Family
.model("claude-3-opus")    // Most capable
.model("claude-3-sonnet")  // Balanced
.model("claude-3-haiku")   // Fastest

When to use:

• Opus: Complex analysis, long documents, creative writing
• Sonnet: General purpose, good balance of speed and quality
• Haiku: Fast responses, simple queries, high throughput

Model Comparison

Temperature and Sampling

Temperature controls randomness in responses:

Temperature Scale

// 0.0 - Deterministic, focused (best for factual tasks)
.temperature(0.0)

// 0.3-0.5 - Slightly varied (good for classification)
.temperature(0.4)

// 0.7 - Balanced (general purpose)
.temperature(0.7)

// 0.9-1.0 - Creative, diverse (brainstorming, creative writing)
.temperature(0.9)

// >1.0 - Very random (experimental, not recommended)
.temperature(1.2)

Use Cases by Temperature

Example: Task-Specific Configuration

// Code Review - Deterministic
let code_reviewer = PaladinBuilder::new(llm_adapter)
    .system_prompt("Review Rust code for safety and best practices.")
    .temperature(0.2)
    .build()?;

// Content Writer - Creative
let writer = PaladinBuilder::new(llm_adapter)
    .system_prompt("Write engaging blog posts about technology.")
    .temperature(0.9)
    .build()?;

// Customer Support - Balanced
let support = PaladinBuilder::new(llm_adapter)
    .system_prompt("Help customers with product questions.")
    .temperature(0.7)
    .build()?;

Stop Words and Termination

Control when a Paladin stops generating:

Basic Stop Words

let paladin = PaladinBuilder::new(llm_adapter)
    .add_stop_word("END").add_stop_word("STOP").add_stop_word("###")
    .build()?;

Use Cases

1. Structured Output

// Stop at delimiter for parsing
.system_prompt("Generate a list of items. End with '---'")
.add_stop_word("---")

2. Multi-Step Reasoning

// Stop when final answer is reached
.system_prompt("Think step by step. When done, output FINAL_ANSWER: <answer>")
.add_stop_word("FINAL_ANSWER:")

3. Dialog Systems

// Stop at turn boundaries
.system_prompt("You are user A in a conversation. End each turn with [END_TURN]")
.add_stop_word("[END_TURN]")

Max Loops

Prevent infinite reasoning loops:

// Default: 3 loops
.max_loops(3)

// For simple tasks: 1 loop
.max_loops(1)

// For complex reasoning: 10+ loops
.max_loops(15)

What is a loop? A loop is one reasoning cycle: prompt → LLM → response → (optional tool calls) → repeat.

Timeout and Retry Settings

Timeout Configuration

use std::time::Duration;

let paladin = PaladinBuilder::new(llm_adapter)
    .timeout_seconds(60)  // 60 second timeout
    .build()?;

Recommended Timeouts:

• Simple queries: 30 seconds
• Complex reasoning: 120 seconds
• With tool calls: 300 seconds

Retry Configuration

let paladin = PaladinBuilder::new(llm_adapter)
    .retry_attempts(3)                        // Retry up to 3 times
    .build()?;

Error Handling

match paladin.execute(input).await {
    Ok(response) => println!("Success: {}", response.content),
    Err(PaladinError::Timeout(secs)) => {
        eprintln!("Request timed out after {} seconds", secs);
        // Increase timeout or simplify prompt
    }
    Err(PaladinError::LlmError(msg)) => {
        eprintln!("LLM error: {}", msg);
        // Check API key, rate limits, model availability
    }
    Err(PaladinError::MaxLoopsExceeded) => {
        eprintln!("Max reasoning loops exceeded");
        // Increase max_loops or refine system prompt
    }
    Err(e) => eprintln!("Other error: {}", e),
}

Advanced Configuration

Configuration from File

use paladin::config::ApplicationSettings;

let config = ApplicationSettings::load_from("config.yml")?;
let paladin = PaladinBuilder::from_config(&config.paladin)?;

config.yml:

paladin:
  name: "Assistant"
  system_prompt: "You are a helpful assistant."
  model: "gpt-4"
  temperature: 0.7
  max_loops: 5
  timeout_seconds: 120
  retry_attempts: 3
  stop_words:
    - "END"
    - "STOP"

Environment-Based Configuration

let model = std::env::var("PALADIN_MODEL").unwrap_or("gpt-3.5-turbo".to_string());
let temperature = std::env::var("PALADIN_TEMPERATURE")
    .ok()
    .and_then(|s| s.parse::<f32>().ok())
    .unwrap_or(0.7);

let paladin = PaladinBuilder::new(llm_adapter)
    .model(&model)
    .temperature(temperature)
    .build()?;

Dynamic Configuration

struct PaladinFactory;

impl PaladinFactory {
    fn create_for_task(task_type: &str, llm_adapter: Arc<dyn LlmPort>) -> Result<Paladin> {
        match task_type {
            "code_review" => Self::create_code_reviewer(llm_adapter),
            "creative_writing" => Self::create_writer(llm_adapter),
            "data_analysis" => Self::create_analyst(llm_adapter),
            _ => Self::create_default(llm_adapter),
        }
    }

    fn create_code_reviewer(llm_adapter: Arc<dyn LlmPort>) -> Result<Paladin> {
        PaladinBuilder::new(llm_adapter)
            .system_prompt("Expert Rust code reviewer")
            .temperature(0.2)
            .model("gpt-4")
            .build()
    }

    // ... other factory methods
}

Configuration Validation

let paladin = PaladinBuilder::new(llm_adapter)
    .temperature(0.7)
    .build()?;  // Validates configuration

// Manual validation
if let Err(e) = paladin.validate() {
    eprintln!("Invalid configuration: {}", e);
}

Configuration Checklist

Before deploying a Paladin, verify:

• System prompt is clear and specific
• Appropriate model selected for task
• Temperature suitable for use case (0.2 for factual, 0.9 for creative)
• Max loops set appropriately (1-3 for simple, 10+ for complex)
• Timeout configured (30-300 seconds)
• Retry logic in place for production
• Stop words defined if needed
• Error handling implemented
• Configuration tested with sample inputs

Performance Tuning

For Throughput

// Fast model, simple prompts
let paladin = PaladinBuilder::new(llm_adapter)
    .model("gpt-3.5-turbo")
    .temperature(0.7)
    .max_loops(1)
    .timeout_seconds(30)
    .build()?;

For Quality

// Best model, detailed prompts
let paladin = PaladinBuilder::new(llm_adapter)
    .model("gpt-4")
    .temperature(0.5)
    .max_loops(10)
    .timeout_seconds(300)
    .build()?;

For Cost Efficiency

// Cheaper model, efficient prompts
let paladin = PaladinBuilder::new(llm_adapter)
    .model("deepseek-chat")
    .temperature(0.7)
    .max_loops(3)
    .build()?;

Next Steps

• Battalion Patterns - Multi-agent orchestration
• Tool Integration - Add capabilities with Arsenal
• Memory Management - Use Garrison for context
• Examples - See configuration in action

Related Documentation

• Quickstart Guide
• API Reference
• Performance Tuning