rstructor: Structured LLM Outputs for Rust

rstructor is a Rust library for extracting structured data from Large Language Models (LLMs) with built-in validation. Define your schemas as Rust structs/enums, and rstructor will handle the rest—generating JSON Schemas, communicating with LLMs, parsing responses, and validating the results.

Think of it as the Rust equivalent of Instructor + Pydantic for Python, bringing the same structured output capabilities to the Rust ecosystem.

✨ Features

• 📝 Type-Safe Definitions: Define data models as standard Rust structs/enums with attributes
• 🔄 JSON Schema Generation: Auto-generates JSON Schema from your Rust types
• ✅ Built-in Validation: Type checking plus custom business rule validation
• 🔌 Multiple LLM Providers: Support for OpenAI, Anthropic, Grok (xAI), and Gemini (Google), with an extensible backend system
• 🧩 Complex Data Structures: Support for nested objects, arrays, optional fields, and deeply nested enums
• 🧠 Schema Fidelity: Heuristic-free JSON Schema generation that preserves nested struct and enum detail
• 🔍 Custom Validation Rules: Add domain-specific validation with automatically detected validate methods
• 🔁 Async API: Fully asynchronous API for efficient operations
• ⚙️ Builder Pattern: Fluent API for configuring LLM clients (temperature retries, timeouts, etc)
• 📊 Feature Flags: Optional backends via feature flags

📦 Installation

Add rstructor to your Cargo.toml:

[dependencies]
rstructor = "0.1.0"
serde = { version = "1.0", features = ["derive"] }
tokio = { version = "1.0", features = ["rt-multi-thread", "macros"] }

🚀 Quick Start

Here's a simple example of extracting structured information about a movie from an LLM:

use rstructor::{Instructor, LLMClient, OpenAIClient, OpenAIModel};
use serde::{Serialize, Deserialize};
use std::env;
use std::time::Duration;

// Define your data model
#[derive(Instructor, Serialize, Deserialize, Debug)]
struct Movie {
    #[llm(description = "Title of the movie")]
    title: String,

    #[llm(description = "Director of the movie")]
    director: String,

    #[llm(description = "Year the movie was released", example = 2010)]
    year: u16,

    #[llm(description = "Brief plot summary")]
    plot: String,
}

#[tokio::main]
async fn main() -> Result<(), Box<dyn std::error::Error>> {
    // Get API key from environment
    let api_key = env::var("OPENAI_API_KEY")?;

    // Create an OpenAI client
    let client = OpenAIClient::new(api_key)?
        .model(OpenAIModel::Gpt4OMini)
        .temperature(0.0)
        .with_timeout(Duration::from_secs(30))  // Optional: set 30 second timeout
        .build();

    // Generate structured information with a simple prompt
    // For production use, prefer generate_struct_with_retry for automatic error recovery
    let movie: Movie = client.generate_struct("Tell me about the movie Inception").await?;

    // Use the structured data
    println!("Title: {}", movie.title);
    println!("Director: {}", movie.director);
    println!("Year: {}", movie.year);
    println!("Plot: {}", movie.plot);

    Ok(())
}

📝 Detailed Examples

Production Example with Automatic Retry

For production use, prefer generate_struct_with_retry which automatically retries on validation errors:

use rstructor::{Instructor, LLMClient, OpenAIClient, OpenAIModel};
use std::time::Duration;
use serde::{Serialize, Deserialize};

#[derive(Instructor, Serialize, Deserialize, Debug)]
#[llm(description = "Information about a movie")]
struct Movie {
    #[llm(description = "Title of the movie")]
    title: String,

    #[llm(description = "Year the movie was released", example = 2010)]
    year: u16,

    #[llm(description = "IMDB rating out of 10", example = 8.5)]
    rating: f32,
}

// Generate with automatic retry (recommended for production)
let movie: Movie = client
    .generate_struct_with_retry::<Movie>(
        "Tell me about Inception",
        Some(3),    // max retries
        Some(true),  // include error feedback in retries
    )
    .await?;

Basic Example with Validation

Add custom validation rules to enforce business logic beyond type checking:

use rstructor::{Instructor, LLMClient, OpenAIClient, OpenAIModel, RStructorError, Result};
use serde::{Serialize, Deserialize};

#[derive(Instructor, Serialize, Deserialize, Debug)]
#[llm(description = "Information about a movie")]
struct Movie {
    #[llm(description = "Title of the movie")]
    title: String,

    #[llm(description = "Year the movie was released", example = 2010)]
    year: u16,

    #[llm(description = "IMDB rating out of 10", example = 8.5)]
    rating: f32,
}

// Add custom validation
impl Movie {
    fn validate(&self) -> Result<()> {
        // Title can't be empty
        if self.title.trim().is_empty() {
            return Err(RStructorError::ValidationError(
                "Movie title cannot be empty".to_string()
            ));
        }

        // Year must be in a reasonable range
        if self.year < 1888 || self.year > 2030 {
            return Err(RStructorError::ValidationError(
                format!("Movie year must be between 1888 and 2030, got {}", self.year)
            ));
        }

        // Rating must be between 0 and 10
        if self.rating < 0.0 || self.rating > 10.0 {
            return Err(RStructorError::ValidationError(
                format!("Rating must be between 0 and 10, got {}", self.rating)
            ));
        }

        Ok(())
    }
}

// The derive macro automatically wires this method into the generated implementation,
// so you won't see `dead_code` warnings even if the method is only called by rstructor.

Complex Nested Structures

rstructor supports complex nested data structures:

use rstructor::{Instructor, LLMClient, OpenAIClient, OpenAIModel};
use std::time::Duration;
use serde::{Serialize, Deserialize};

// Define a nested data model for a recipe
#[derive(Instructor, Serialize, Deserialize, Debug)]
struct Ingredient {
    #[llm(description = "Name of the ingredient", example = "flour")]
    name: String,

    #[llm(description = "Amount of the ingredient", example = 2.5)]
    amount: f32,

    #[llm(description = "Unit of measurement", example = "cups")]
    unit: String,
}

#[derive(Instructor, Serialize, Deserialize, Debug)]
struct Step {
    #[llm(description = "Order number of this step", example = 1)]
    number: u16,

    #[llm(description = "Description of this step",
          example = "Mix the flour and sugar together")]
    description: String,
}

#[derive(Instructor, Serialize, Deserialize, Debug)]
#[llm(description = "A cooking recipe with ingredients and instructions")]
struct Recipe {
    #[llm(description = "Name of the recipe", example = "Chocolate Chip Cookies")]
    name: String,

    #[llm(description = "List of ingredients needed")]
    ingredients: Vec<Ingredient>,

    #[llm(description = "Step-by-step cooking instructions")]
    steps: Vec<Step>,
}

// Usage:
// let recipe: Recipe = client.generate_struct("Give me a recipe for chocolate chip cookies").await?;

Working with Enums

rstructor supports both simple enums and enums with associated data.

Simple Enums

Use enums for categorical data:

use rstructor::{Instructor, LLMClient, AnthropicClient, AnthropicModel};
use serde::{Serialize, Deserialize};

// Define an enum for sentiment analysis
#[derive(Instructor, Serialize, Deserialize, Debug)]
#[llm(description = "The sentiment of a text")]
enum Sentiment {
    #[llm(description = "Positive or favorable sentiment")]
    Positive,

    #[llm(description = "Negative or unfavorable sentiment")]
    Negative,

    #[llm(description = "Neither clearly positive nor negative")]
    Neutral,
}

#[derive(Instructor, Serialize, Deserialize, Debug)]
struct SentimentAnalysis {
    #[llm(description = "The text to analyze")]
    text: String,

    #[llm(description = "The detected sentiment of the text")]
    sentiment: Sentiment,

    #[llm(description = "Confidence score between 0.0 and 1.0",
          example = 0.85)]
    confidence: f32,
}

// Usage:
// let analysis: SentimentAnalysis = client.generate_struct("Analyze the sentiment of: I love this product!").await?;

Enums with Associated Data (Tagged Unions)

rstructor also supports more complex enums with associated data:

use rstructor::{Instructor, SchemaType};
use serde::{Deserialize, Serialize};

// Enum with different types of associated data
#[derive(Instructor, Serialize, Deserialize, Debug)]
enum UserStatus {
    #[llm(description = "The user is online")]
    Online,

    #[llm(description = "The user is offline")]
    Offline,

    #[llm(description = "The user is away with an optional message")]
    Away(String),

    #[llm(description = "The user is busy until a specific time in minutes")]
    Busy(u32),
}

// Using struct variants for more complex associated data
#[derive(Instructor, Serialize, Deserialize, Debug)]
enum PaymentMethod {
    #[llm(description = "Payment with credit card")]
    Card {
        #[llm(description = "Credit card number")]
        number: String,

        #[llm(description = "Expiration date in MM/YY format")]
        expiry: String,
    },

    #[llm(description = "Payment via PayPal account")]
    PayPal(String),

    #[llm(description = "Payment will be made on delivery")]
    CashOnDelivery,
}

// Usage:
// let user_status: UserStatus = client.generate_struct("What's the user's status?").await?;

Nested Enums Across Structs

Enums can be freely nested inside other enums and structs—#[derive(Instructor)] now generates the correct schema without requiring manual SchemaType implementations:

#[derive(Instructor, Serialize, Deserialize, Debug)]
enum TaskState {
    #[llm(description = "Task is pending with a priority level")]
    Pending { priority: Priority },
    #[llm(description = "Task is in progress")]
    InProgress { priority: Priority, status: Status },
    #[llm(description = "Task is completed")]
    Completed { status: Status },
}

#[derive(Instructor, Serialize, Deserialize, Debug)]
struct Task {
    #[llm(description = "Task title")]
    title: String,

    #[llm(description = "Current task state with nested enums")]
    state: TaskState,
}

// Works automatically – TaskState::schema() includes the nested enum structure.

See examples/nested_enum_example.rs for a complete runnable walkthrough that also exercises deserialization of nested enum variants.

When serialized to JSON, these enum variants with data become tagged unions:

// UserStatus::Away("Back in 10 minutes")
{
  "Away": "Back in 10 minutes"
}

// PaymentMethod::Card { number: "4111...", expiry: "12/25" }
{
  "Card": {
    "number": "4111 1111 1111 1111",
    "expiry": "12/25"
  }
}

Working with Custom Types (Dates, UUIDs, etc.)

rstructor provides the CustomTypeSchema trait to handle types that don't have direct JSON representations but need specific schema formats. This is particularly useful for:

• Date/time types (e.g., chrono::DateTime)
• UUIDs (e.g., uuid::Uuid)
• Email addresses
• URLs
• Custom domain-specific types

Basic Implementation

use rstructor::{Instructor, schema::CustomTypeSchema};
use serde::{Serialize, Deserialize};
use chrono::{DateTime, Utc};
use serde_json::json;
use uuid::Uuid;

// Implement CustomTypeSchema for chrono::DateTime<Utc>
impl CustomTypeSchema for DateTime<Utc> {
    fn schema_type() -> &'static str {
        "string"
    }

    fn schema_format() -> Option<&'static str> {
        Some("date-time")
    }

    fn schema_description() -> Option<String> {
        Some("ISO-8601 formatted date and time".to_string())
    }
}

// Implement CustomTypeSchema for UUID
impl CustomTypeSchema for Uuid {
    fn schema_type() -> &'static str {
        "string"
    }

    fn schema_format() -> Option<&'static str> {
        Some("uuid")
    }
}

Usage in Structs

Once implemented, these custom types can be used directly in your structs:

#[derive(Instructor, Serialize, Deserialize, Debug)]
struct Event {
    #[llm(description = "Unique identifier for the event")]
    id: Uuid,

    #[llm(description = "Name of the event")]
    name: String,

    #[llm(description = "When the event starts")]
    start_time: DateTime<Utc>,

    #[llm(description = "When the event ends (optional)")]
    end_time: Option<DateTime<Utc>>,

    #[llm(description = "Recurring event dates")]
    recurring_dates: Vec<DateTime<Utc>>, // Even works with arrays!
}

Advanced Customization

You can add additional schema properties for more complex validation:

impl CustomTypeSchema for EmailAddress {
    fn schema_type() -> &'static str {
        "string"
    }

    fn schema_format() -> Option<&'static str> {
        Some("email")
    }

    fn schema_additional_properties() -> Option<Value> {
        Some(json!({
            "pattern": "^[a-zA-Z0-9._%+-]+@[a-zA-Z0-9.-]+\\.[a-zA-Z]{2,}$",
            "examples": ["user@example.com", "contact@company.org"]
        }))
    }
}

The macro automatically detects these custom types and generates appropriate JSON Schema with format specifications that guide LLMs to produce correctly formatted values. The library includes built-in recognition of common date and UUID types, but you can implement the trait for any custom type.

Configuring Different LLM Providers

Choose between different providers:

// Using OpenAI
let openai_client = OpenAIClient::new(openai_api_key)?
    .model(OpenAIModel::Gpt5)
    .temperature(0.2)
    .max_tokens(1500)
    .with_timeout(Duration::from_secs(60))  // Optional: set 60 second timeout
    .build();

// Using Anthropic
let anthropic_client = AnthropicClient::new(anthropic_api_key)?
    .model(AnthropicModel::ClaudeSonnet4)
    .temperature(0.0)
    .max_tokens(2000)
    .with_timeout(Duration::from_secs(60))  // Optional: set 60 second timeout
    .build();

// Using Grok (xAI) - reads from XAI_API_KEY environment variable
let grok_client = GrokClient::from_env()?  // Reads from XAI_API_KEY env var
    .model(GrokModel::Grok4)
    .temperature(0.0)
    .max_tokens(1500)
    .with_timeout(Duration::from_secs(60))  // Optional: set 60 second timeout
    .build();

// Using Gemini (Google) - reads from GEMINI_API_KEY environment variable
let gemini_client = GeminiClient::from_env()?  // Reads from GEMINI_API_KEY env var
    .model(GeminiModel::Gemini25Flash)
    .temperature(0.0)
    .max_tokens(1500)
    .with_timeout(Duration::from_secs(60))  // Optional: set 60 second timeout
    .build();

Configuring Request Timeouts

All clients (OpenAIClient, AnthropicClient, GrokClient, and GeminiClient) support configurable timeouts for HTTP requests using the builder pattern:

use std::time::Duration;

let client = OpenAIClient::new(api_key)?
    .model(OpenAIModel::Gpt4O)
    .temperature(0.0)
    .with_timeout(Duration::from_secs(30))  // Set 30 second timeout
    .build();

Timeout Behavior:

• The timeout applies to each HTTP request made by the client
• If a request exceeds the timeout, it will return RStructorError::Timeout
• If no timeout is specified, the client uses reqwest's default timeout behavior
• Timeout values are specified as std::time::Duration (e.g., Duration::from_secs(30) or Duration::from_millis(2500))

Example with timeout error handling:

use rstructor::{OpenAIClient, OpenAIModel, RStructorError};
use std::time::Duration;

match client.generate_struct::<Movie>("prompt").await {
    Ok(movie) => println!("Success: {:?}", movie),
    Err(RStructorError::Timeout) => eprintln!("Request timed out"),
    Err(e) => eprintln!("Other error: {}", e),
}

Handling Container-Level Attributes

Add metadata and examples at the container level:

#[derive(Instructor, Serialize, Deserialize, Debug)]
#[llm(description = "Detailed information about a movie",
      title = "MovieDetails",
      examples = [
        ::serde_json::json!({
            "title": "The Matrix",
            "director": "Lana and Lilly Wachowski",
            "year": 1999,
            "genres": ["Sci-Fi", "Action"],
            "rating": 8.7,
            "plot": "A computer hacker learns from mysterious rebels about the true nature of his reality and his role in the war against its controllers."
        })
      ])]
struct Movie {
    // fields...
}

📚 API Reference

Instructor Trait

The Instructor trait is the core of rstructor. It's implemented automatically via the derive macro and provides schema generation and validation:

pub trait Instructor: SchemaType + DeserializeOwned + Serialize {
    fn validate(&self) -> Result<()> {
        Ok(())
    }
}

Override the validate method to add custom validation logic.

CustomTypeSchema Trait

The CustomTypeSchema trait allows you to define JSON Schema representations for types that don't have direct JSON equivalents, like dates and UUIDs:

pub trait CustomTypeSchema {
    /// Returns the JSON Schema type for this custom type
    ///
    /// This is typically "string" for dates, UUIDs, etc.
    fn schema_type() -> &'static str;

    /// Returns the JSON Schema format for this custom type
    ///
    /// Common formats include "date-time", "uuid", "email", etc.
    fn schema_format() -> Option<&'static str> {
        None
    }

    /// Returns a description of this custom type for documentation
    fn schema_description() -> Option<String> {
        None
    }

    /// Returns any additional JSON Schema properties for this type
    ///
    /// This can include patterns, examples, minimum/maximum values, etc.
    fn schema_additional_properties() -> Option<Value> {
        None
    }

    /// Generate a complete JSON Schema object for this type
    fn json_schema() -> Value {
        // Default implementation that combines all properties
        // (You don't normally need to override this)
        let mut schema = json!({
            "type": Self::schema_type(),
        });

        // Add format if present
        if let Some(format) = Self::schema_format() {
            schema.as_object_mut().unwrap()
                .insert("format".to_string(), Value::String(format.to_string()));
        }

        // Add description if present
        if let Some(description) = Self::schema_description() {
            schema.as_object_mut().unwrap()
                .insert("description".to_string(), Value::String(description));
        }

        // Add any additional properties
        if let Some(additional) = Self::schema_additional_properties() {
            // Merge additional properties into the schema
            if let Some(additional_obj) = additional.as_object() {
                for (key, value) in additional_obj {
                    schema.as_object_mut().unwrap()
                        .insert(key.clone(), value.clone());
                }
            }
        }

        schema
    }
}

Implement this trait for custom types like DateTime<Utc> or Uuid to control their JSON Schema representation. Most implementations only need to specify schema_type() and schema_format(), with the remaining methods providing additional schema customization when needed.

LLMClient Trait

The LLMClient trait defines the interface for all LLM providers:

#[async_trait]
pub trait LLMClient {
    /// Generate a structured object from a prompt (single attempt)
    async fn generate_struct<T>(&self, prompt: &str) -> Result<T>
    where
        T: Instructor + DeserializeOwned + Send + 'static;

    /// Generate a structured object with automatic retry on validation errors
    ///
    /// This is the recommended method for production use as it automatically
    /// retries failed generations with error feedback to improve success rates.
    async fn generate_struct_with_retry<T>(
        &self,
        prompt: &str,
        max_retries: Option<usize>,
        include_errors: Option<bool>,
    ) -> Result<T>
    where
        T: Instructor + DeserializeOwned + Send + 'static;

    /// Generate raw text without structure
    async fn generate(&self, prompt: &str) -> Result<String>;
}

Note: For production applications, prefer generate_struct_with_retry over generate_struct as it automatically handles validation errors by retrying with error feedback. This significantly improves success rates with complex schemas.

Supported Attributes

Field Attributes

• description: Text description of the field
• example: A single example value
• examples: Multiple example values

Container Attributes

• description: Text description of the struct or enum
• title: Custom title for the JSON Schema
• examples: Example instances as JSON objects

🔧 Feature Flags

Configure rstructor with feature flags:

[dependencies]
rstructor = { version = "0.1.0", features = ["openai", "anthropic", "grok", "gemini"] }

Available features:

• openai: Include the OpenAI client
• anthropic: Include the Anthropic client
• grok: Include the Grok (xAI) client
• gemini: Include the Gemini (Google) client
• derive: Include the derive macro (enabled by default)
• logging: Enable tracing integration with default subscriber

📊 Logging and Tracing

rstructor includes structured logging via the tracing crate:

use rstructor::logging::{init_logging, LogLevel};

// Initialize with desired level
init_logging(LogLevel::Debug);

// Or use filter strings for granular control
// init_logging_with_filter("rstructor=info,rstructor::backend=trace");

Override with environment variables:

RSTRUCTOR_LOG=debug cargo run

Validation errors, retries, and API interactions are thoroughly logged at appropriate levels.

📋 Examples

See the examples/ directory for complete, working examples:

• structured_movie_info.rs: Basic example of getting movie information with validation
• nested_objects_example.rs: Working with complex nested structures for recipe data
• news_article_categorizer.rs: Using enums for categorization
• enum_with_data_example.rs: Working with enums that have associated data (tagged unions)
• event_planner.rs: Interactive event planning with user input
• weather_example.rs: Simple model with validation demonstration
• validation_example.rs: Demonstrates custom validation without dead code warnings
• custom_type_example.rs: Using custom types like dates and UUIDs with JSON Schema format support
• logging_example.rs: Demonstrates tracing integration with custom log levels
• nested_enum_example.rs: Shows automatic schema generation for nested enums inside structs

▶️ Running the Examples

# Set environment variables
export OPENAI_API_KEY=your_openai_key_here
# or
export ANTHROPIC_API_KEY=your_anthropic_key_here
# or
export XAI_API_KEY=your_xai_key_here
# or
export GEMINI_API_KEY=your_gemini_key_here

# Run examples
cargo run --example structured_movie_info
cargo run --example news_article_categorizer

⚠️ Current Limitations

rstructor currently focuses on single-turn, synchronous structured output generation. The following features are planned but not yet implemented:

• Streaming Responses: Real-time streaming of partial results as they're generated
• Conversation History: Multi-turn conversations with message history (currently only single prompts supported)
• System Messages: Explicit system prompts for role-based interactions
• Response Modes: Different validation strategies (strict, partial, etc.)
• Rate Limiting: Built-in rate limit handling and backoff strategies

🛣️ Roadmap

• Core traits and interfaces
• OpenAI backend implementation
• Anthropic backend implementation
• Procedural macro for deriving Instructor
• Schema generation functionality
• Custom validation capabilities
• Support for nested structures
• Rich validation API with custom domain rules
• Support for enums with associated data (tagged unions)
• Support for custom types (dates, UUIDs, etc.)
• Structured logging and tracing
• Automatic retry with validation error feedback
• Streaming responses
• Conversation history / multi-turn support
• System messages and role-based prompts
• Response modes (strict, partial, retry)
• Rate limiting and backoff strategies
• Support for additional LLM providers
• Integration with web frameworks (Axum, Actix)

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

👥 Contributing

Contributions are welcome! Please feel free to submit a Pull Request.