realizar 0.3.2

//! APR Format Loading Example
//!
//! Demonstrates loading and using Aprender's native .apr format,
//! the PRIMARY inference format for the sovereign AI stack.
//!
//! ## Format Priority (Sovereign Stack)
//!
//! 1. `.apr` (PRIMARY) - Aprender native format with CRC32, Ed25519, AES-256-GCM
//! 2. `.gguf` (FALLBACK) - GGUF quantized models (llama.cpp/Ollama)
//! 3. `.safetensors` (FALLBACK) - HuggingFace format
//!
//! ## Usage
//!
//! ```bash
//! cargo run --example apr_loading
//! ```
//!
//! ## Integration
//!
//! For end-to-end evaluation with .apr models, see:
//! - [`single-shot-eval`](https://github.com/paiml/single-shot-eval) - SLM Pareto Frontier Evaluation

use realizar::apr::{
    detect_format, AprHeader, AprMetadata, AprModelType, ModelWeights, HEADER_SIZE, MAGIC,
};

fn main() {
    println!("=== APR Format Loading Demo ===\n");

    // 1. Show format constants
    println!("APR Format Specification:");
    println!("  Magic: {:?} (\"APRN\")", MAGIC);
    println!("  Header Size: {} bytes", HEADER_SIZE);
    println!();

    // 2. Demonstrate model types
    println!("Supported Model Types:");
    let model_types = [
        (AprModelType::LinearRegression, "Linear Regression"),
        (AprModelType::LogisticRegression, "Logistic Regression"),
        (AprModelType::DecisionTree, "Decision Tree"),
        (AprModelType::RandomForest, "Random Forest"),
        (AprModelType::GradientBoosting, "Gradient Boosting"),
        (AprModelType::KMeans, "K-Means Clustering"),
        (AprModelType::Pca, "Principal Component Analysis"),
        (AprModelType::NaiveBayes, "Naive Bayes"),
        (AprModelType::Knn, "K-Nearest Neighbors"),
        (AprModelType::Svm, "Support Vector Machine"),
        (AprModelType::NgramLm, "N-gram Language Model"),
        (AprModelType::Tfidf, "TF-IDF Vectorizer"),
        (
            AprModelType::NeuralSequential,
            "Neural Network (Sequential)",
        ),
        (AprModelType::MixtureOfExperts, "Mixture of Experts"),
    ];

    for (model_type, name) in &model_types {
        println!("  0x{:04X} - {}", model_type.as_u16(), name);
    }
    println!();

    // 3. Create a test model in-memory
    println!("Creating test 2-layer neural network...\n");

    let weights = ModelWeights {
        weights: vec![
            vec![0.5, -0.3, 0.8, -0.1, 0.2, -0.4], // 3x2 input layer
            vec![0.7, -0.5],                       // 2x1 output layer
        ],
        biases: vec![vec![0.1, 0.1], vec![0.0]],
        dimensions: vec![3, 2, 1], // 3 inputs -> 2 hidden -> 1 output
    };

    let metadata = AprMetadata {
        name: Some("demo-mlp".to_string()),
        description: Some("Demo MLP for APR format example".to_string()),
        trained_at: Some("2024-12-09T00:00:00Z".to_string()),
        framework_version: Some("aprender-0.16.0".to_string()),
        custom: std::collections::HashMap::new(),
    };

    // In a real scenario, you would use AprModel::load("model.apr")
    // For this demo, we construct directly
    let model = create_demo_model(weights, metadata);

    println!("Model Info:");
    println!("  Type: {:?}", model.model_type());
    println!("  Parameters: {}", model.num_parameters());
    if let Some(name) = &model.metadata().name {
        println!("  Name: {}", name);
    }
    if let Some(desc) = &model.metadata().description {
        println!("  Description: {}", desc);
    }
    println!();

    // 4. Run inference
    println!("Running Inference:");
    let inputs = [
        vec![1.0, 0.5, 0.2],
        vec![0.0, 1.0, 0.0],
        vec![0.5, 0.5, 0.5],
    ];

    for (i, input) in inputs.iter().enumerate() {
        match model.predict(input) {
            Ok(output) => {
                println!("  Input {}: {:?} -> Output: {:.4}", i + 1, input, output[0]);
            },
            Err(e) => {
                println!("  Input {}: Error - {}", i + 1, e);
            },
        }
    }
    println!();

    // 5. Demonstrate format detection
    println!("Format Detection:");
    let test_paths = [
        ("model.apr", "apr"),
        ("model.gguf", "gguf"),
        ("model.safetensors", "safetensors"),
        ("model.unknown", "unknown"),
    ];

    for (path, expected) in &test_paths {
        let detected = detect_format(path);
        let status = if detected == *expected {
            "OK"
        } else {
            "MISMATCH"
        };
        println!("  {} -> {} [{}]", path, detected, status);
    }
    println!();

    // 6. Header parsing demo
    println!("Header Parsing Demo:");
    let mut header_bytes = vec![0u8; HEADER_SIZE];
    header_bytes[0..4].copy_from_slice(&MAGIC);
    header_bytes[4] = 1; // version major
    header_bytes[5] = 0; // version minor
    header_bytes[6] = 0x05; // flags: compressed + signed
    header_bytes[8..10].copy_from_slice(&AprModelType::NeuralSequential.as_u16().to_le_bytes());
    header_bytes[10..14].copy_from_slice(&100u32.to_le_bytes()); // metadata_len
    header_bytes[14..18].copy_from_slice(&1000u32.to_le_bytes()); // payload_len

    match AprHeader::from_bytes(&header_bytes) {
        Ok(header) => {
            println!("  Version: {}.{}", header.version.0, header.version.1);
            println!("  Model Type: {:?}", header.model_type);
            println!("  Compressed: {}", header.is_compressed());
            println!("  Encrypted: {}", header.is_encrypted());
            println!("  Signed: {}", header.is_signed());
            println!("  Metadata Length: {} bytes", header.metadata_len);
            println!("  Payload Length: {} bytes", header.payload_len);
        },
        Err(e) => {
            println!("  Header parse error: {}", e);
        },
    }
    println!();

    // 7. Batch inference demo
    println!("Batch Inference:");
    let batch_inputs: Vec<Vec<f32>> = vec![
        vec![1.0, 0.0, 0.0],
        vec![0.0, 1.0, 0.0],
        vec![0.0, 0.0, 1.0],
        vec![1.0, 1.0, 1.0],
    ];

    match model.predict_batch(&batch_inputs) {
        Ok(outputs) => {
            println!("  Batch size: {}", batch_inputs.len());
            for (i, output) in outputs.iter().enumerate() {
                println!("  Sample {}: {:.4}", i + 1, output[0]);
            }
        },
        Err(e) => {
            println!("  Batch error: {}", e);
        },
    }
    println!();

    println!("=== APR Demo Complete! ===");
    println!();
    println!("For production use with .apr models:");
    println!("  let model = AprModel::load(\"path/to/model.apr\")?;");
    println!("  let output = model.predict(&input)?;");
    println!();
    println!("For SLM evaluation with Pareto frontier analysis:");
    println!("  See: https://github.com/paiml/single-shot-eval");
}

/// Create a demo model (helper for the example)
/// In production, use AprModel::load() instead
fn create_demo_model(weights: ModelWeights, metadata: AprMetadata) -> DemoAprModel {
    DemoAprModel {
        model_type: AprModelType::NeuralSequential,
        weights,
        metadata,
    }
}

/// Demo wrapper that mirrors AprModel's public interface
struct DemoAprModel {
    model_type: AprModelType,
    weights: ModelWeights,
    metadata: AprMetadata,
}

impl DemoAprModel {
    fn model_type(&self) -> AprModelType {
        self.model_type
    }

    fn metadata(&self) -> &AprMetadata {
        &self.metadata
    }

    fn num_parameters(&self) -> usize {
        let weight_params: usize = self.weights.weights.iter().map(Vec::len).sum();
        let bias_params: usize = self.weights.biases.iter().map(Vec::len).sum();
        weight_params + bias_params
    }

    fn predict(&self, input: &[f32]) -> Result<Vec<f32>, String> {
        if self.weights.dimensions.is_empty() {
            return Err("Model has no layers".to_string());
        }

        let expected_input_dim = self.weights.dimensions[0];
        if input.len() != expected_input_dim {
            return Err(format!(
                "Input dimension mismatch: expected {}, got {}",
                expected_input_dim,
                input.len()
            ));
        }

        let mut current = input.to_vec();

        for (i, (weights, biases)) in self
            .weights
            .weights
            .iter()
            .zip(self.weights.biases.iter())
            .enumerate()
        {
            let in_dim = self.weights.dimensions[i];
            let out_dim = self.weights.dimensions[i + 1];

            let mut output = vec![0.0; out_dim];

            for (j, out_val) in output.iter_mut().enumerate() {
                let mut sum = biases.get(j).copied().unwrap_or(0.0);
                for (k, &in_val) in current.iter().enumerate() {
                    let weight_idx = j * in_dim + k;
                    if let Some(&w) = weights.get(weight_idx) {
                        sum += in_val * w;
                    }
                }
                *out_val = sum;
            }

            // ReLU for hidden layers
            if i < self.weights.weights.len() - 1 {
                for val in &mut output {
                    *val = val.max(0.0);
                }
            }

            current = output;
        }

        Ok(current)
    }

    fn predict_batch(&self, inputs: &[Vec<f32>]) -> Result<Vec<Vec<f32>>, String> {
        inputs.iter().map(|input| self.predict(input)).collect()
    }
}