oxirs_vec/

adaptive_compression.rs

//! Adaptive compression techniques for vector data
//!
//! This module provides intelligent compression strategies that adapt to data characteristics,
//! optimizing compression ratio and decompression speed based on vector patterns and usage.

use crate::{
    compression::{create_compressor, CompressionMethod, VectorCompressor},
    Vector, VectorError,
};
use anyhow::Result;
use std::collections::HashMap;
use std::sync::{Arc, RwLock};
use std::time::{Duration, Instant};

// Random functionality now provided by scirs2-core

/// Context information for compression decisions
#[derive(Debug, Clone)]
pub struct CompressionContext {
    pub domain: VectorDomain,
    pub access_frequency: AccessFrequency,
    pub quality_requirement: QualityRequirement,
    pub resource_constraints: ResourceConstraints,
    pub temporal_patterns: TemporalPatterns,
}

/// Vector domain types for domain-specific optimization
#[derive(Debug, Clone, PartialEq, Eq, Hash)]
pub enum VectorDomain {
    TextEmbeddings,
    ImageFeatures,
    AudioFeatures,
    KnowledgeGraph,
    TimeSeriesData,
    Unknown,
}

/// Access frequency patterns
#[derive(Debug, Clone)]
pub enum AccessFrequency {
    VeryHigh, // Accessed multiple times per second
    High,     // Accessed multiple times per minute
    Moderate, // Accessed multiple times per hour
    Low,      // Accessed daily
    Archive,  // Rarely accessed
}

/// Quality requirements for different use cases
#[derive(Debug, Clone)]
pub enum QualityRequirement {
    Lossless,    // No quality loss acceptable
    HighQuality, // Minimal quality loss (>99% accuracy)
    Balanced,    // Moderate quality loss (>95% accuracy)
    Compressed,  // Higher compression priority (>90% accuracy)
    Aggressive,  // Maximum compression (<90% accuracy)
}

/// Resource constraints for compression decisions
#[derive(Debug, Clone)]
pub struct ResourceConstraints {
    pub cpu_usage_limit: f32,    // 0.0 to 1.0
    pub memory_usage_limit: f32, // 0.0 to 1.0
    pub compression_time_limit: Duration,
    pub decompression_time_limit: Duration,
}

/// Temporal patterns for time-aware compression
#[derive(Debug, Clone)]
pub struct TemporalPatterns {
    pub time_of_day_factor: f32, // Compression aggressiveness based on time
    pub load_factor: f32,        // Current system load
    pub seasonal_factor: f32,    // Long-term usage patterns
}

/// Enhanced statistics about vector data characteristics
#[derive(Debug, Clone)]
pub struct VectorStats {
    pub dimensions: usize,
    pub mean: f32,
    pub std_dev: f32,
    pub min_val: f32,
    pub max_val: f32,
    pub entropy: f32,
    pub sparsity: f32,                 // Fraction of near-zero values
    pub correlation: f32,              // Average correlation between dimensions
    pub intrinsic_dimension: f32,      // Estimated intrinsic dimensionality
    pub clustering_tendency: f32,      // Hopkins statistic
    pub temporal_stability: f32,       // Stability over time
    pub domain_affinity: VectorDomain, // Detected domain type
}

impl Default for CompressionContext {
    fn default() -> Self {
        Self {
            domain: VectorDomain::Unknown,
            access_frequency: AccessFrequency::Moderate,
            quality_requirement: QualityRequirement::Balanced,
            resource_constraints: ResourceConstraints {
                cpu_usage_limit: 0.7,
                memory_usage_limit: 0.8,
                compression_time_limit: Duration::from_millis(100),
                decompression_time_limit: Duration::from_millis(50),
            },
            temporal_patterns: TemporalPatterns {
                time_of_day_factor: 1.0,
                load_factor: 1.0,
                seasonal_factor: 1.0,
            },
        }
    }
}

impl VectorStats {
    /// Calculate enhanced statistics for a vector with context
    pub fn from_vector(vector: &Vector) -> Result<Self, VectorError> {
        Self::from_vector_with_context(vector, &CompressionContext::default())
    }

    /// Calculate statistics for a vector with compression context
    pub fn from_vector_with_context(
        vector: &Vector,
        context: &CompressionContext,
    ) -> Result<Self, VectorError> {
        let values = vector.as_f32();
        let n = values.len();

        if n == 0 {
            return Err(VectorError::InvalidDimensions("Empty vector".to_string()));
        }

        // Basic statistics
        let sum: f32 = values.iter().sum();
        let mean = sum / n as f32;

        let variance: f32 = values.iter().map(|x| (x - mean).powi(2)).sum::<f32>() / n as f32;
        let std_dev = variance.sqrt();

        let min_val = values.iter().fold(f32::INFINITY, |a, &b| a.min(b));
        let max_val = values.iter().fold(f32::NEG_INFINITY, |a, &b| a.max(b));

        // Enhanced entropy estimation with domain-specific bins
        let bin_count = match context.domain {
            VectorDomain::TextEmbeddings => 128,
            VectorDomain::ImageFeatures => 256,
            VectorDomain::KnowledgeGraph => 64,
            _ => 256,
        };

        let mut histogram = vec![0u32; bin_count];
        let range = max_val - min_val;
        if range > 0.0 {
            for val in &values {
                let bucket = ((val - min_val) / range * (bin_count - 1) as f32)
                    .clamp(0.0, (bin_count - 1) as f32) as usize;
                histogram[bucket] += 1;
            }
        }

        let entropy = histogram
            .iter()
            .filter(|&&count| count > 0)
            .map(|&count| {
                let p = count as f32 / n as f32;
                -p * p.log2()
            })
            .sum();

        // Sparsity (fraction of values near zero)
        let threshold = std_dev * 0.1;
        let sparse_count = values.iter().filter(|&&x| x.abs() < threshold).count();
        let sparsity = sparse_count as f32 / n as f32;

        // Enhanced correlation analysis
        let correlation = Self::calculate_enhanced_correlation(&values);

        // Intrinsic dimensionality estimation using correlation dimension
        let intrinsic_dimension = Self::estimate_intrinsic_dimension(&values);

        // Clustering tendency using Hopkins statistic
        let clustering_tendency = Self::calculate_hopkins_statistic(&values);

        // Temporal stability (placeholder for now)
        let temporal_stability = 1.0;

        // Domain detection based on statistical patterns
        let domain_affinity = Self::detect_domain(&values, entropy, sparsity, correlation);

        Ok(VectorStats {
            dimensions: n,
            mean,
            std_dev,
            min_val,
            max_val,
            entropy,
            sparsity,
            correlation,
            intrinsic_dimension,
            clustering_tendency,
            temporal_stability,
            domain_affinity,
        })
    }

    /// Enhanced correlation analysis with multiple window sizes
    fn calculate_enhanced_correlation(values: &[f32]) -> f32 {
        let n = values.len();
        if n <= 1 {
            return 0.0;
        }

        let window_sizes = [5, 10, 20].iter().map(|&w| w.min(n / 2).max(2));
        let mut total_corr = 0.0;
        let mut total_count = 0;

        for window_size in window_sizes {
            if window_size >= n {
                continue;
            }

            for i in 0..(n - window_size) {
                let window1 = &values[i..i + window_size];
                let window2 = &values[i + 1..i + window_size + 1];

                let mean1: f32 = window1.iter().sum::<f32>() / window_size as f32;
                let mean2: f32 = window2.iter().sum::<f32>() / window_size as f32;

                let covariance: f32 = window1
                    .iter()
                    .zip(window2)
                    .map(|(a, b)| (a - mean1) * (b - mean2))
                    .sum();
                let var1: f32 = window1.iter().map(|x| (x - mean1).powi(2)).sum();
                let var2: f32 = window2.iter().map(|x| (x - mean2).powi(2)).sum();

                if var1 > 0.0 && var2 > 0.0 {
                    let corr = covariance / (var1.sqrt() * var2.sqrt());
                    total_corr += corr.abs();
                    total_count += 1;
                }
            }
        }

        if total_count > 0 {
            total_corr / total_count as f32
        } else {
            0.0
        }
    }

    /// Estimate intrinsic dimensionality using correlation dimension method
    fn estimate_intrinsic_dimension(values: &[f32]) -> f32 {
        let n = values.len();
        if n < 10 {
            return n as f32;
        }

        // Sample points for correlation dimension calculation
        let sample_size = n.min(100);
        let step = n / sample_size;
        let sampled: Vec<f32> = (0..sample_size).map(|i| values[i * step]).collect();

        // Calculate correlation dimension with multiple radii
        let mut log_radii = Vec::new();
        let mut log_counts = Vec::new();

        let max_val = sampled.iter().fold(f32::NEG_INFINITY, |a, &b| a.max(b));
        let min_val = sampled.iter().fold(f32::INFINITY, |a, &b| a.min(b));
        let range = max_val - min_val;

        if range <= 0.0 {
            return 1.0;
        }

        for radius_factor in [0.001, 0.01, 0.1, 0.5] {
            let radius = range * radius_factor;
            let mut count = 0;

            for i in 0..sampled.len() {
                for j in (i + 1)..sampled.len() {
                    if (sampled[i] - sampled[j]).abs() < radius {
                        count += 1;
                    }
                }
            }

            if count > 0 {
                log_radii.push(radius.ln());
                log_counts.push((count as f32).ln());
            }
        }

        // Linear regression to estimate dimension
        if log_radii.len() < 2 {
            return n as f32;
        }

        let mean_log_r: f32 = log_radii.iter().sum::<f32>() / log_radii.len() as f32;
        let mean_log_c: f32 = log_counts.iter().sum::<f32>() / log_counts.len() as f32;

        let numerator: f32 = log_radii
            .iter()
            .zip(&log_counts)
            .map(|(r, c)| (r - mean_log_r) * (c - mean_log_c))
            .sum();
        let denominator: f32 = log_radii.iter().map(|r| (r - mean_log_r).powi(2)).sum();

        if denominator > 0.0 {
            let slope = numerator / denominator;
            slope.abs().min(n as f32).max(1.0)
        } else {
            n as f32
        }
    }

    /// Calculate Hopkins statistic for clustering tendency
    fn calculate_hopkins_statistic(values: &[f32]) -> f32 {
        let n = values.len();
        if n < 10 {
            return 0.5; // Neutral value
        }

        let sample_size = (n / 10).clamp(5, 50);
        let min_val = values.iter().fold(f32::INFINITY, |a, &b| a.min(b));
        let max_val = values.iter().fold(f32::NEG_INFINITY, |a, &b| a.max(b));

        if max_val <= min_val {
            return 0.5;
        }

        let mut w_sum = 0.0; // Sum of distances to nearest neighbor in data
        let mut u_sum = 0.0; // Sum of distances to nearest neighbor in uniform random data

        // Sample from actual data
        for i in 0..sample_size {
            let idx = (i * n / sample_size) % n;
            let point = values[idx];

            let mut min_dist = f32::INFINITY;
            for &other in values {
                if other != point {
                    let dist = (point - other).abs();
                    min_dist = min_dist.min(dist);
                }
            }
            w_sum += min_dist;
        }

        // Generate uniform random points and find nearest neighbors in data
        use std::collections::hash_map::DefaultHasher;
        use std::hash::{Hash, Hasher};

        let mut hasher = DefaultHasher::new();
        42u64.hash(&mut hasher);
        let mut rng_state = hasher.finish();

        for _ in 0..sample_size {
            rng_state = rng_state.wrapping_mul(1103515245).wrapping_add(12345);
            let random_point = min_val + (max_val - min_val) * (rng_state as f32 / u64::MAX as f32);

            let mut min_dist = f32::INFINITY;
            for &data_point in values {
                let dist = (random_point - data_point).abs();
                min_dist = min_dist.min(dist);
            }
            u_sum += min_dist;
        }

        if w_sum + u_sum > 0.0 {
            u_sum / (w_sum + u_sum)
        } else {
            0.5
        }
    }

    /// Detect vector domain based on statistical patterns
    fn detect_domain(
        _values: &[f32],
        entropy: f32,
        sparsity: f32,
        correlation: f32,
    ) -> VectorDomain {
        // Text embeddings: moderate entropy, low sparsity, moderate correlation
        if entropy > 6.0
            && entropy < 8.0
            && sparsity < 0.3
            && correlation > 0.2
            && correlation < 0.6
        {
            return VectorDomain::TextEmbeddings;
        }

        // Image features: high entropy, variable sparsity, low correlation
        if entropy > 7.0 && correlation < 0.3 {
            return VectorDomain::ImageFeatures;
        }

        // Knowledge graph: lower entropy, higher sparsity, specific patterns
        if entropy < 6.0 && sparsity > 0.4 {
            return VectorDomain::KnowledgeGraph;
        }

        // Time series: high correlation, moderate entropy
        if correlation > 0.7 && entropy > 5.0 && entropy < 7.0 {
            return VectorDomain::TimeSeriesData;
        }

        VectorDomain::Unknown
    }

    /// Calculate aggregate statistics from multiple vectors
    pub fn from_vectors(vectors: &[Vector]) -> Result<Self, VectorError> {
        Self::from_vectors_with_context(vectors, &CompressionContext::default())
    }

    /// Calculate aggregate statistics from multiple vectors with context
    pub fn from_vectors_with_context(
        vectors: &[Vector],
        context: &CompressionContext,
    ) -> Result<Self, VectorError> {
        if vectors.is_empty() {
            return Err(VectorError::InvalidDimensions(
                "No vectors provided".to_string(),
            ));
        }

        let individual_stats: Result<Vec<_>, _> = vectors
            .iter()
            .map(|v| Self::from_vector_with_context(v, context))
            .collect();
        let stats = individual_stats?;

        let n = stats.len() as f32;

        Ok(VectorStats {
            dimensions: stats[0].dimensions,
            mean: stats.iter().map(|s| s.mean).sum::<f32>() / n,
            std_dev: stats.iter().map(|s| s.std_dev).sum::<f32>() / n,
            min_val: stats
                .iter()
                .map(|s| s.min_val)
                .fold(f32::INFINITY, |a, b| a.min(b)),
            max_val: stats
                .iter()
                .map(|s| s.max_val)
                .fold(f32::NEG_INFINITY, |a, b| a.max(b)),
            entropy: stats.iter().map(|s| s.entropy).sum::<f32>() / n,
            sparsity: stats.iter().map(|s| s.sparsity).sum::<f32>() / n,
            correlation: stats.iter().map(|s| s.correlation).sum::<f32>() / n,
            intrinsic_dimension: stats.iter().map(|s| s.intrinsic_dimension).sum::<f32>() / n,
            clustering_tendency: stats.iter().map(|s| s.clustering_tendency).sum::<f32>() / n,
            temporal_stability: stats.iter().map(|s| s.temporal_stability).sum::<f32>() / n,
            domain_affinity: Self::aggregate_domain_affinity(&stats),
        })
    }

    /// Aggregate domain affinity from multiple statistics
    fn aggregate_domain_affinity(stats: &[VectorStats]) -> VectorDomain {
        let mut domain_counts = HashMap::new();

        for stat in stats {
            *domain_counts
                .entry(stat.domain_affinity.clone())
                .or_insert(0) += 1;
        }

        domain_counts
            .into_iter()
            .max_by_key(|(_, count)| *count)
            .map(|(domain, _)| domain)
            .unwrap_or(VectorDomain::Unknown)
    }
}

/// Performance metrics for compression methods
#[derive(Debug, Clone)]
pub struct CompressionMetrics {
    pub method: CompressionMethod,
    pub compression_ratio: f32,
    pub compression_time: Duration,
    pub decompression_time: Duration,
    pub reconstruction_error: f32,
    pub usage_count: u64,
    pub avg_performance_score: f32,
}

impl CompressionMetrics {
    pub fn new(method: CompressionMethod) -> Self {
        Self {
            method,
            compression_ratio: 1.0,
            compression_time: Duration::ZERO,
            decompression_time: Duration::ZERO,
            reconstruction_error: 0.0,
            usage_count: 0,
            avg_performance_score: 0.0,
        }
    }

    /// Calculate performance score (higher is better)
    pub fn calculate_score(&self, priorities: &CompressionPriorities) -> f32 {
        let ratio_score = self.compression_ratio.min(0.9); // Cap at 90% compression
        let speed_score = 1.0 / (1.0 + self.compression_time.as_millis() as f32 / 1000.0);
        let accuracy_score = 1.0 / (1.0 + self.reconstruction_error);

        priorities.compression_weight * ratio_score
            + priorities.speed_weight * speed_score
            + priorities.accuracy_weight * accuracy_score
    }

    /// Update metrics with new measurement
    pub fn update(
        &mut self,
        compression_ratio: f32,
        comp_time: Duration,
        decomp_time: Duration,
        error: f32,
        priorities: &CompressionPriorities,
    ) {
        let alpha = 0.1; // Learning rate for exponential moving average

        self.compression_ratio = self.compression_ratio * (1.0 - alpha) + compression_ratio * alpha;
        self.compression_time = Duration::from_nanos(
            (self.compression_time.as_nanos() as f32 * (1.0 - alpha)
                + comp_time.as_nanos() as f32 * alpha) as u64,
        );
        self.decompression_time = Duration::from_nanos(
            (self.decompression_time.as_nanos() as f32 * (1.0 - alpha)
                + decomp_time.as_nanos() as f32 * alpha) as u64,
        );
        self.reconstruction_error = self.reconstruction_error * (1.0 - alpha) + error * alpha;
        self.usage_count += 1;

        self.avg_performance_score = self.calculate_score(priorities);
    }
}

/// Priorities for compression strategy selection
#[derive(Debug, Clone)]
pub struct CompressionPriorities {
    pub compression_weight: f32, // Importance of compression ratio
    pub speed_weight: f32,       // Importance of compression/decompression speed
    pub accuracy_weight: f32,    // Importance of reconstruction accuracy
}

impl Default for CompressionPriorities {
    fn default() -> Self {
        Self {
            compression_weight: 0.4,
            speed_weight: 0.3,
            accuracy_weight: 0.3,
        }
    }
}

/// Multi-level compression strategy
#[derive(Debug, Clone)]
pub struct MultiLevelCompression {
    pub levels: Vec<CompressionMethod>,
    pub thresholds: Vec<f32>, // Quality thresholds for each level
}

impl Default for MultiLevelCompression {
    fn default() -> Self {
        Self::new()
    }
}

impl MultiLevelCompression {
    pub fn new() -> Self {
        Self {
            levels: vec![
                CompressionMethod::None,
                CompressionMethod::Quantization { bits: 16 },
                CompressionMethod::Quantization { bits: 8 },
                CompressionMethod::Pca { components: 0 }, // Will be set adaptively
                CompressionMethod::Zstd { level: 3 },
            ],
            thresholds: vec![0.0, 0.1, 0.3, 0.6, 0.8],
        }
    }

    /// Select compression level based on requirements
    pub fn select_level(&self, required_compression: f32) -> &CompressionMethod {
        for (i, &threshold) in self.thresholds.iter().enumerate() {
            if required_compression <= threshold {
                return &self.levels[i];
            }
        }
        self.levels.last().unwrap()
    }
}

/// Adaptive compression engine that learns optimal strategies
pub struct AdaptiveCompressor {
    /// Current compression priorities
    priorities: CompressionPriorities,
    /// Performance metrics for each compression method
    metrics: Arc<RwLock<HashMap<String, CompressionMetrics>>>,
    /// Multi-level compression strategies
    multi_level: MultiLevelCompression,
    /// Cache of trained compressors
    compressor_cache: Arc<RwLock<HashMap<String, Box<dyn VectorCompressor + Send + Sync>>>>,
    /// Statistics cache for similar vectors
    stats_cache: Arc<RwLock<HashMap<String, (VectorStats, Instant)>>>,
    /// Learning parameters
    exploration_rate: f32,
    cache_ttl: Duration,
}

impl AdaptiveCompressor {
    pub fn new() -> Self {
        Self::new_with_priorities(CompressionPriorities::default())
    }

    pub fn new_with_priorities(priorities: CompressionPriorities) -> Self {
        Self {
            priorities,
            metrics: Arc::new(RwLock::new(HashMap::new())),
            multi_level: MultiLevelCompression::new(),
            compressor_cache: Arc::new(RwLock::new(HashMap::new())),
            stats_cache: Arc::new(RwLock::new(HashMap::new())),
            exploration_rate: 0.1,
            cache_ttl: Duration::from_secs(3600), // 1 hour cache TTL
        }
    }

    /// Analyze vector characteristics and recommend compression strategy
    pub fn analyze_and_recommend(
        &mut self,
        vectors: &[Vector],
    ) -> Result<CompressionMethod, VectorError> {
        let stats = VectorStats::from_vectors(vectors)?;
        let stats_key = self.generate_stats_key(&stats);

        // Check cache first
        {
            let cache = self.stats_cache.read().unwrap();
            if let Some((cached_stats, timestamp)) = cache.get(&stats_key) {
                if timestamp.elapsed() < self.cache_ttl {
                    return Ok(self.recommend_from_stats(cached_stats));
                }
            }
        }

        // Cache the stats
        {
            let mut cache = self.stats_cache.write().unwrap();
            cache.insert(stats_key, (stats.clone(), Instant::now()));
        }

        Ok(self.recommend_from_stats(&stats))
    }

    /// Recommend compression method based on vector statistics
    fn recommend_from_stats(&self, stats: &VectorStats) -> CompressionMethod {
        // High sparsity -> prefer quantization or PCA
        if stats.sparsity > 0.7 {
            return CompressionMethod::Quantization { bits: 4 };
        }

        // High correlation -> PCA works well
        if stats.correlation > 0.6 && stats.dimensions > 20 {
            let components = (stats.dimensions as f32 * 0.7) as usize;
            return CompressionMethod::Pca { components };
        }

        // Low entropy -> Zstd compression is effective
        if stats.entropy < 4.0 {
            return CompressionMethod::Zstd { level: 9 };
        }

        // High variance -> quantization with more bits
        if stats.std_dev > stats.mean.abs() {
            return CompressionMethod::Quantization { bits: 12 };
        }

        // Default: moderate quantization
        CompressionMethod::Quantization { bits: 8 }
    }

    /// Compress vector with adaptive strategy selection
    pub fn compress_adaptive(&mut self, vector: &Vector) -> Result<Vec<u8>, VectorError> {
        let stats = VectorStats::from_vector(vector)?;
        let method = self.recommend_from_stats(&stats);

        // Check if we should explore alternative methods
        if self.should_explore() {
            let alternative = self.get_alternative_method(&method);
            return self.compress_with_method(vector, &alternative);
        }

        self.compress_with_method(vector, &method)
    }

    /// Compress with specific method and update metrics
    pub fn compress_with_method(
        &mut self,
        vector: &Vector,
        method: &CompressionMethod,
    ) -> Result<Vec<u8>, VectorError> {
        let method_key = format!("{method:?}");
        let compressor = self.get_or_create_compressor(method)?;

        let start_time = Instant::now();
        let compressed = compressor.compress(vector)?;
        let compression_time = start_time.elapsed();

        // Measure reconstruction error
        let decompressed = compressor.decompress(&compressed, vector.dimensions)?;
        let error = self.calculate_reconstruction_error(vector, &decompressed)?;

        let compression_ratio = compressed.len() as f32 / (vector.dimensions * 4) as f32; // Assuming f32 vectors

        // Update metrics
        {
            let mut metrics = self.metrics.write().unwrap();
            let metric = metrics
                .entry(method_key)
                .or_insert_with(|| CompressionMetrics::new(method.clone()));
            metric.update(
                compression_ratio,
                compression_time,
                Duration::ZERO,
                error,
                &self.priorities,
            );
        }

        Ok(compressed)
    }

    /// Multi-level compression for extreme compression ratios
    pub fn compress_multi_level(
        &mut self,
        vector: &Vector,
        target_ratio: f32,
    ) -> Result<Vec<u8>, VectorError> {
        let mut current_vector = vector.clone();
        let mut compression_steps = Vec::new();
        let mut total_ratio = 1.0;

        while total_ratio > target_ratio && compression_steps.len() < 3 {
            let remaining_ratio = target_ratio / total_ratio;
            let method = self.multi_level.select_level(remaining_ratio);

            let compressor = self.get_or_create_compressor(method)?;
            let compressed = compressor.compress(&current_vector)?;

            let step_ratio = compressed.len() as f32 / (current_vector.dimensions * 4) as f32;
            total_ratio *= step_ratio;

            compression_steps.push((method.clone(), compressed.clone()));

            // Prepare for next level if needed
            if total_ratio > target_ratio {
                current_vector = compressor.decompress(&compressed, current_vector.dimensions)?;
            }
        }

        // Serialize the compression steps
        self.serialize_multi_level_result(compression_steps)
    }

    /// Get best performing compression method based on current metrics
    pub fn get_best_method(&self) -> CompressionMethod {
        let metrics = self.metrics.read().unwrap();
        let best = metrics.values().max_by(|a, b| {
            a.avg_performance_score
                .partial_cmp(&b.avg_performance_score)
                .unwrap()
        });

        best.map(|m| m.method.clone())
            .unwrap_or(CompressionMethod::Quantization { bits: 8 })
    }

    /// Get compression performance statistics
    pub fn get_performance_stats(&self) -> HashMap<String, CompressionMetrics> {
        self.metrics.read().unwrap().clone()
    }

    /// Update compression priorities
    pub fn update_priorities(&mut self, priorities: CompressionPriorities) {
        self.priorities = priorities;

        // Recalculate scores for all metrics
        let mut metrics = self.metrics.write().unwrap();
        for metric in metrics.values_mut() {
            metric.avg_performance_score = metric.calculate_score(&self.priorities);
        }
    }

    /// Clear caches and reset learning
    pub fn reset(&mut self) {
        self.metrics.write().unwrap().clear();
        self.compressor_cache.write().unwrap().clear();
        self.stats_cache.write().unwrap().clear();
    }

    // Private helper methods

    fn get_or_create_compressor(
        &self,
        method: &CompressionMethod,
    ) -> Result<Box<dyn VectorCompressor>, VectorError> {
        let method_key = format!("{method:?}");

        {
            let cache = self.compressor_cache.read().unwrap();
            if cache.contains_key(&method_key) {
                // Note: We can't return a reference here due to trait object limitations
                // So we create a new instance
            }
        }

        // Create new compressor (existing create_compressor function)
        let compressor = create_compressor(method);

        // Cache it (though we can't use it directly due to trait object limitations)
        {
            let _cache = self.compressor_cache.write().unwrap();
            // Note: This is a placeholder for caching logic
            // In practice, we might need to redesign this for trait objects
        }

        Ok(compressor)
    }

    fn calculate_reconstruction_error(
        &self,
        original: &Vector,
        reconstructed: &Vector,
    ) -> Result<f32, VectorError> {
        let orig_values = original.as_f32();
        let recon_values = reconstructed.as_f32();

        if orig_values.len() != recon_values.len() {
            return Err(VectorError::InvalidDimensions(
                "Dimension mismatch".to_string(),
            ));
        }

        let mse: f32 = orig_values
            .iter()
            .zip(recon_values.iter())
            .map(|(a, b)| (a - b).powi(2))
            .sum::<f32>()
            / orig_values.len() as f32;

        Ok(mse.sqrt()) // RMSE
    }

    fn generate_stats_key(&self, stats: &VectorStats) -> String {
        format!(
            "{}_{:.2}_{:.2}_{:.2}_{:.2}",
            stats.dimensions, stats.entropy, stats.sparsity, stats.correlation, stats.std_dev
        )
    }

    fn should_explore(&self) -> bool {
        use scirs2_core::random::{Random, Rng};
        let mut rng = Random::seed(42);
        rng.gen_range(0.0..1.0) < self.exploration_rate
    }

    fn get_alternative_method(&self, current: &CompressionMethod) -> CompressionMethod {
        match current {
            CompressionMethod::None => CompressionMethod::Quantization { bits: 8 },
            CompressionMethod::Quantization { bits } => {
                if *bits > 8 {
                    CompressionMethod::Quantization { bits: bits - 2 }
                } else {
                    CompressionMethod::Pca { components: 16 }
                }
            }
            CompressionMethod::Pca { components: _ } => CompressionMethod::Zstd { level: 6 },
            CompressionMethod::Zstd { level } => {
                if *level < 15 {
                    CompressionMethod::Zstd { level: level + 3 }
                } else {
                    CompressionMethod::Quantization { bits: 4 }
                }
            }
            _ => CompressionMethod::None,
        }
    }

    fn serialize_multi_level_result(
        &self,
        steps: Vec<(CompressionMethod, Vec<u8>)>,
    ) -> Result<Vec<u8>, VectorError> {
        use std::io::Write;

        let mut result = Vec::new();

        // Write number of steps
        result.write_all(&(steps.len() as u32).to_le_bytes())?;

        // Write each step
        for (method, data) in steps {
            // Serialize method (simplified)
            let method_id = match method {
                CompressionMethod::None => 0u8,
                CompressionMethod::Zstd { .. } => 1u8,
                CompressionMethod::Quantization { .. } => 2u8,
                CompressionMethod::Pca { .. } => 3u8,
                CompressionMethod::ProductQuantization { .. } => 4u8,
                CompressionMethod::Adaptive { .. } => 5u8,
            };
            result.push(method_id);

            // Write data length and data
            result.write_all(&(data.len() as u32).to_le_bytes())?;
            result.extend_from_slice(&data);
        }

        Ok(result)
    }
}

impl Default for AdaptiveCompressor {
    fn default() -> Self {
        Self::new()
    }
}

/// Domain-specific compression profiles
pub struct CompressionProfiles {
    profiles: HashMap<VectorDomain, CompressionPriorities>,
}

impl Default for CompressionProfiles {
    fn default() -> Self {
        Self::new()
    }
}

impl CompressionProfiles {
    pub fn new() -> Self {
        let mut profiles = HashMap::new();

        // Text embeddings: balance compression and quality
        profiles.insert(
            VectorDomain::TextEmbeddings,
            CompressionPriorities {
                compression_weight: 0.3,
                speed_weight: 0.4,
                accuracy_weight: 0.3,
            },
        );

        // Image features: favor compression due to redundancy
        profiles.insert(
            VectorDomain::ImageFeatures,
            CompressionPriorities {
                compression_weight: 0.5,
                speed_weight: 0.2,
                accuracy_weight: 0.3,
            },
        );

        // Knowledge graph: favor accuracy
        profiles.insert(
            VectorDomain::KnowledgeGraph,
            CompressionPriorities {
                compression_weight: 0.2,
                speed_weight: 0.3,
                accuracy_weight: 0.5,
            },
        );

        // Time series: balance speed and accuracy
        profiles.insert(
            VectorDomain::TimeSeriesData,
            CompressionPriorities {
                compression_weight: 0.3,
                speed_weight: 0.4,
                accuracy_weight: 0.3,
            },
        );

        // Audio features: favor compression
        profiles.insert(
            VectorDomain::AudioFeatures,
            CompressionPriorities {
                compression_weight: 0.4,
                speed_weight: 0.3,
                accuracy_weight: 0.3,
            },
        );

        Self { profiles }
    }

    pub fn get_profile(&self, domain: &VectorDomain) -> CompressionPriorities {
        self.profiles.get(domain).cloned().unwrap_or_default()
    }

    pub fn update_profile(&mut self, domain: VectorDomain, priorities: CompressionPriorities) {
        self.profiles.insert(domain, priorities);
    }
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn test_vector_stats() {
        let vector = Vector::new(vec![1.0, 2.0, 3.0, 4.0, 5.0]);
        let stats = VectorStats::from_vector(&vector).unwrap();

        assert_eq!(stats.dimensions, 5);
        assert_eq!(stats.mean, 3.0);
        assert!(stats.std_dev > 0.0);
    }

    #[test]
    fn test_adaptive_compression() {
        let vectors = vec![
            Vector::new(vec![1.0, 2.0, 3.0, 4.0]),
            Vector::new(vec![2.0, 3.0, 4.0, 5.0]),
            Vector::new(vec![3.0, 4.0, 5.0, 6.0]),
        ];

        let mut compressor = AdaptiveCompressor::new();
        let recommended = compressor.analyze_and_recommend(&vectors).unwrap();

        // Should recommend some compression method
        assert!(!matches!(recommended, CompressionMethod::None));
    }

    #[test]
    fn test_compression_metrics() {
        let method = CompressionMethod::Quantization { bits: 8 };
        let mut metrics = CompressionMetrics::new(method);
        let priorities = CompressionPriorities::default();

        metrics.update(
            0.5,
            Duration::from_millis(10),
            Duration::from_millis(5),
            0.01,
            &priorities,
        );

        assert!(metrics.avg_performance_score > 0.0);
        assert_eq!(metrics.usage_count, 1);
    }

    #[test]
    fn test_multi_level_compression() {
        let mut compressor = AdaptiveCompressor::new();
        // Use a larger vector with repetitive patterns that compress well
        let values: Vec<f32> = (0..256).map(|i| (i % 16) as f32).collect();
        let vector = Vector::new(values);

        let compressed = compressor.compress_multi_level(&vector, 0.1).unwrap();

        // Should achieve significant compression on this larger, repetitive vector
        // Original size would be 256 * 4 = 1024 bytes
        // Multi-level compression includes metadata overhead, so expect reasonable compression
        println!(
            "Compressed size: {} bytes, original size: {} bytes",
            compressed.len(),
            vector.dimensions * 4
        );
        assert!(compressed.len() < vector.dimensions * 4); // At least some compression
        assert!(compressed.len() < 900); // Should achieve at least 12% compression
    }

    #[test]
    fn test_stats_aggregation() {
        let vectors = vec![
            Vector::new(vec![1.0, 2.0]),
            Vector::new(vec![3.0, 4.0]),
            Vector::new(vec![5.0, 6.0]),
        ];

        let stats = VectorStats::from_vectors(&vectors).unwrap();
        assert_eq!(stats.dimensions, 2);
        assert!(stats.mean > 0.0);
    }
}