9. Estrategias de Caché para Servicios de IA

Introducción

El caché es esencial para optimizar costos y rendimiento en aplicaciones de IA. Las llamadas a LLMs son costosas en tiempo y dinero. En este tutorial aprenderás estrategias efectivas de caché para servicios de IA.

¿Por Qué Cachear?

1. Reducción de costos: Menos llamadas al API = menor gasto
2. Mejor rendimiento: Respuestas instantáneas desde caché
3. Resiliencia: Funciona aunque el servicio esté caído
4. Reducción de latencia: Microsegundos vs segundos

Tipos de Caché

1. Caché en Memoria (IMemoryCache)

using Microsoft.Extensions.Caching.Memory;

public class MemoryCachedAIService
{
    private readonly IMemoryCache _cache;
    private readonly IAIService _aiService;
    
    public MemoryCachedAIService(IMemoryCache cache, IAIService aiService)
    {
        _cache = cache;
        _aiService = aiService;
    }
    
    public async Task<string> GetResponseAsync(string prompt)
    {
        var cacheKey = GenerateCacheKey(prompt);
        
        if (_cache.TryGetValue<string>(cacheKey, out var cachedResponse))
        {
            return cachedResponse;
        }
        
        var response = await _aiService.GenerateAsync(prompt);
        
        var cacheOptions = new MemoryCacheEntryOptions
        {
            AbsoluteExpirationRelativeToNow = TimeSpan.FromHours(24),
            SlidingExpiration = TimeSpan.FromHours(6),
            Priority = CacheItemPriority.Normal
        };
        
        _cache.Set(cacheKey, response, cacheOptions);
        
        return response;
    }
    
    private string GenerateCacheKey(string prompt)
    {
        using var sha256 = SHA256.Create();
        var hashBytes = sha256.ComputeHash(Encoding.UTF8.GetBytes(prompt));
        return $"ai_{BitConverter.ToString(hashBytes).Replace("-", "").Substring(0, 16)}";
    }
}

2. Caché Distribuido (Redis)

using Microsoft.Extensions.Caching.Distributed;
using System.Text.Json;

public class DistributedCachedAIService
{
    private readonly IDistributedCache _cache;
    private readonly IAIService _aiService;
    private readonly ILogger<DistributedCachedAIService> _logger;
    
    public DistributedCachedAIService(
        IDistributedCache cache,
        IAIService aiService,
        ILogger<DistributedCachedAIService> logger)
    {
        _cache = cache;
        _aiService = aiService;
        _logger = logger;
    }
    
    public async Task<AIResponse> GetResponseAsync(
        string prompt,
        CancellationToken cancellationToken = default)
    {
        var cacheKey = GenerateCacheKey(prompt);
        
        // Intentar obtener del caché
        var cachedBytes = await _cache.GetAsync(cacheKey, cancellationToken);
        
        if (cachedBytes != null)
        {
            _logger.LogInformation("Cache HIT: {CacheKey}", cacheKey);
            var cachedResponse = JsonSerializer.Deserialize<AIResponse>(cachedBytes);
            if (cachedResponse != null)
            {
                cachedResponse.FromCache = true;
                return cachedResponse;
            }
        }
        
        _logger.LogInformation("Cache MISS: {CacheKey}", cacheKey);
        
        // Generar respuesta
        var response = await _aiService.GenerateAsync(prompt, cancellationToken);
        
        // Guardar en caché
        var responseBytes = JsonSerializer.SerializeToUtf8Bytes(response);
        
        var cacheOptions = new DistributedCacheEntryOptions
        {
            AbsoluteExpirationRelativeToNow = TimeSpan.FromDays(7),
            SlidingExpiration = TimeSpan.FromDays(1)
        };
        
        await _cache.SetAsync(cacheKey, responseBytes, cacheOptions, cancellationToken);
        
        response.FromCache = false;
        return response;
    }
    
    private string GenerateCacheKey(string prompt)
    {
        using var sha256 = SHA256.Create();
        var hashBytes = sha256.ComputeHash(Encoding.UTF8.GetBytes(prompt));
        return $"ai_{BitConverter.ToString(hashBytes).Replace("-", "")}";
    }
}

public class AIResponse
{
    public required string Content { get; set; }
    public DateTime GeneratedAt { get; set; } = DateTime.UtcNow;
    public bool FromCache { get; set; }
    public Dictionary<string, string>? Metadata { get; set; }
}

3. Caché de Embeddings

public class EmbeddingCacheService
{
    private readonly IDistributedCache _cache;
    private readonly ITextEmbeddingGenerationService _embeddingService;
    
    public async Task<ReadOnlyMemory<float>> GetEmbeddingAsync(
        string text,
        CancellationToken cancellationToken = default)
    {
        var cacheKey = $"emb_{ComputeHash(text)}";
        
        var cachedBytes = await _cache.GetAsync(cacheKey, cancellationToken);
        
        if (cachedBytes != null)
        {
            // Deserializar embedding
            var floatArray = new float[cachedBytes.Length / sizeof(float)];
            Buffer.BlockCopy(cachedBytes, 0, floatArray, 0, cachedBytes.Length);
            return new ReadOnlyMemory<float>(floatArray);
        }
        
        // Generar embedding
        var embeddings = await _embeddingService.GenerateEmbeddingsAsync(
            new[] { text },
            kernel: null,
            cancellationToken);
        
        var embedding = embeddings.First();
        
        // Serializar y cachear
        var floats = embedding.ToArray();
        var bytes = new byte[floats.Length * sizeof(float)];
        Buffer.BlockCopy(floats, 0, bytes, 0, bytes.Length);
        
        await _cache.SetAsync(
            cacheKey,
            bytes,
            new DistributedCacheEntryOptions
            {
                AbsoluteExpirationRelativeToNow = TimeSpan.FromDays(30)
            },
            cancellationToken);
        
        return embedding;
    }
    
    private string ComputeHash(string text)
    {
        using var sha256 = SHA256.Create();
        var hashBytes = sha256.ComputeHash(Encoding.UTF8.GetBytes(text));
        return BitConverter.ToString(hashBytes).Replace("-", "");
    }
}

Caché Inteligente

Caché con Versioning

public class VersionedCacheService
{
    private readonly IDistributedCache _cache;
    private readonly string _version;
    
    public VersionedCacheService(IDistributedCache cache, string version = "v1")
    {
        _cache = cache;
        _version = version;
    }
    
    private string GenerateCacheKey(string key)
    {
        return $"{_version}_{key}";
    }
    
    public async Task InvalidateVersionAsync()
    {
        // Cambiar versión invalida todo el caché anterior
        // Implementar incremento de versión
    }
}

Caché con TTL Dinámico

public class DynamicTTLCacheService
{
    private readonly IDistributedCache _cache;
    
    public async Task<string> GetWithDynamicTTLAsync(
        string prompt,
        Func<string, TimeSpan> ttlCalculator)
    {
        var cacheKey = GenerateCacheKey(prompt);
        var cached = await _cache.GetStringAsync(cacheKey);
        
        if (cached != null)
        {
            return cached;
        }
        
        var response = await GenerateResponseAsync(prompt);
        
        // TTL basado en características de la respuesta
        var ttl = ttlCalculator(response);
        
        await _cache.SetStringAsync(
            cacheKey,
            response,
            new DistributedCacheEntryOptions
            {
                AbsoluteExpirationRelativeToNow = ttl
            });
        
        return response;
    }
}

// Uso
var response = await cacheService.GetWithDynamicTTLAsync(
    prompt,
    response => response.Length > 1000 
        ? TimeSpan.FromDays(7)  // Respuestas largas: TTL largo
        : TimeSpan.FromHours(1)); // Respuestas cortas: TTL corto

Caché por Similitud Semántica

public class SemanticCacheService
{
    private readonly IDistributedCache _cache;
    private readonly ITextEmbeddingGenerationService _embeddingService;
    private readonly double _similarityThreshold = 0.95;
    
    public async Task<(bool Found, string? Response)> TryGetSimilarAsync(
        string prompt,
        CancellationToken cancellationToken = default)
    {
        // Generar embedding del prompt
        var embeddings = await _embeddingService.GenerateEmbeddingsAsync(
            new[] { prompt },
            kernel: null,
            cancellationToken);
        
        var promptEmbedding = embeddings.First();
        
        // Buscar prompts similares en caché (simplificado)
        // En producción, usar una base de datos vectorial
        var cachedPrompts = await GetCachedPromptsAsync();
        
        foreach (var cachedPrompt in cachedPrompts)
        {
            var similarity = CalculateCosineSimilarity(
                promptEmbedding,
                cachedPrompt.Embedding);
            
            if (similarity >= _similarityThreshold)
            {
                var response = await _cache.GetStringAsync(cachedPrompt.Key);
                if (response != null)
                {
                    return (true, response);
                }
            }
        }
        
        return (false, null);
    }
    
    private double CalculateCosineSimilarity(
        ReadOnlyMemory<float> v1,
        ReadOnlyMemory<float> v2)
    {
        // Implementación de similitud coseno
        return 0.0;
    }
    
    private Task<List<CachedPrompt>> GetCachedPromptsAsync()
    {
        // Implementación
        return Task.FromResult(new List<CachedPrompt>());
    }
}

public class CachedPrompt
{
    public required string Key { get; init; }
    public required ReadOnlyMemory<float> Embedding { get; init; }
}

Estrategias de Invalidación

Invalidación por Tiempo

public class TimeBasedInvalidation
{
    private readonly IMemoryCache _cache;
    
    public void SetWithExpiration<T>(string key, T value, TimeSpan expiration)
    {
        _cache.Set(key, value, new MemoryCacheEntryOptions
        {
            AbsoluteExpirationRelativeToNow = expiration
        });
    }
}

Invalidación por Evento

public class EventBasedInvalidation
{
    private readonly IMemoryCache _cache;
    
    public void InvalidateOnEvent(string pattern)
    {
        // Invalidar todas las claves que coincidan con el patrón
        // Nota: IMemoryCache no soporta pattern matching nativamente
        // Necesitarías mantener un registro de claves
    }
    
    public void InvalidateRelated(string entityId)
    {
        // Invalidar caché relacionado con una entidad
        var relatedKeys = new[]
        {
            $"entity_{entityId}",
            $"list_with_{entityId}",
            $"summary_of_{entityId}"
        };
        
        foreach (var key in relatedKeys)
        {
            _cache.Remove(key);
        }
    }
}

Invalidación por Tamaño

public class SizeLimitedCache
{
    private readonly IMemoryCache _cache;
    private readonly long _maxSizeInBytes;
    
    public SizeLimitedCache(IMemoryCache cache, long maxSizeInBytes)
    {
        _cache = cache;
        _maxSizeInBytes = maxSizeInBytes;
    }
    
    public void Set<T>(string key, T value, long estimatedSize)
    {
        _cache.Set(key, value, new MemoryCacheEntryOptions
        {
            Size = estimatedSize,
            Priority = CacheItemPriority.Normal
        });
    }
}

Caché en Múltiples Niveles

public class MultiLevelCacheService
{
    private readonly IMemoryCache _l1Cache;      // Nivel 1: Memoria local
    private readonly IDistributedCache _l2Cache; // Nivel 2: Redis
    private readonly IAIService _aiService;       // Nivel 3: Servicio IA
    
    public async Task<string> GetResponseAsync(
        string prompt,
        CancellationToken cancellationToken = default)
    {
        var cacheKey = GenerateCacheKey(prompt);
        
        // Nivel 1: Memoria local
        if (_l1Cache.TryGetValue<string>(cacheKey, out var l1Response))
        {
            return l1Response;
        }
        
        // Nivel 2: Caché distribuido
        var l2Response = await _l2Cache.GetStringAsync(cacheKey, cancellationToken);
        if (l2Response != null)
        {
            // Poblar L1
            _l1Cache.Set(cacheKey, l2Response, TimeSpan.FromMinutes(5));
            return l2Response;
        }
        
        // Nivel 3: Generar desde IA
        var response = await _aiService.GenerateAsync(prompt, cancellationToken);
        
        // Poblar ambos niveles
        _l1Cache.Set(cacheKey, response, TimeSpan.FromMinutes(5));
        await _l2Cache.SetStringAsync(
            cacheKey,
            response,
            new DistributedCacheEntryOptions
            {
                AbsoluteExpirationRelativeToNow = TimeSpan.FromHours(24)
            },
            cancellationToken);
        
        return response;
    }
    
    private string GenerateCacheKey(string prompt)
    {
        using var sha256 = SHA256.Create();
        var hashBytes = sha256.ComputeHash(Encoding.UTF8.GetBytes(prompt));
        return BitConverter.ToString(hashBytes).Replace("-", "");
    }
}

Monitoreo de Caché

public class CacheMonitor
{
    private long _hits;
    private long _misses;
    private readonly ILogger<CacheMonitor> _logger;
    
    public void RecordHit()
    {
        Interlocked.Increment(ref _hits);
    }
    
    public void RecordMiss()
    {
        Interlocked.Increment(ref _misses);
    }
    
    public CacheStatistics GetStatistics()
    {
        var totalRequests = _hits + _misses;
        var hitRate = totalRequests > 0 ? (double)_hits / totalRequests : 0;
        
        return new CacheStatistics
        {
            Hits = _hits,
            Misses = _misses,
            TotalRequests = totalRequests,
            HitRate = hitRate
        };
    }
    
    public void LogStatistics()
    {
        var stats = GetStatistics();
        _logger.LogInformation(
            "Cache Stats: {Hits} hits, {Misses} misses, {HitRate:P2} hit rate",
            stats.Hits,
            stats.Misses,
            stats.HitRate);
    }
}

public class CacheStatistics
{
    public long Hits { get; init; }
    public long Misses { get; init; }
    public long TotalRequests { get; init; }
    public double HitRate { get; init; }
}

Mejores Prácticas

1. Definir TTL Apropiado

// ✅ TTL basado en volatilidad de datos
var staticDataTTL = TimeSpan.FromDays(7);
var dynamicDataTTL = TimeSpan.FromMinutes(5);
var realtimeDataTTL = TimeSpan.FromSeconds(30);

2. Considerar Tamaño de Caché

// ✅ Limitar tamaño para evitar problemas de memoria
services.AddMemoryCache(options =>
{
    options.SizeLimit = 1024; // Límite en unidades arbitrarias
});

3. Caché Selectivo

// ✅ Solo cachear operaciones costosas
public async Task<string> GetDataAsync(string id, bool useCache = true)
{
    if (!useCache || IsRealtimeRequired(id))
    {
        return await FetchFromSourceAsync(id);
    }
    
    return await GetFromCacheAsync(id);
}

4. Warming del Caché

public class CacheWarmer : IHostedService
{
    public async Task StartAsync(CancellationToken cancellationToken)
    {
        // Pre-cargar datos frecuentes al iniciar
        await WarmFrequentQueriesAsync(cancellationToken);
    }
    
    private async Task WarmFrequentQueriesAsync(CancellationToken cancellationToken)
    {
        var frequentQueries = await GetFrequentQueriesAsync();
        
        foreach (var query in frequentQueries)
        {
            await _cachedService.GetResponseAsync(query, cancellationToken);
        }
    }
    
    public Task StopAsync(CancellationToken cancellationToken) => Task.CompletedTask;
}

Conclusión

El caché es esencial para aplicaciones de IA eficientes. Implementa caché en múltiples niveles, monitorea hit rates, y ajusta TTLs basándote en patrones de uso. Una estrategia de caché bien diseñada puede reducir costos hasta 80% y mejorar significativamente el rendimiento.

Palabras clave: caching strategies, distributed cache, Redis, memory cache, AI optimization, performance, cost reduction