5a2fdacebe
- Add Node.js heap limit (256MB default) via OMNIROUTE_MEMORY_MB - Convert LRU caches to byte-based limits (2-4MB) - Reduce in-memory buffer sizes: 500→200 entries - Add .env file loading in CLI for global npm installs - Add PM2 ecosystem.config.js example for deployment - Document memory tuning env vars in USER_GUIDE.md Estimated RAM reduction: ~60-70% (from ~512MB+ to ~150-200MB)
301 lines
8.3 KiB
TypeScript
301 lines
8.3 KiB
TypeScript
/**
|
|
* Semantic Cache — Phase 9.1
|
|
*
|
|
* Caches non-streaming LLM responses (temperature=0) to reduce cost and latency.
|
|
* Two-tier: in-memory LRU (fast) + SQLite (persistent across restarts).
|
|
*
|
|
* Cache key = SHA-256(model + normalized messages + temperature + top_p)
|
|
* Bypass: X-OmniRoute-No-Cache: true
|
|
*
|
|
* @module lib/semanticCache
|
|
*/
|
|
|
|
import crypto from "node:crypto";
|
|
import { LRUCache } from "./cacheLayer";
|
|
import { getDbInstance } from "./db/core";
|
|
|
|
// ─── Singleton ─────────────────
|
|
|
|
let memoryCache: LRUCache | null = null;
|
|
let stats = { hits: 0, misses: 0, tokensSaved: 0 };
|
|
|
|
function getMemoryCache() {
|
|
if (!memoryCache) {
|
|
memoryCache = new LRUCache({
|
|
maxSize: parseInt(process.env.SEMANTIC_CACHE_MAX_SIZE || "100", 10),
|
|
maxBytes: parseInt(process.env.SEMANTIC_CACHE_MAX_BYTES || String(4 * 1024 * 1024), 10),
|
|
defaultTTL: parseInt(process.env.SEMANTIC_CACHE_TTL_MS || "1800000", 10),
|
|
});
|
|
}
|
|
return memoryCache;
|
|
}
|
|
|
|
// ─── Signature Generation ─────────────────
|
|
|
|
/**
|
|
* Generate deterministic cache signature from request params.
|
|
* @param {string} model
|
|
* @param {Array} messages - Normalized messages array
|
|
* @param {number} temperature
|
|
* @param {number} topP
|
|
* @returns {string} hex signature
|
|
*/
|
|
export function generateSignature(model, messages, temperature = 0, topP = 1) {
|
|
const payload = JSON.stringify({
|
|
model,
|
|
messages: normalizeMessages(messages),
|
|
temperature,
|
|
top_p: topP,
|
|
});
|
|
return crypto.createHash("sha256").update(payload).digest("hex");
|
|
}
|
|
|
|
/**
|
|
* Normalize messages for consistent hashing.
|
|
* Strips metadata, keeps only role + content.
|
|
*/
|
|
function normalizeMessages(messages) {
|
|
if (!Array.isArray(messages)) return [];
|
|
return messages.map((m) => ({
|
|
role: m.role || "user",
|
|
content: typeof m.content === "string" ? m.content : JSON.stringify(m.content),
|
|
}));
|
|
}
|
|
|
|
// ─── Cache Operations ─────────────────
|
|
|
|
/**
|
|
* Check if a cached response exists for the given signature.
|
|
* Checks memory first, then SQLite.
|
|
* @param {string} signature
|
|
* @returns {object|null} Cached response or null
|
|
*/
|
|
export function getCachedResponse(signature) {
|
|
// 1. Check memory cache
|
|
const memResult = getMemoryCache().get(signature);
|
|
if (memResult) {
|
|
stats.hits++;
|
|
stats.tokensSaved += memResult.tokensSaved || 0;
|
|
return memResult.response;
|
|
}
|
|
|
|
// 2. Check SQLite
|
|
try {
|
|
const db = getDbInstance();
|
|
const row = db
|
|
.prepare(
|
|
"SELECT response, tokens_saved FROM semantic_cache WHERE signature = ? AND expires_at > datetime('now')"
|
|
)
|
|
.get(signature);
|
|
|
|
if (row) {
|
|
const parsed = JSON.parse(row.response);
|
|
// Promote to memory cache
|
|
getMemoryCache().set(signature, {
|
|
response: parsed,
|
|
tokensSaved: row.tokens_saved,
|
|
});
|
|
// Update hit count in DB
|
|
db.prepare("UPDATE semantic_cache SET hit_count = hit_count + 1 WHERE signature = ?").run(
|
|
signature
|
|
);
|
|
|
|
stats.hits++;
|
|
stats.tokensSaved += row.tokens_saved || 0;
|
|
return parsed;
|
|
}
|
|
} catch {
|
|
// DB not available — fail open
|
|
}
|
|
|
|
stats.misses++;
|
|
return null;
|
|
}
|
|
|
|
/**
|
|
* Store a response in cache.
|
|
* @param {string} signature
|
|
* @param {string} model
|
|
* @param {object} response - The API response to cache
|
|
* @param {number} tokensSaved - Estimated tokens saved
|
|
* @param {number} [ttlMs] - TTL in ms (default: 1 hour)
|
|
*/
|
|
export function setCachedResponse(signature, model, response, tokensSaved = 0, ttlMs = 3600000) {
|
|
const ttl = parseInt(process.env.SEMANTIC_CACHE_TTL_MS || String(ttlMs), 10);
|
|
|
|
// 1. Memory cache
|
|
getMemoryCache().set(signature, { response, tokensSaved }, ttl);
|
|
|
|
// 2. SQLite
|
|
try {
|
|
const db = getDbInstance();
|
|
const id = crypto.randomUUID();
|
|
const promptHash = signature.slice(0, 16);
|
|
const now = new Date().toISOString();
|
|
const expiresAt = new Date(Date.now() + ttl).toISOString();
|
|
|
|
db.prepare(
|
|
`INSERT OR REPLACE INTO semantic_cache (id, signature, model, prompt_hash, response, tokens_saved, hit_count, created_at, expires_at)
|
|
VALUES (?, ?, ?, ?, ?, ?, 0, ?, ?)`
|
|
).run(id, signature, model, promptHash, JSON.stringify(response), tokensSaved, now, expiresAt);
|
|
} catch {
|
|
// DB write failed — cache still in memory
|
|
}
|
|
}
|
|
|
|
// ─── Maintenance ─────────────────
|
|
|
|
/**
|
|
* Remove expired entries from SQLite.
|
|
* @returns {number} Number of entries removed
|
|
*/
|
|
export function cleanExpiredEntries() {
|
|
try {
|
|
const db = getDbInstance();
|
|
const result = db
|
|
.prepare("DELETE FROM semantic_cache WHERE expires_at <= datetime('now')")
|
|
.run();
|
|
return result.changes;
|
|
} catch {
|
|
return 0;
|
|
}
|
|
}
|
|
|
|
/**
|
|
* Invalidate cache entries by model name.
|
|
* Useful when a model is updated/changed and cached responses are stale.
|
|
* @param {string} model - Model name to invalidate (exact match)
|
|
* @returns {number} Number of entries removed
|
|
*/
|
|
export function invalidateByModel(model: string): number {
|
|
getMemoryCache().clear(); // Memory cache doesn't track model; full clear
|
|
try {
|
|
const db = getDbInstance();
|
|
const result = db
|
|
.prepare("DELETE FROM semantic_cache WHERE model = ?")
|
|
.run(model);
|
|
return result.changes || 0;
|
|
} catch {
|
|
return 0;
|
|
}
|
|
}
|
|
|
|
/**
|
|
* Invalidate a single cache entry by its signature.
|
|
* @param {string} signature - Cache signature to invalidate
|
|
* @returns {boolean} Whether the entry was found and removed
|
|
*/
|
|
export function invalidateBySignature(signature: string): boolean {
|
|
getMemoryCache().delete(signature);
|
|
try {
|
|
const db = getDbInstance();
|
|
const result = db
|
|
.prepare("DELETE FROM semantic_cache WHERE signature = ?")
|
|
.run(signature);
|
|
return (result.changes || 0) > 0;
|
|
} catch {
|
|
return false;
|
|
}
|
|
}
|
|
|
|
/**
|
|
* Invalidate entries older than a given age.
|
|
* @param {number} maxAgeMs - Maximum age in milliseconds
|
|
* @returns {number} Number of entries removed
|
|
*/
|
|
export function invalidateStale(maxAgeMs: number): number {
|
|
getMemoryCache().clear();
|
|
try {
|
|
const db = getDbInstance();
|
|
const cutoff = new Date(Date.now() - maxAgeMs).toISOString();
|
|
const result = db
|
|
.prepare("DELETE FROM semantic_cache WHERE created_at < ?")
|
|
.run(cutoff);
|
|
return result.changes || 0;
|
|
} catch {
|
|
return 0;
|
|
}
|
|
}
|
|
|
|
// ── Auto-cleanup timer ──
|
|
|
|
let _cleanupTimer: ReturnType<typeof setInterval> | null = null;
|
|
|
|
/**
|
|
* Start periodic auto-cleanup of expired entries.
|
|
* @param {number} intervalMs - Cleanup interval (default: 5 minutes)
|
|
*/
|
|
export function startAutoCleanup(intervalMs = 300_000): void {
|
|
stopAutoCleanup();
|
|
_cleanupTimer = setInterval(() => {
|
|
const removed = cleanExpiredEntries();
|
|
if (removed > 0) {
|
|
console.log(`[SemanticCache] Auto-cleaned ${removed} expired entries`);
|
|
}
|
|
}, intervalMs);
|
|
}
|
|
|
|
/**
|
|
* Stop periodic auto-cleanup.
|
|
*/
|
|
export function stopAutoCleanup(): void {
|
|
if (_cleanupTimer) {
|
|
clearInterval(_cleanupTimer);
|
|
_cleanupTimer = null;
|
|
}
|
|
}
|
|
|
|
/**
|
|
* Clear all cache entries.
|
|
*/
|
|
export function clearCache() {
|
|
getMemoryCache().clear();
|
|
try {
|
|
const db = getDbInstance();
|
|
db.prepare("DELETE FROM semantic_cache").run();
|
|
} catch {
|
|
// DB not available
|
|
}
|
|
stats = { hits: 0, misses: 0, tokensSaved: 0 };
|
|
}
|
|
|
|
// ─── Stats ─────────────────
|
|
|
|
/**
|
|
* Get cache statistics.
|
|
*/
|
|
export function getCacheStats() {
|
|
const memStats = getMemoryCache().getStats();
|
|
let dbSize = 0;
|
|
try {
|
|
const db = getDbInstance();
|
|
const row = db
|
|
.prepare("SELECT COUNT(*) as count FROM semantic_cache WHERE expires_at > datetime('now')")
|
|
.get();
|
|
dbSize = row?.count || 0;
|
|
} catch {
|
|
// DB not available
|
|
}
|
|
|
|
const total = stats.hits + stats.misses;
|
|
return {
|
|
memoryEntries: memStats.size,
|
|
dbEntries: dbSize,
|
|
hits: stats.hits,
|
|
misses: stats.misses,
|
|
hitRate: total > 0 ? ((stats.hits / total) * 100).toFixed(1) : "0.0",
|
|
tokensSaved: stats.tokensSaved,
|
|
};
|
|
}
|
|
|
|
/**
|
|
* Check if a request is cacheable.
|
|
* Only non-streaming, deterministic (temperature=0) requests.
|
|
*/
|
|
export function isCacheable(body, headers) {
|
|
if (headers?.get?.("x-omniroute-no-cache") === "true") return false;
|
|
if (body.stream !== false) return false;
|
|
if ((body.temperature ?? 0) !== 0) return false;
|
|
return true;
|
|
}
|