Skip to content

Never stop coding. Free MIT AI gateway: one endpoint, 375 providers (150+ free), 1,400+ models Kimi, Claude, GPT, Gemini, GLM, DeepSeek, MiniMax. Works with Claude Code, Codex, Cursor, OpenCode, Cline & Copilot. Quota-aware auto-fallback, RTK+Caveman compression saves 15-95% tokens, MCP/A2A, Desktop/PWA. Built by 900+ contributors

21.6 KB
Raw Blame
/**
 * Semantic Cache — Phase 9.1
 *
 * Caches LLM responses (temperature=0) to reduce cost and latency.
 * Supports both streaming and non-streaming requests: streaming responses
 * are cached after assembly; cache hits always return JSON.
 * Two-tier: in-memory LRU (fast) + SQLite (persistent across restarts).
 *
 * Cache key = SHA-256(model + normalized messages + temperature + top_p
 *             + output contract, when present — see outputContractOf, #12307)
 * Bypass: X-OmniRoute-No-Cache: true
 *
 * @module lib/semanticCache
 */

import crypto from "crypto";
import { LRUCache } from "./cacheLayer";
import { getDbInstance } from "./db/core";
import { toNumber } from "@/shared/utils/numeric";
import { conversationSignatureFields } from "../../open-sse/utils/conversationSignatureFields";

type JsonRecord = Record<string, unknown>;

function asRecord(value: unknown): JsonRecord {
  return value && typeof value === "object" && !Array.isArray(value) ? (value as JsonRecord) : {};
}

/**
 * Current time in the same format `expires_at` and `created_at` are written in.
 *
 * `expires_at` is a TEXT column holding an ISO-8601 string ("2026-08-26T14:00:00.000Z"),
 * so it has to be compared against another ISO-8601 string. Comparing it to SQLite's
 * `datetime('now')` ("2026-08-26 14:00:00") is a lexicographic comparison that diverges
 * at index 10, where the stored value has 'T' (0x54) and `datetime('now')` has ' ' (0x20).
 * 'T' sorts after ' ', so every row whose UTC calendar date was today compared as
 * unexpired regardless of its real time-of-day TTL.
 */
function isoNow(): string {
  return new Date().toISOString();
}

function ensureCacheMetricsTable() {
  try {
    const db = getDbInstance();
    db.prepare(
      `CREATE TABLE IF NOT EXISTS cache_metrics (
        key TEXT PRIMARY KEY,
        value INTEGER NOT NULL DEFAULT 0,
        updated_at TEXT NOT NULL DEFAULT (datetime('now'))
      )`
    ).run();
    db.prepare(
      `INSERT OR IGNORE INTO cache_metrics (key, value) VALUES ('hits', 0), ('misses', 0), ('tokens_saved', 0)`
    ).run();
  } catch {
    // DB not available
  }
}

function incrementMetric(metric: "hits" | "misses" | "tokens_saved", amount = 1) {
  try {
    const db = getDbInstance();
    db.prepare(
      `UPDATE cache_metrics SET value = value + ?, updated_at = datetime('now') WHERE key = ?`
    ).run(amount, metric);
  } catch {
    // DB not available — fall back to in-memory
  }
}

function getMetricValue(metric: string): number {
  try {
    const db = getDbInstance();
    const row = db.prepare(`SELECT value FROM cache_metrics WHERE key = ?`).get(metric);
    return row ? toNumber(asRecord(row).value, 0) : 0;
  } catch {
    return 0;
  }
}

function getHeaderValue(
  headers: { get?: (name: string) => string | null } | Record<string, unknown> | null | undefined,
  name: string
): string | null {
  if (!headers) return null;

  if (typeof headers.get === "function") {
    return headers.get(name);
  }

  const needle = name.toLowerCase();
  for (const [key, value] of Object.entries(headers)) {
    if (key.toLowerCase() !== needle) continue;
    return typeof value === "string" ? value : null;
  }

  return null;
}

// ─── Singleton ─────────────────

let memoryCache: LRUCache | null = null;

function getMemoryCache() {
  if (!memoryCache) {
    memoryCache = new LRUCache({
      maxSize: parseInt(process.env.SEMANTIC_CACHE_MAX_SIZE || "50", 10),
      maxBytes: parseInt(process.env.SEMANTIC_CACHE_MAX_BYTES || String(2 * 1024 * 1024), 10),
      defaultTTL: parseInt(process.env.SEMANTIC_CACHE_TTL_MS || "1800000", 10),
    });
    ensureCacheMetricsTable();
  }
  return memoryCache;
}

/**
 * In-memory LRU stats for the semantic cache.
 *
 * Exposed for `/api/cache/stats`, which used to report `getPromptCache()` — an
 * LRU that nothing writes to, so it always answered 0 hit / 0 miss. Same shape
 * as `LRUCache.getStats()`, so callers do not have to change.
 */
export function getMemoryCacheStats(): ReturnType<LRUCache["getStats"]> {
  return getMemoryCache().getStats();
}

/** Drop the in-memory LRU without touching the `semantic_cache` table. */
export function clearMemoryCache(): void {
  getMemoryCache().clear();
}

// ─── Signature Generation ─────────────────

/**
 * Behavior-changing generation constraints that MUST be folded into the cache signature
 * (#12734). Without these, a cached response produced under one `tool_choice`/`tools`/
 * `response_format` could be replayed for a later request that forbids or changes that
 * behavior (e.g. a cached `tool_calls` response served to a `tool_choice: "none"` request).
 *
 * The snake_case fields mirror the raw request body shape and are what `outputContractOf`
 * (#12307) fills in; the camelCase fields are the pre-existing (#12734) call-site shape.
 * `generateSignature` folds both spellings in so neither call style silently drops a field.
 *
 * Generation params (#15149): `reasoning`/`reasoning_effort`, `max_tokens`/
 * `max_completion_tokens`, `top_k`, `seed`, `stop`, penalties and `logit_bias` change the
 * generated output, so two temp=0 requests that disagree on them must not share an entry
 * (an effort=none request would otherwise be served an effort=max replay, or vice versa).
 * These are stored under their raw request-body key only — no camelCase mirror exists,
 * because every caller passes `outputContractOf(body)` through untouched.
 */
export interface SignatureConstraints {
  toolChoice?: unknown;
  tools?: unknown;
  responseFormat?: unknown;
  tool_choice?: unknown;
  response_format?: unknown;
  text_format?: unknown;
  reasoning?: unknown;
  reasoning_effort?: unknown;
  max_tokens?: unknown;
  max_completion_tokens?: unknown;
  top_k?: unknown;
  seed?: unknown;
  stop?: unknown;
  presence_penalty?: unknown;
  frequency_penalty?: unknown;
  logit_bias?: unknown;
}

/**
 * Request-body fields that change what the model generates for the SAME conversation.
 * Listed verbatim as body keys (#15149); `!= null` keeps falsy-but-meaningful values
 * (seed 0, top_k 0, stop "").
 */
const GENERATION_PARAM_KEYS = [
  "reasoning",
  "reasoning_effort",
  "max_tokens",
  "max_completion_tokens",
  "top_k",
  "seed",
  "stop",
  "presence_penalty",
  "frequency_penalty",
  "logit_bias",
] as const;

/**
 * The parts of a request that decide what a *valid response* looks like.
 * Two calls that agree on the conversation but disagree here are not
 * interchangeable and must not share a cache entry (#12307): a request for
 * {color, wheels} must not be served a stored {value: "..."} body, and a
 * tool-calling request must not be served the body of one without tools.
 * Generation params (reasoning effort, max_tokens, …) count the same way (#15149).
 *
 * Returns null when the request carries none of these, so plain-chat
 * signatures — and every cache entry already written for them — are unchanged.
 */
export function outputContractOf(body: unknown): SignatureConstraints | null {
  const record = asRecord(body);
  const text = asRecord(record.text);
  const contract: SignatureConstraints = {};
  // Both spellings are set for each field so callers built against either the
  // pre-existing (#12734) camelCase constraints shape or this snake_case one
  // (matching the raw request body) can read the field they expect.
  if (record.response_format != null) {
    contract.response_format = record.response_format;
    contract.responseFormat = record.response_format;
  }
  if (text.format != null) contract.text_format = text.format;
  if (record.tools != null) contract.tools = record.tools;
  if (record.tool_choice != null) {
    contract.tool_choice = record.tool_choice;
    contract.toolChoice = record.tool_choice;
  }
  for (const key of GENERATION_PARAM_KEYS) {
    if (record[key] != null) contract[key] = record[key];
  }
  return Object.keys(contract).length > 0 ? contract : null;
}

/** Normalize a single tool definition, keeping only the fields that define its policy. */
function normalizeTool(tool: unknown): unknown {
  const record = asRecord(tool);
  const fn = asRecord(record.function);
  if (Object.keys(fn).length === 0 && Object.keys(record).length === 0) return tool;
  return {
    type: typeof record.type === "string" ? record.type : "function",
    function: {
      name: fn.name,
      description: fn.description,
      parameters: fn.parameters,
    },
  };
}

/**
 * Normalize `tools` for consistent hashing (mirrors `normalizeConversation` for messages):
 * strips volatile/irrelevant fields while keeping name/description/parameters, which are
 * what actually define the tool policy a cached response was generated under.
 */
function normalizeTools(tools: unknown): unknown {
  if (!Array.isArray(tools) || tools.length === 0) return undefined;
  return tools.map(normalizeTool);
}

/**
 * Pick the #15149 generation params off a constraints object, skipping null/undefined
 * so they drop out of `JSON.stringify` and a request without them hashes byte-identically
 * to the legacy payload.
 */
function pickGenerationParams(constraints?: SignatureConstraints | null): Record<string, unknown> {
  if (!constraints) return {};
  const picked: Record<string, unknown> = {};
  for (const key of GENERATION_PARAM_KEYS) {
    if (constraints[key] != null) picked[key] = constraints[key];
  }
  return picked;
}

/**
 * Generate deterministic cache signature from request params.
 * @param {string} model
 * @param {Array} messages - Normalized messages array
 * @param {number} temperature
 * @param {number} topP
 * @param {string} [apiKeyId] - API key ID for per-key isolation (prevents cross-user cache hits)
 * @param {SignatureConstraints} [constraints] - tool_choice/tools/response_format (#12734)
 *   plus the Responses-API `text.format` spelling (#12307): these change model behavior
 *   and must not collide with a signature computed without them. Also carries the
 *   generation params (reasoning/max_tokens/top_k/seed/stop/penalties/logit_bias, #15149).
 * @returns {string} hex signature
 */
export function generateSignature(
  model,
  conversation,
  temperature = 0,
  topP = 1,
  apiKeyId?: string,
  constraints?: SignatureConstraints | null
) {
  const payload = JSON.stringify({
    model,
    messages: normalizeConversation(conversation),
    temperature,
    top_p: topP,
    tool_choice: constraints?.toolChoice ?? constraints?.tool_choice,
    tools: normalizeTools(constraints?.tools),
    response_format: constraints?.responseFormat ?? constraints?.response_format,
    text_format: constraints?.text_format,
    ...pickGenerationParams(constraints),
  });
  const digest = crypto.createHash("sha256").update(payload).digest("hex");
  // Per-key cache isolation (#3740) namespaces the signature with the apiKeyId as a
  // PLAINTEXT prefix instead of folding it into the digest. The apiKeyId is an internal
  // identifier used purely to namespace a deterministic cache key — not a stored
  // credential — so it must never flow into a hash (a salted/slow password hash would
  // also break the determinism the cache relies on). Keeping it out of the digest
  // preserves isolation + determinism and avoids the false-positive
  // CodeQL js/insufficient-password-hash on a cache signature.
  return apiKeyId ? `${apiKeyId}.${digest}` : digest;
}

function stringifyForSignature(value: unknown): string {
  if (typeof value === "string") return value;
  try {
    return JSON.stringify(value);
  } catch {
    return String(value);
  }
}

/**
 * Normalize conversation items for consistent hashing.
 * Supports both Chat Completions `messages[]` and Responses API `input[]`.
 */
function normalizeConversation(conversation: unknown) {
  if (typeof conversation === "string") {
    return [{ role: "user", content: conversation }];
  }
  if (!Array.isArray(conversation)) return [];

  return conversation.map((item: Record<string, unknown>) => ({
    role: typeof item?.role === "string" && item.role.trim().length > 0 ? item.role : "user",
    content: stringifyForSignature(item?.content),
    ...conversationSignatureFields(item),
  }));
}

// ─── 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) {
    incrementMetric("hits");
    incrementMetric("tokens_saved", 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 > ?"
      )
      .get(signature, isoNow());

    if (row) {
      const record = asRecord(row);
      const responsePayload = typeof record.response === "string" ? record.response : null;
      if (!responsePayload) {
        incrementMetric("misses");
        return null;
      }
      const parsed = JSON.parse(responsePayload);
      const tokensSaved = toNumber(record.tokens_saved, 0);
      // Promote to memory cache
      getMemoryCache().set(signature, {
        response: parsed,
        tokensSaved,
      });
      // Update hit count in DB
      db.prepare("UPDATE semantic_cache SET hit_count = hit_count + 1 WHERE signature = ?").run(
        signature
      );

      incrementMetric("hits");
      incrementMetric("tokens_saved", tokensSaved);
      return parsed;
    }
  } catch {
    // DB not available — fail open
  }

  incrementMetric("misses");
  return null;
}

/**
 * Record a semantic cache hit: increments hit count for the entry in SQLite
 * and increments global hit metrics (hits and tokens_saved).
 */
export function recordSemanticCacheHit(signature: string, tokensSaved = 0): void {
  try {
    const db = getDbInstance();
    if (signature) {
      db.prepare(
        "UPDATE semantic_cache SET hit_count = hit_count + 1 WHERE signature = ? OR prompt_hash = ?"
      ).run(signature, signature.slice(0, 16));
    }
    incrementMetric("hits");
    if (tokensSaved > 0) {
      incrementMetric("tokens_saved", tokensSaved);
    }
  } catch {
    // DB not available — fail open
  }
}

/**
 * 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
  }
}

/**
 * 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;
  }
}

/**
 * Clear all cache entries.
 */
export function clearCache(): number {
  getMemoryCache().clear();
  let removed = 0;
  try {
    const db = getDbInstance();
    const result = db.prepare("DELETE FROM semantic_cache").run();
    removed = result.changes || 0;
    db.prepare("UPDATE cache_metrics SET value = 0").run();
  } catch {
    // DB not available
  }
  return removed;
}

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 > ?")
      .get(isoNow());
    dbSize = toNumber(asRecord(row).count, 0);
  } catch {
    // DB not available
  }

  const hits = getMetricValue("hits");
  const misses = getMetricValue("misses");
  const tokensSaved = getMetricValue("tokens_saved");
  const total = hits + misses;

  return {
    memoryEntries: memStats.size,
    dbEntries: dbSize,
    hits,
    misses,
    hitRate: total > 0 ? ((hits / total) * 100).toFixed(1) : "0.0",
    tokensSaved,
  };
}

/**
 * Check if a cached response can be served for this request.
 * Works for both streaming and non-streaming requests (cache hit returns JSON).
 * Requires explicit numeric `temperature: 0` — omitted temperature is NOT cached
 * because the provider default may be non-deterministic (e.g. random/creative tasks).
 */
export function isCacheableForRead(body, headers) {
  if ((getHeaderValue(headers, "x-omniroute-no-cache") || "").toLowerCase() === "true") {
    return false;
  }
  const cacheControl = (getHeaderValue(headers, "cache-control") || "").toLowerCase();
  if (cacheControl.includes("no-cache")) {
    return false;
  }
  if (typeof body.temperature !== "number" || body.temperature !== 0) return false;
  return true;
}

/**
 * Check if a response should be stored in cache after completion.
 * Works for both streaming and non-streaming responses.
 * Requires explicit `temperature: 0` — omitted temperature is NOT cacheable
 * because the provider default may be non-deterministic.
 */
export function isCacheableForWrite(body, headers) {
  if ((getHeaderValue(headers, "x-omniroute-no-cache") || "").toLowerCase() === "true") {
    return false;
  }
  const cacheControl = (getHeaderValue(headers, "cache-control") || "").toLowerCase();
  if (cacheControl.includes("no-cache")) {
    return false;
  }
  if (body.temperature !== 0) return false;
  return true;
}

/**
 * A response cut short by the output-token ceiling is a partial answer, not a
 * reusable one. Caching it under a temperature:0 signature pins the truncation
 * for every later identical request — the caller sees a mid-sentence reply that
 * no retry clears, because each retry is served the same poisoned entry.
 *
 * Only `length` (and its Claude-side spelling `max_tokens`) is treated as
 * truncation. `stop`, `tool_calls`, and a missing/unknown reason are complete
 * responses and stay cacheable, so this never narrows the cache beyond the bug.
 */
const TRUNCATED_FINISH_REASONS = new Set(["length", "max_tokens"]);

export function isTruncatedCompletion(response: unknown): boolean {
  if (!response || typeof response !== "object") return false;
  const r = response as {
    choices?: Array<{ finish_reason?: unknown }>;
    stop_reason?: unknown;
  };
  if (Array.isArray(r.choices)) {
    for (const choice of r.choices) {
      const reason = choice?.finish_reason;
      if (typeof reason === "string" && TRUNCATED_FINISH_REASONS.has(reason)) return true;
    }
  }
  // Claude-format responses carry the reason at the top level instead.
  if (typeof r.stop_reason === "string" && TRUNCATED_FINISH_REASONS.has(r.stop_reason)) {
    return true;
  }
  return false;
}

/**
 * Streaming variant: the assembled SSE body is scanned for a truncating
 * finish_reason. Parsing is intentionally tolerant — an unparseable chunk is
 * treated as "not known to be truncated" so a malformed frame never silently
 * disables caching.
 */
export function isTruncatedStreamBody(streamBody: unknown): boolean {
  // chatCore hands the streaming store the *assembled* body (an object with
  // `choices[].finish_reason`), not raw SSE text — so the object shape must be
  // checked too or the streaming guard is a no-op in production (#14159).
  if (streamBody && typeof streamBody === "object") return isTruncatedCompletion(streamBody);
  if (typeof streamBody !== "string" || streamBody.length === 0) return false;
  for (const line of streamBody.split("\n")) {
    const trimmed = line.trim();
    if (!trimmed.startsWith("data:")) continue;
    const payload = trimmed.slice(5).trim();
    if (!payload || payload === "[DONE]") continue;
    try {
      if (isTruncatedCompletion(JSON.parse(payload))) return true;
    } catch {
      // Non-JSON frame — ignore.
    }
  }
  return false;
}