@denizsvnc/ai-rag-engine 2.2.1 → 2.2.3

package/dist/core/sml-engine.d.ts

@@ -11,17 +11,53 @@ export interface EmbeddingExtractor {
 export declare class CircuitBreakerOpenError extends Error {
     constructor(message: string);
 }
+/**
+ * Industry-standard ONNX embedding model presets for the Small Machine Learning (SML) engine.
+ * Powered by `@huggingface/transformers` without requiring Python or external API keys.
+ */
+export declare const SMLModelPresets: {
+    /**
+     * Ultra-fast English embedding model (~30MB q8 ONNX).
+     * Ideal for English-only workloads, low-latency semantic routing, and minimal RAM footprint.
+     */
+    readonly ENGLISH_FAST: 'Xenova/all-MiniLM-L6-v2';
+    /**
+     * Highly recommended balanced multilingual model (~117MB q8 ONNX).
+     * Provides exceptional semantic similarity across 50+ languages including Turkish, German, English, French, and Spanish.
+     */
+    readonly MULTILINGUAL: 'Xenova/paraphrase-multilingual-MiniLM-L12-v2';
+    /**
+     * Heavyweight high-accuracy multilingual foundation model (1024-dim, long context, 100+ languages).
+     * Highest precision across mixed-language corpora, but requires higher memory and CPU allocation.
+     */
+    readonly MULTILINGUAL_ACCURATE: 'Xenova/bge-m3';
+};
+export type SMLModelPreset = typeof SMLModelPresets[keyof typeof SMLModelPresets];
 export interface SMLEngineConfig {
+    /** Directory where downloaded ONNX models are cached. Defaults to process.env.RAG_MODEL_CACHE_DIR or './data/models'. */
     cacheDir?: string | undefined;
+    /** When true, only loads models already present on local disk and prevents external network fetching. */
     localFilesOnly?: boolean | undefined;
-    modelName?: string | undefined;
+    /**
+     * Hugging Face ONNX model identifier.
+     * Use a preset (e.g. `SMLModelPresets.MULTILINGUAL`) or pass any custom HuggingFace model string (e.g. `'my-org/my-custom-model'`).
+     * Defaults to `'Xenova/all-MiniLM-L6-v2'`.
+     */
+    modelName?: string | SMLModelPreset | undefined;
+    /** Maximum number of precomputed query/document embeddings retained in RAM LRU cache. Defaults to 1000. */
     cacheSize?: number | undefined;
+    /** Model quantization data type: 'q8' (default), 'fp32', 'fp16', or 'q4'. */
     dtype?: 'q8' | 'fp32' | 'fp16' | 'q4' | undefined;
+    /** Legacy flag for quantization: false forces 'fp32'. */
     quantized?: boolean | undefined;
+    /** Number of consecutive load failures before tripping the circuit breaker OPEN. Defaults to 3. */
     circuitBreakerThreshold?: number | undefined;
+    /** Cooldown period in milliseconds before re-testing model recovery. Defaults to 15000 (15s). */
     circuitBreakerCooldownMs?: number | undefined;
 }
 export declare class SMLEngine {
+    private static initLockChain;
+    private static acquireInitLock;
     private readonly modelName;
     private readonly cacheDir;
     private readonly localFilesOnly;

package/dist/index.d.ts

@@ -8,7 +8,7 @@ export * from './interfaces/cache.interface.js';
 export * from './interfaces/llm.interface.js';
 export * from './core/security/security.types.js';
 export { RAGEngine } from './core/engine.js';
-export { SMLEngine, type SMLEngineConfig, type EmbeddingExtractor, CircuitBreakerOpenError } from './core/sml-engine.js';
+export { SMLEngine, SMLModelPresets, type SMLModelPreset, type SMLEngineConfig, type EmbeddingExtractor, CircuitBreakerOpenError } from './core/sml-engine.js';
 export { ToolRegistry } from './core/tool-registry.js';
 export { CacheManager } from './core/cache-manager.js';
 export { PromptBuilder, type PromptContext } from './core/prompt-builder.js';
@@ -31,15 +31,15 @@ export { RedisCacheAdapter, type IRedisClientLike, type RedisCacheOptions } from
 export { createRagMiddleware } from './api/middleware.js';
 export { createApiRoutes, type ApiRouteOptions } from './api/routes.js';
 export { createAuthToken, verifyAuthToken, createAuthMiddleware, requireRoles, type AuthTokenPayload, type AuthMiddlewareOptions } from './api/auth.js';
-export { OpenAIAdapter } from './adapters/llm/openai.adapter.js';
-export { GeminiAdapter } from './adapters/llm/gemini.adapter.js';
-export { OpenRouterAdapter } from './adapters/llm/openrouter.adapter.js';
+export { OpenAIAdapter, type OpenAIConfig } from './adapters/llm/openai.adapter.js';
+export { GeminiAdapter, type GeminiConfig } from './adapters/llm/gemini.adapter.js';
+export { OpenRouterAdapter, type OpenRouterConfig } from './adapters/llm/openrouter.adapter.js';
 export { CustomLLMAdapter, type CustomLLMConfig } from './adapters/llm/custom.adapter.js';
 export { OllamaAdapter, type OllamaConfig } from './adapters/llm/ollama.adapter.js';
 export { FallbackLLMProvider, type FallbackProviderOptions } from './adapters/llm/fallback.adapter.js';
 export { TieredLLMProvider, type TieredLLMConfig } from './adapters/llm/tiered.adapter.js';
 export { MemoryVectorAdapter } from './adapters/vector/memory-vector.adapter.js';
-export { StaticRulesAdapter } from './adapters/rules/static-rules.adapter.js';
+export { StaticRulesAdapter, type RuleInput } from './adapters/rules/static-rules.adapter.js';
 export { UrlFaqAdapter } from './adapters/faq/url-faq.adapter.js';
 export { LocalFaqAdapter, type LocalFaqAdapterOptions } from './adapters/faq/local-faq.adapter.js';
 export { RedisConfirmationNonceStore, MemoryConfirmationNonceStore, type ConfirmationNonceStore } from './core/security/nonce-store.js';