@denizsvnc/ai-rag-engine 2.2.1 → 2.2.3

package/CHANGELOG.md

@@ -2,6 +2,37 @@
 
 This file follows Semantic Versioning.
 
+## 2.2.3 — 2026-11-10
+
+### Fixed
+
+- **Transformers.js Concurrency Lock:** Implemented static Promise lock chain (`acquireInitLock`) in `SMLEngine` to serialize `init()` calls across independent instances and isolate mutable Transformers.js global `env`. Model inference remains fully concurrent.
+- **LLM Adapter Constructor Flexibility:** Dual constructor signatures added to `GeminiAdapter`, `OpenAIAdapter`, and `OpenRouterAdapter` accepting both `{ apiKey, model, temperature }` configuration objects and positional parameters.
+- **Document Chunking Bounds & Streaming UTF-8 Decoder:** Enforced `RangeError` on invalid chunk boundaries (`chunkSize > 0`, `overlap < chunkSize`). Added `StringDecoder('utf-8')` to prevent split multi-byte characters (e.g. Turkish `ş`, `ğ`) across stream chunks.
+- **Native zlib PDF FlateDecode Ingestion:** Integrated Node's native `zlib.inflateSync` & `inflateRawSync` for zero-dependency decompression of vector PDF streams.
+- **RAGEvaluator Heuristics & Polarity:** Ensured empty answers score 0.0; deduplicated retrieved IDs; added English & Turkish negation and polarity contradiction checks.
+
+### Added
+
+- **Bun Runtime Compatibility:** Dedicated Bun test suite (`test/bun/bun-compat.test.ts`) and npm script `npm run test:bun` verifying ONNX embeddings, PII masking, vector math, and RAG execution under Bun v1.4.4+.
+- **CI Workflow Hardening:** Added automated Bun compatibility verification job in GitHub Actions CI workflow.
+- **Engineering & Security Audits:** Comprehensive audit documentation covering Baseline, Architecture, Security (OWASP Top 10 LLM), Invariants, Performance Benchmarks, RAG Evaluation, and Node/Bun Compatibility in `docs/audits/`.
+
+## 2.2.2 — 2026-11-10
+
+### Fixed
+
+- **Document Ingestion:** Resolved infinite decaying chunk loop in `DocumentIngestionService.chunkText` when `endIndex >= text.length`.
+- **Security & Prompt Injection:** Resolved regex boundary and preposition issues in `InjectionDetector` for English DAN jailbreak attacks (`in DAN mode`) and Turkish Unicode word boundaries (`ş, ç`).
+- **LLM Failover Logging:** Replaced generic `Object failed` error message in `FallbackLLMProvider` with dynamic provider, model, or constructor resolution.
+- **Rules Adapter:** Extended `StaticRulesAdapter` to accept structured rule objects (`{ id, rule }`) alongside plain string arrays via polymorphic `RuleInput`.
+- **Performance & SML Sharing:** Prevented redundant ONNX model downloads and duplicate memory footprint by sharing `SMLEngine` instances across `RAGEngine`, `LocalFaqAdapter`, and `MemoryVectorAdapter`.
+
+### Added
+
+- Exported `JobStatusValues` constants in `job-queue.interface.ts` to prevent casing mismatches.
+- Exported `RuleInput` type in root exports.
+
 ## 2.2.1 — 2026-11-10
 
 ### Documentation

package/dist/adapters/faq/local-faq.adapter.d.ts

@@ -16,6 +16,8 @@ export declare class LocalFaqAdapter implements IFAQAdapter {
     private items;
     private isLoaded;
     constructor(options?: LocalFaqAdapterOptions);
+    setSMLEngine(engine: SMLEngine): void;
+    getSMLEngine(): SMLEngine;
     /**
      * Initializes the adapter by reading the JSON file and caching embeddings in RAM.
      */

package/dist/adapters/llm/fallback.adapter.d.ts

@@ -6,5 +6,6 @@ export declare class FallbackLLMProvider implements ILLMProvider {
     private providers;
     private logFailover;
     constructor(providers: ILLMProvider[], options?: FallbackProviderOptions);
+    private getProviderName;
     generate(request: ILLMRequest): Promise<ILLMResponse>;
 }

package/dist/adapters/llm/gemini.adapter.d.ts

@@ -1,8 +1,17 @@
 import type { ILLMProvider, ILLMRequest, ILLMResponse } from '../../interfaces/llm.interface.js';
+export interface GeminiConfig {
+    /** Google Gemini API Key. Defaults to process.env.GEMINI_API_KEY. */
+    apiKey?: string | undefined;
+    /** Model identifier (e.g. 'gemini-2.0-flash', 'gemini-1.5-flash', 'gemini-1.5-pro'). Defaults to process.env.GEMINI_MODEL || 'gemini-2.0-flash'. */
+    model?: string | undefined;
+    /** Sampling temperature. Defaults to 0.1 for deterministic RAG answers. */
+    temperature?: number | undefined;
+}
 export declare class GeminiAdapter implements ILLMProvider {
     private apiKey;
     private model;
-    constructor(apiKey?: string, model?: string);
+    private temperature;
+    constructor(configOrApiKey?: string | GeminiConfig, model?: string);
     static createFast(apiKey?: string): GeminiAdapter;
     static createSmart(apiKey?: string): GeminiAdapter;
     generate(request: ILLMRequest): Promise<ILLMResponse>;

package/dist/adapters/llm/openai.adapter.d.ts

@@ -1,9 +1,20 @@
 import type { ILLMProvider, ILLMRequest, ILLMResponse } from '../../interfaces/llm.interface.js';
+export interface OpenAIConfig {
+    /** OpenAI API Key. Defaults to process.env.OPENAI_API_KEY. */
+    apiKey?: string | undefined;
+    /** Model identifier (e.g. 'gpt-4o-mini', 'gpt-4o'). Defaults to process.env.OPENAI_MODEL || 'gpt-4o-mini'. */
+    model?: string | undefined;
+    /** Base API endpoint. Defaults to 'https://api.openai.com/v1/chat/completions'. */
+    endpoint?: string | undefined;
+    /** Sampling temperature. Defaults to 0.1 for deterministic RAG answers. */
+    temperature?: number | undefined;
+}
 export declare class OpenAIAdapter implements ILLMProvider {
     protected apiKey: string;
     protected model: string;
     protected endpoint: string;
-    constructor(apiKey?: string, model?: string, endpoint?: string);
+    protected temperature: number;
+    constructor(configOrApiKey?: string | OpenAIConfig, model?: string, endpoint?: string);
     static createFast(apiKey?: string): OpenAIAdapter;
     static createSmart(apiKey?: string): OpenAIAdapter;
     generate(request: ILLMRequest): Promise<ILLMResponse>;

package/dist/adapters/llm/openrouter.adapter.d.ts

@@ -1,6 +1,8 @@
-import { OpenAIAdapter } from './openai.adapter.js';
+import { OpenAIAdapter, type OpenAIConfig } from './openai.adapter.js';
+export interface OpenRouterConfig extends OpenAIConfig {
+}
 export declare class OpenRouterAdapter extends OpenAIAdapter {
-    constructor(apiKey?: string, model?: string, endpoint?: string);
+    constructor(configOrApiKey?: string | OpenRouterConfig, model?: string, endpoint?: string);
     static createFast(apiKey?: string): OpenRouterAdapter;
     static createSmart(apiKey?: string): OpenRouterAdapter;
 }

package/dist/adapters/rules/static-rules.adapter.d.ts

@@ -1,6 +1,11 @@
 import type { IRulesAdapter } from '../../interfaces/rules.interface.js';
+export type RuleInput = string | {
+    id?: string;
+    rule: string;
+    [key: string]: any;
+};
 export declare class StaticRulesAdapter implements IRulesAdapter {
     private rules;
-    constructor(rules: string[]);
+    constructor(rules?: Array<RuleInput>);
     getActiveRules(tenantId?: string): Promise<string[]>;
 }

package/dist/adapters/vector/memory-vector.adapter.d.ts

@@ -6,6 +6,8 @@ export declare class MemoryVectorAdapter implements IVectorStoreAdapter {
     private documents;
     private bm25Index;
     constructor(smlEngine?: SMLEngine);
+    setSMLEngine(engine: SMLEngine): void;
+    getSMLEngine(): SMLEngine;
     /**
      * Adds documents to memory, computes their embeddings, and indexes them in BM25.
      */