@juspay/neurolink 9.68.9 → 9.68.11

package/dist/lib/providers/huggingFace.d.ts

@@ -1,79 +1,31 @@
-import type { ZodType } from "zod";
 import type { AIProviderName } from "../constants/enums.js";
-import { BaseProvider } from "../core/baseProvider.js";
-import type { NeurolinkCredentials, StreamOptions, StreamResult } from "../types/index.js";
-import type { LanguageModel, Schema } from "../types/index.js";
+import type { NeurolinkCredentials } from "../types/index.js";
+import { OpenAIChatCompletionsProvider } from "./openaiChatCompletionsBase.js";
 /**
- * HuggingFace Provider - BaseProvider Implementation
- * Using AI SDK with HuggingFace's OpenAI-compatible endpoint
+ * HuggingFace Provider — direct HTTP, no AI SDK.
+ *
+ * OpenAI-compatible chat completions at router.huggingface.co/v1 (unified
+ * router endpoint, 2025). Supports the full HuggingFace model hub including
+ * Llama 3.x, Qwen 2.5, Mistral, DeepSeek, and tool-calling capable variants.
+ * All request/stream/tool-loop orchestration lives in
+ * `OpenAIChatCompletionsProvider`; this class only declares configuration
+ * and provider-specific error mapping.
+ *
+ * @see https://huggingface.co/docs/api-inference/index
  */
-export declare class HuggingFaceProvider extends BaseProvider {
-    private model;
-    constructor(modelName?: string, _sdk?: unknown, credentials?: NeurolinkCredentials["huggingFace"]);
+export declare class HuggingFaceProvider extends OpenAIChatCompletionsProvider {
+    constructor(modelName?: string, sdk?: unknown, credentials?: NeurolinkCredentials["huggingFace"]);
+    protected getProviderName(): AIProviderName;
+    protected getDefaultModel(): string;
+    protected getFallbackModelName(): string;
     /**
-     * HuggingFace Tool Calling Support (Enhanced 2025)
-     *
-     * **Supported Models (Tool Calling Enabled):**
-     * - meta-llama/Llama-3.1-8B-Instruct - Post-trained for tool calling
-     * - meta-llama/Llama-3.1-70B-Instruct - Advanced tool calling capabilities
-     * - meta-llama/Llama-3.1-405B-Instruct - Full tool calling support
-     * - nvidia/Llama-3.1-Nemotron-Ultra-253B-v1 - Optimized for tool calling
-     * - NousResearch/Hermes-3-Llama-3.2-3B - Function calling trained
-     * - codellama/CodeLlama-34b-Instruct-hf - Code-focused tool calling
-     * - mistralai/Mistral-7B-Instruct-v0.3 - Basic tool support
-     *
-     * **Unsupported Models (Tool Calling Disabled):**
-     * - microsoft/DialoGPT-* - Treats tools as conversation context
-     * - gpt2, bert, roberta variants - No tool calling training
-     * - Most pre-2024 models - Limited function calling capabilities
-     *
-     * **Implementation Details:**
-     * - Intelligent model detection based on known capabilities
-     * - Custom tool schema formatting for HuggingFace models
-     * - Enhanced response parsing for function call extraction
-     * - Graceful fallback for unsupported models
-     *
-     * @returns true for supported models, false for unsupported models
+     * HuggingFace serves a huge variety of models, many of which reject the
+     * OpenAI `tools` field. The base reports supportsTools() === true
+     * unconditionally, which would merge tools into requests for models
+     * (including the default DialoGPT-medium) that don't accept them. Preserve
+     * the pre-migration allowlist: only known tool-calling-capable model
+     * families opt in; everything else runs tool-free.
      */
     supportsTools(): boolean;
-    protected executeStream(options: StreamOptions, analysisSchema?: ZodType | Schema<unknown>): Promise<StreamResult>;
-    /**
-     * Prepare stream options with HuggingFace-specific enhancements
-     * Handles tool calling optimizations and model-specific formatting
-     */
-    private prepareStreamOptions;
-    /**
-     * Enhance system prompt with tool calling instructions for HuggingFace models
-     * Many HF models benefit from explicit tool calling guidance
-     */
-    private enhanceSystemPromptForTools;
-    /**
-     * Format tools for HuggingFace model compatibility
-     * Some models require specific tool schema formatting
-     */
-    private formatToolsForHuggingFace;
-    /**
-     * Get recommendations for tool-calling capable HuggingFace models
-     * Provides guidance for users who want to use function calling
-     */
-    static getToolCallingRecommendations(): {
-        recommended: string[];
-        performance: Record<string, {
-            speed: number;
-            quality: number;
-            cost: number;
-        }>;
-        notes: Record<string, string>;
-    };
-    /**
-     * Enhanced error handling with HuggingFace-specific guidance
-     */
-    formatProviderError(error: unknown): Error;
-    getProviderName(): AIProviderName;
-    getDefaultModel(): string;
-    /**
-     * Returns the Vercel AI SDK model instance for HuggingFace
-     */
-    getAISDKModel(): LanguageModel;
+    protected formatProviderError(error: unknown): Error;
 }
-export default HuggingFaceProvider;

package/dist/lib/providers/huggingFace.js

@@ -1,311 +1,77 @@
-import { createOpenAI } from "@ai-sdk/openai";
-import { BaseProvider } from "../core/baseProvider.js";
-import { DEFAULT_MAX_STEPS } from "../core/constants.js";
-import { createProxyFetch } from "../proxy/proxyFetch.js";
 import { AuthenticationError, InvalidModelError, NetworkError, ProviderError, RateLimitError, } from "../types/index.js";
-import { emitToolEndFromStepFinish } from "../utils/toolEndEmitter.js";
 import { logger } from "../utils/logger.js";
-import { buildNoOutputSentinel, detectPostStreamNoOutput, stampNoOutputSpan, } from "../utils/noOutputSentinel.js";
 import { createHuggingFaceConfig, getProviderModel, validateApiKey, } from "../utils/providerConfig.js";
-import { composeAbortSignals, createTimeoutController, TimeoutError, } from "../utils/timeout.js";
-import { resolveToolChoice } from "../utils/toolChoice.js";
-import { NoOutputGeneratedError } from "../utils/generationErrors.js";
-import { stepCountIs } from "../utils/tool.js";
-import { streamText } from "../utils/generation.js";
-// Configuration helpers - now using consolidated utility
-const getHuggingFaceApiKey = () => {
-    return validateApiKey(createHuggingFaceConfig());
-};
-const getDefaultHuggingFaceModel = () => {
-    return getProviderModel("HUGGINGFACE_MODEL", "microsoft/DialoGPT-medium");
-};
-// Note: hasNeurolinkCredentials["huggingFace"] now directly imported from consolidated utility
+import { TimeoutError } from "../utils/timeout.js";
+import { OpenAIChatCompletionsProvider } from "./openaiChatCompletionsBase.js";
+const HUGGINGFACE_DEFAULT_BASE_URL = "https://router.huggingface.co/v1";
+const getHuggingFaceApiKey = () => validateApiKey(createHuggingFaceConfig());
+const getDefaultHuggingFaceModel = () => getProviderModel("HUGGINGFACE_MODEL", "microsoft/DialoGPT-medium");
 /**
- * HuggingFace Provider - BaseProvider Implementation
- * Using AI SDK with HuggingFace's OpenAI-compatible endpoint
+ * HuggingFace Provider — direct HTTP, no AI SDK.
+ *
+ * OpenAI-compatible chat completions at router.huggingface.co/v1 (unified
+ * router endpoint, 2025). Supports the full HuggingFace model hub including
+ * Llama 3.x, Qwen 2.5, Mistral, DeepSeek, and tool-calling capable variants.
+ * All request/stream/tool-loop orchestration lives in
+ * `OpenAIChatCompletionsProvider`; this class only declares configuration
+ * and provider-specific error mapping.
+ *
+ * @see https://huggingface.co/docs/api-inference/index
  */
-export class HuggingFaceProvider extends BaseProvider {
-    model;
-    constructor(modelName, _sdk, credentials) {
-        super(modelName, "huggingface");
-        // Get API key and validate
-        const apiKey = credentials?.apiKey ?? getHuggingFaceApiKey();
-        // Create HuggingFace provider using unified router endpoint (2025) with proxy support
-        const huggingface = createOpenAI({
-            apiKey: apiKey,
-            baseURL: credentials?.baseURL ?? "https://router.huggingface.co/v1",
-            fetch: createProxyFetch(),
-        });
-        // Initialize model
-        this.model = huggingface(this.modelName);
+export class HuggingFaceProvider extends OpenAIChatCompletionsProvider {
+    constructor(modelName, sdk, credentials) {
+        const apiKey = credentials?.apiKey?.trim()
+            ? credentials.apiKey.trim()
+            : getHuggingFaceApiKey();
+        // Treat blank/whitespace overrides as unset so an empty
+        // `credentials.baseURL` or `HUGGINGFACE_BASE_URL=` cannot override the
+        // default with "" (mirrors the apiKey precedence above).
+        const baseURL = credentials?.baseURL?.trim() ||
+            process.env.HUGGINGFACE_BASE_URL?.trim() ||
+            HUGGINGFACE_DEFAULT_BASE_URL;
+        super("huggingface", modelName, sdk, { baseURL, apiKey });
         logger.debug("HuggingFaceProvider initialized", {
-            model: this.modelName,
-            provider: this.providerName,
+            modelName: this.modelName,
+            providerName: this.providerName,
+            baseURL: this.config.baseURL,
         });
     }
-    // ===================
-    // ABSTRACT METHOD IMPLEMENTATIONS
-    // ===================
+    getProviderName() {
+        return "huggingface";
+    }
+    getDefaultModel() {
+        return getDefaultHuggingFaceModel();
+    }
+    getFallbackModelName() {
+        return "meta-llama/Llama-3.1-8B-Instruct";
+    }
     /**
-     * HuggingFace Tool Calling Support (Enhanced 2025)
-     *
-     * **Supported Models (Tool Calling Enabled):**
-     * - meta-llama/Llama-3.1-8B-Instruct - Post-trained for tool calling
-     * - meta-llama/Llama-3.1-70B-Instruct - Advanced tool calling capabilities
-     * - meta-llama/Llama-3.1-405B-Instruct - Full tool calling support
-     * - nvidia/Llama-3.1-Nemotron-Ultra-253B-v1 - Optimized for tool calling
-     * - NousResearch/Hermes-3-Llama-3.2-3B - Function calling trained
-     * - codellama/CodeLlama-34b-Instruct-hf - Code-focused tool calling
-     * - mistralai/Mistral-7B-Instruct-v0.3 - Basic tool support
-     *
-     * **Unsupported Models (Tool Calling Disabled):**
-     * - microsoft/DialoGPT-* - Treats tools as conversation context
-     * - gpt2, bert, roberta variants - No tool calling training
-     * - Most pre-2024 models - Limited function calling capabilities
-     *
-     * **Implementation Details:**
-     * - Intelligent model detection based on known capabilities
-     * - Custom tool schema formatting for HuggingFace models
-     * - Enhanced response parsing for function call extraction
-     * - Graceful fallback for unsupported models
-     *
-     * @returns true for supported models, false for unsupported models
+     * HuggingFace serves a huge variety of models, many of which reject the
+     * OpenAI `tools` field. The base reports supportsTools() === true
+     * unconditionally, which would merge tools into requests for models
+     * (including the default DialoGPT-medium) that don't accept them. Preserve
+     * the pre-migration allowlist: only known tool-calling-capable model
+     * families opt in; everything else runs tool-free.
      */
     supportsTools() {
         const modelName = this.modelName.toLowerCase();
-        // Check if model is in the list of known tool-calling capable models
         const toolCapableModels = [
-            // Llama 3.1 series (post-trained for tool calling)
             "llama-3.1-8b-instruct",
             "llama-3.1-70b-instruct",
             "llama-3.1-405b-instruct",
             "llama-3.1-nemotron-ultra",
-            // Hermes series (function calling trained)
             "hermes-3-llama-3.2",
             "hermes-2-pro",
-            // Code Llama (code-focused tool calling)
             "codellama-34b-instruct",
             "codellama-13b-instruct",
-            // Mistral series (basic tool support)
             "mistral-7b-instruct-v0.3",
             "mistral-8x7b-instruct",
-            // Other known tool-capable models
             "nous-hermes",
             "openchat",
             "wizardcoder",
         ];
-        // Check if current model matches tool-capable model patterns
-        const isToolCapable = toolCapableModels.some((capableModel) => modelName.includes(capableModel));
-        if (isToolCapable) {
-            logger.debug("HuggingFace tool calling enabled", {
-                model: this.modelName,
-                reason: "Model supports function calling",
-            });
-            return true;
-        }
-        // Log why tools are disabled for transparency
-        logger.debug("HuggingFace tool calling disabled", {
-            model: this.modelName,
-            reason: "Model not in tool-capable list",
-            suggestion: "Consider using Llama-3.1-* or Hermes-3-* models for tool calling",
-        });
-        return false;
-    }
-    // executeGenerate removed - BaseProvider handles all generation with tools
-    async executeStream(options, analysisSchema) {
-        this.validateStreamOptions(options);
-        const timeout = this.getTimeout(options);
-        const timeoutController = createTimeoutController(timeout, this.providerName, "stream");
-        try {
-            // Get tools - options.tools is pre-merged by BaseProvider.stream()
-            const shouldUseTools = !options.disableTools && this.supportsTools();
-            const allTools = shouldUseTools
-                ? options.tools || (await this.getAllTools())
-                : {};
-            // Enhanced tool handling for HuggingFace models
-            const streamOptions = this.prepareStreamOptions(options, analysisSchema);
-            // Build message array from options with multimodal support
-            // Using protected helper from BaseProvider to eliminate code duplication
-            // Pass the enhanced system prompt (with tool-calling instructions) so it
-            // actually reaches the model instead of being silently discarded.
-            const messagesOptions = streamOptions.system
-                ? { ...options, systemPrompt: streamOptions.system }
-                : options;
-            const messages = await this.buildMessagesForStream(messagesOptions);
-            // Reviewer follow-up: capture upstream provider errors via onError
-            // so the post-stream NoOutput detect can propagate the real cause
-            // into the sentinel's providerError / modelResponseRaw.
-            let capturedProviderError;
-            const result = await streamText({
-                model: this.model,
-                messages: messages,
-                temperature: options.temperature,
-                maxOutputTokens: options.maxTokens, // No default limit - unlimited unless specified
-                stopWhen: stepCountIs(options.maxSteps || DEFAULT_MAX_STEPS),
-                tools: (shouldUseTools
-                    ? streamOptions.tools || allTools
-                    : {}),
-                toolChoice: resolveToolChoice(options, (shouldUseTools ? streamOptions.tools || allTools : {}), shouldUseTools),
-                abortSignal: composeAbortSignals(options.abortSignal, timeoutController?.controller.signal),
-                experimental_telemetry: this.telemetryHandler.getTelemetryConfig(options),
-                experimental_repairToolCall: this.getToolCallRepairFn(options),
-                onError: (event) => {
-                    capturedProviderError = event.error;
-                    logger.error("HuggingFace: Stream error", {
-                        error: event.error instanceof Error
-                            ? event.error.message
-                            : String(event.error),
-                    });
-                },
-                onStepFinish: ({ toolCalls, toolResults }) => {
-                    emitToolEndFromStepFinish(this.neurolink?.getEventEmitter(), toolResults);
-                    this.handleToolExecutionStorage(toolCalls, toolResults, options, new Date()).catch((error) => {
-                        logger.warn("[HuggingFaceProvider] Failed to store tool executions", {
-                            provider: this.providerName,
-                            error: error instanceof Error ? error.message : String(error),
-                        });
-                    });
-                },
-            });
-            timeoutController?.cleanup();
-            // Transform stream to match StreamResult interface with enhanced tool call parsing
-            const transformedStream = async function* () {
-                let chunkCount = 0;
-                try {
-                    for await (const chunk of result.textStream) {
-                        chunkCount++;
-                        yield { content: chunk };
-                    }
-                }
-                catch (streamError) {
-                    if (NoOutputGeneratedError.isInstance(streamError)) {
-                        logger.warn("HuggingFace: Stream produced no output (NoOutputGeneratedError) — caught from textStream");
-                        const sentinel = await buildNoOutputSentinel(streamError, result, capturedProviderError);
-                        stampNoOutputSpan(sentinel);
-                        yield sentinel;
-                        return;
-                    }
-                    throw streamError;
-                }
-                // Curator P3-6 (round-2 fix): production trigger comes through
-                // the result.finishReason rejection, not textStream throws.
-                if (chunkCount === 0) {
-                    const detected = await detectPostStreamNoOutput(result, capturedProviderError);
-                    if (detected) {
-                        logger.warn("HuggingFace: Stream produced no output (NoOutputGeneratedError) — caught from finishReason rejection");
-                        stampNoOutputSpan(detected.sentinel);
-                        yield detected.sentinel;
-                    }
-                }
-            };
-            return {
-                stream: transformedStream(),
-                provider: this.providerName,
-                model: this.modelName,
-            };
-        }
-        catch (error) {
-            timeoutController?.cleanup();
-            throw this.handleProviderError(error);
-        }
-    }
-    /**
-     * Prepare stream options with HuggingFace-specific enhancements
-     * Handles tool calling optimizations and model-specific formatting
-     */
-    prepareStreamOptions(options, _analysisSchema) {
-        const modelSupportsTools = this.supportsTools();
-        // If model doesn't support tools, disable them completely
-        if (!modelSupportsTools) {
-            return {
-                prompt: options.input.text ?? "",
-                system: options.systemPrompt,
-                tools: undefined,
-                toolChoice: undefined,
-            };
-        }
-        // For tool-capable models, enhance the prompt with tool calling instructions
-        const enhancedSystemPrompt = this.enhanceSystemPromptForTools(options.systemPrompt, options.tools);
-        // Format tools using HuggingFace-compatible schema if tools are provided
-        const formattedTools = options.tools
-            ? this.formatToolsForHuggingFace(options.tools)
-            : undefined;
-        return {
-            prompt: options.input.text ?? "",
-            system: enhancedSystemPrompt,
-            tools: formattedTools,
-            toolChoice: formattedTools ? (options.toolChoice ?? "auto") : undefined,
-        };
-    }
-    /**
-     * Enhance system prompt with tool calling instructions for HuggingFace models
-     * Many HF models benefit from explicit tool calling guidance
-     */
-    enhanceSystemPromptForTools(originalSystemPrompt, tools) {
-        if (!tools || !this.supportsTools()) {
-            return originalSystemPrompt || "";
-        }
-        const toolInstructions = `
-You have access to function tools. When you need to use a tool to answer the user's request:
-1. Identify the appropriate tool from the available functions
-2. Call the function with the correct parameters in JSON format
-3. Use the function results to provide a comprehensive answer
-
-Available tools will be provided in the function calling format. Use them when they can help answer the user's question.
-`;
-        return originalSystemPrompt
-            ? `${originalSystemPrompt}\n\n${toolInstructions}`
-            : toolInstructions;
-    }
-    /**
-     * Format tools for HuggingFace model compatibility
-     * Some models require specific tool schema formatting
-     */
-    formatToolsForHuggingFace(tools) {
-        // For now, pass through tools as-is since we're using OpenAI-compatible endpoint
-        // Future enhancement: Add model-specific tool formatting if needed
-        return tools;
+        return toolCapableModels.some((capable) => modelName.includes(capable));
     }
-    /**
-     * Get recommendations for tool-calling capable HuggingFace models
-     * Provides guidance for users who want to use function calling
-     */
-    static getToolCallingRecommendations() {
-        return {
-            recommended: [
-                "meta-llama/Llama-3.1-8B-Instruct",
-                "meta-llama/Llama-3.1-70B-Instruct",
-                "nvidia/Llama-3.1-Nemotron-Ultra-253B-v1",
-                "NousResearch/Hermes-3-Llama-3.2-3B",
-                "codellama/CodeLlama-34b-Instruct-hf",
-            ],
-            performance: {
-                "meta-llama/Llama-3.1-8B-Instruct": { speed: 3, quality: 2, cost: 3 },
-                "meta-llama/Llama-3.1-70B-Instruct": { speed: 2, quality: 3, cost: 2 },
-                "nvidia/Llama-3.1-Nemotron-Ultra-253B-v1": {
-                    speed: 2,
-                    quality: 3,
-                    cost: 1,
-                },
-                "NousResearch/Hermes-3-Llama-3.2-3B": { speed: 3, quality: 2, cost: 3 },
-                "codellama/CodeLlama-34b-Instruct-hf": {
-                    speed: 2,
-                    quality: 3,
-                    cost: 2,
-                },
-            },
-            notes: {
-                "meta-llama/Llama-3.1-8B-Instruct": "Best balance of speed and tool calling capability",
-                "meta-llama/Llama-3.1-70B-Instruct": "High-quality tool calling, slower inference",
-                "nvidia/Llama-3.1-Nemotron-Ultra-253B-v1": "Optimized for tool calling, requires more resources",
-                "NousResearch/Hermes-3-Llama-3.2-3B": "Lightweight with good tool calling support",
-                "codellama/CodeLlama-34b-Instruct-hf": "Excellent for code-related tool calling",
-            },
-        };
-    }
-    /**
-     * Enhanced error handling with HuggingFace-specific guidance
-     */
     formatProviderError(error) {
         if (error instanceof TimeoutError) {
             return new NetworkError(`Request timed out: ${error.message}`, "huggingface");
@@ -330,19 +96,5 @@ Available tools will be provided in the function calling format. Use them when t
         }
         return new ProviderError(`HuggingFace Provider Error: ${message}`, "huggingface");
     }
-    getProviderName() {
-        return "huggingface";
-    }
-    getDefaultModel() {
-        return getDefaultHuggingFaceModel();
-    }
-    /**
-     * Returns the Vercel AI SDK model instance for HuggingFace
-     */
-    getAISDKModel() {
-        return this.model;
-    }
 }
-// Export for factory registration
-export default HuggingFaceProvider;
 //# sourceMappingURL=huggingFace.js.map

package/dist/lib/providers/togetherAi.d.ts

@@ -1,33 +1,23 @@
 import type { AIProviderName } from "../constants/enums.js";
-import { BaseProvider } from "../core/baseProvider.js";
-import type { NeurolinkCredentials, StreamOptions, StreamResult, ValidationSchema } from "../types/index.js";
-import type { LanguageModel } from "../types/index.js";
+import type { NeurolinkCredentials } from "../types/index.js";
+import { OpenAIChatCompletionsProvider } from "./openaiChatCompletionsBase.js";
 /**
- * Together AI Provider
+ * Together AI Provider — direct HTTP, no AI SDK.
  *
  * Hosted open-model gateway at api.together.xyz/v1 (OpenAI-compatible).
  * Llama / Mistral / Qwen / DeepSeek / Gemma / WizardLM available
  * server-less; pass any catalog id via `--model`.
+ * All request/stream/tool-loop orchestration lives in
+ * `OpenAIChatCompletionsProvider`; this class only declares configuration
+ * and provider-specific error mapping.
  *
  * @see https://docs.together.ai/docs/openai-api-compatibility
  */
-export declare class TogetherAIProvider extends BaseProvider {
-    private model;
-    private apiKey;
-    private baseURL;
+export declare class TogetherAIProvider extends OpenAIChatCompletionsProvider {
     constructor(modelName?: string, sdk?: unknown, _region?: string, credentials?: NeurolinkCredentials["together"]);
-    protected executeStream(options: StreamOptions, _analysisSchema?: ValidationSchema): Promise<StreamResult>;
-    private executeStreamInner;
     protected getProviderName(): AIProviderName;
     protected getDefaultModel(): string;
-    protected getAISDKModel(): LanguageModel;
+    protected getFallbackModelName(): string;
+    protected getFallbackModels(): string[];
     protected formatProviderError(error: unknown): Error;
-    validateConfiguration(): Promise<boolean>;
-    getConfiguration(): {
-        provider: AIProviderName;
-        model: string;
-        defaultModel: string;
-        baseURL: string;
-    };
 }
-export default TogetherAIProvider;