@juspay/neurolink 9.68.9 → 9.68.11

package/dist/lib/providers/togetherAi.js

@@ -1,144 +1,60 @@
-import { createOpenAI } from "@ai-sdk/openai";
 import { TogetherAIModels } from "../constants/enums.js";
-import { BaseProvider } from "../core/baseProvider.js";
-import { DEFAULT_MAX_STEPS } from "../core/constants.js";
-import { streamAnalyticsCollector } from "../core/streamAnalytics.js";
-import { isNeuroLink } from "../neurolink.js";
-import { createLoggingFetch } from "../utils/loggingFetch.js";
-import { tracers, ATTR, withClientStreamSpan } from "../telemetry/index.js";
 import { AuthenticationError, InvalidModelError, NetworkError, ProviderError, RateLimitError, } from "../types/index.js";
 import { logger } from "../utils/logger.js";
 import { createTogetherAIConfig, getProviderModel, validateApiKey, } from "../utils/providerConfig.js";
-import { composeAbortSignals, createTimeoutController, TimeoutError, } from "../utils/timeout.js";
-import { emitToolEndFromStepFinish } from "../utils/toolEndEmitter.js";
-import { resolveToolChoice } from "../utils/toolChoice.js";
-import { toAnalyticsStreamResult } from "./providerTypeUtils.js";
-import { stepCountIs } from "../utils/tool.js";
-import { streamText } from "../utils/generation.js";
+import { TimeoutError } from "../utils/timeout.js";
+import { OpenAIChatCompletionsProvider } from "./openaiChatCompletionsBase.js";
 const TOGETHER_DEFAULT_BASE_URL = "https://api.together.xyz/v1";
 const getTogetherApiKey = () => validateApiKey(createTogetherAIConfig());
 const getDefaultTogetherModel = () => getProviderModel("TOGETHER_MODEL", TogetherAIModels.LLAMA_3_3_70B_INSTRUCT_TURBO);
 /**
- * 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 class TogetherAIProvider extends BaseProvider {
-    model;
-    apiKey;
-    baseURL;
+export class TogetherAIProvider extends OpenAIChatCompletionsProvider {
     constructor(modelName, sdk, _region, credentials) {
-        const validatedNeurolink = isNeuroLink(sdk) ? sdk : undefined;
-        super(modelName, "together-ai", validatedNeurolink);
         const overrideApiKey = credentials?.apiKey?.trim();
-        this.apiKey =
-            overrideApiKey && overrideApiKey.length > 0
-                ? overrideApiKey
-                : getTogetherApiKey();
-        this.baseURL =
-            credentials?.baseURL ??
-                process.env.TOGETHER_BASE_URL ??
-                TOGETHER_DEFAULT_BASE_URL;
-        const together = createOpenAI({
-            apiKey: this.apiKey,
-            baseURL: this.baseURL,
-            fetch: createLoggingFetch("together-ai"),
-        });
-        this.model = together.chat(this.modelName);
+        const apiKey = overrideApiKey && overrideApiKey.length > 0
+            ? overrideApiKey
+            : getTogetherApiKey();
+        const baseURL = credentials?.baseURL?.trim() ||
+            process.env.TOGETHER_BASE_URL?.trim() ||
+            TOGETHER_DEFAULT_BASE_URL;
+        super("together-ai", modelName, sdk, { baseURL, apiKey });
         logger.debug("Together AI Provider initialized", {
             modelName: this.modelName,
             providerName: this.providerName,
-            baseURL: this.baseURL,
+            baseURL: this.config.baseURL,
         });
     }
-    async executeStream(options, _analysisSchema) {
-        return withClientStreamSpan({
-            name: "neurolink.provider.stream",
-            tracer: tracers.provider,
-            attributes: {
-                [ATTR.GEN_AI_SYSTEM]: "together-ai",
-                [ATTR.GEN_AI_MODEL]: this.modelName,
-                [ATTR.GEN_AI_OPERATION]: "stream",
-                [ATTR.NL_STREAM_MODE]: true,
-            },
-        }, async () => this.executeStreamInner(options), (r) => r.stream, (r, wrapped) => ({ ...r, stream: wrapped }));
-    }
-    async executeStreamInner(options) {
-        this.validateStreamOptions(options);
-        // Resolve per-call credentials first, then fall back to instance-level.
-        const perCallCreds = options.credentials?.together;
-        const effectiveApiKey = perCallCreds?.apiKey?.trim() || this.apiKey;
-        const effectiveBaseURL = perCallCreds?.baseURL || this.baseURL;
-        const startTime = Date.now();
-        const timeout = this.getTimeout(options);
-        const timeoutController = createTimeoutController(timeout, this.providerName, "stream");
-        try {
-            const shouldUseTools = !options.disableTools && this.supportsTools();
-            const tools = shouldUseTools
-                ? options.tools || (await this.getAllTools())
-                : {};
-            const messages = await this.buildMessagesForStream(options);
-            // When per-call credentials differ from instance, build a fresh client.
-            const hasDifferentCreds = effectiveApiKey !== this.apiKey || effectiveBaseURL !== this.baseURL;
-            const model = hasDifferentCreds
-                ? createOpenAI({
-                    apiKey: effectiveApiKey,
-                    baseURL: effectiveBaseURL,
-                    fetch: createLoggingFetch("together-ai"),
-                }).chat(this.modelName)
-                : await this.getAISDKModelWithMiddleware(options);
-            const result = await streamText({
-                model,
-                messages,
-                temperature: options.temperature,
-                maxOutputTokens: options.maxTokens,
-                tools,
-                stopWhen: stepCountIs(options.maxSteps || DEFAULT_MAX_STEPS),
-                toolChoice: resolveToolChoice(options, tools, shouldUseTools),
-                abortSignal: composeAbortSignals(options.abortSignal, timeoutController?.controller.signal),
-                experimental_telemetry: this.telemetryHandler.getTelemetryConfig(options),
-                experimental_repairToolCall: this.getToolCallRepairFn(options),
-                onStepFinish: ({ toolCalls, toolResults }) => {
-                    emitToolEndFromStepFinish(this.neurolink?.getEventEmitter(), toolResults);
-                    this.handleToolExecutionStorage(toolCalls, toolResults, options, new Date()).catch((error) => {
-                        logger.warn("[TogetherAIProvider] Failed to store tool executions", {
-                            provider: this.providerName,
-                            error: error instanceof Error ? error.message : String(error),
-                        });
-                    });
-                },
-            });
-            timeoutController?.cleanup();
-            const transformedStream = this.createTextStream(result);
-            const analyticsPromise = streamAnalyticsCollector.createAnalytics(this.providerName, this.modelName, toAnalyticsStreamResult(result), Date.now() - startTime, {
-                requestId: `together-stream-${Date.now()}`,
-                streamingMode: true,
-            });
-            return {
-                stream: transformedStream,
-                provider: this.providerName,
-                model: this.modelName,
-                analytics: analyticsPromise,
-                metadata: { startTime, streamId: `together-${Date.now()}` },
-            };
-        }
-        catch (error) {
-            timeoutController?.cleanup();
-            throw this.handleProviderError(error);
-        }
-    }
     getProviderName() {
-        return this.providerName;
+        return "together-ai";
     }
     getDefaultModel() {
         return getDefaultTogetherModel();
     }
-    getAISDKModel() {
-        return this.model;
+    getFallbackModelName() {
+        return TogetherAIModels.LLAMA_3_1_8B_INSTRUCT_TURBO;
+    }
+    getFallbackModels() {
+        return [
+            TogetherAIModels.LLAMA_3_3_70B_INSTRUCT_TURBO,
+            TogetherAIModels.LLAMA_3_1_405B_INSTRUCT_TURBO,
+            TogetherAIModels.LLAMA_3_1_70B_INSTRUCT_TURBO,
+            TogetherAIModels.LLAMA_3_1_8B_INSTRUCT_TURBO,
+            TogetherAIModels.MIXTRAL_8X22B_INSTRUCT,
+            TogetherAIModels.QWEN_2_5_72B_INSTRUCT_TURBO,
+            TogetherAIModels.DEEPSEEK_R1,
+            TogetherAIModels.DEEPSEEK_V3,
+        ];
     }
     formatProviderError(error) {
         if (error instanceof TimeoutError) {
@@ -161,17 +77,5 @@ export class TogetherAIProvider extends BaseProvider {
         }
         return new ProviderError(`Together AI error: ${message}`, "together-ai");
     }
-    async validateConfiguration() {
-        return typeof this.apiKey === "string" && this.apiKey.trim().length > 0;
-    }
-    getConfiguration() {
-        return {
-            provider: this.providerName,
-            model: this.modelName,
-            defaultModel: getDefaultTogetherModel(),
-            baseURL: this.baseURL,
-        };
-    }
 }
-export default TogetherAIProvider;
 //# sourceMappingURL=togetherAi.js.map

package/dist/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;