@juspay/neurolink 12.14.9 → 12.14.11

package/dist/cli/commands/config.js

@@ -87,7 +87,7 @@ const ConfigSchema = z.object({
         huggingface: z
             .object({
             apiKey: z.string().optional(),
-            model: z.string().default("microsoft/DialoGPT-large"),
+            model: z.string().default("Qwen/Qwen2.5-72B-Instruct"),
         })
             .optional(),
         ollama: z
@@ -783,7 +783,7 @@ export class ConfigManager {
                 type: "input",
                 name: "model",
                 message: "Model name:",
-                default: "microsoft/DialoGPT-large",
+                default: "Qwen/Qwen2.5-72B-Instruct",
             },
         ]);
         this.config.providers.huggingface = answers;

package/dist/cli/commands/setup-huggingface.js

@@ -92,26 +92,34 @@ export const handleHuggingFaceSetup = async (argv) => {
                 type: "select",
                 name: "modelChoice",
                 message: "Select a Hugging Face model:",
+                // Router-served ids only. The previous five options
+                // (DialoGPT-large/medium/small, blenderbot-400M-distill) are all
+                // absent from the router catalogue and answer 400 on every request,
+                // so this wizard wrote a dead model into the user's .env.
                 choices: [
                     {
-                        name: "Skip - use default (microsoft/DialoGPT-large)",
+                        name: "Skip - use default (Qwen/Qwen2.5-72B-Instruct)",
                         value: "default",
                     },
                     {
-                        name: "microsoft/DialoGPT-large (Conversational AI - Recommended)",
-                        value: "microsoft/DialoGPT-large",
+                        name: "Qwen/Qwen2.5-72B-Instruct (Recommended - tool-capable)",
+                        value: "Qwen/Qwen2.5-72B-Instruct",
                     },
                     {
-                        name: "microsoft/DialoGPT-medium (Faster, smaller)",
-                        value: "microsoft/DialoGPT-medium",
+                        name: "meta-llama/Llama-3.1-8B-Instruct (Fastest)",
+                        value: "meta-llama/Llama-3.1-8B-Instruct",
                     },
                     {
-                        name: "facebook/blenderbot-400M-distill (General purpose)",
-                        value: "facebook/blenderbot-400M-distill",
+                        name: "meta-llama/Llama-3.3-70B-Instruct (Stronger reasoning)",
+                        value: "meta-llama/Llama-3.3-70B-Instruct",
                     },
                     {
-                        name: "microsoft/DialoGPT-small (Fastest)",
-                        value: "microsoft/DialoGPT-small",
+                        name: "deepseek-ai/DeepSeek-V3 (Highest quality)",
+                        value: "deepseek-ai/DeepSeek-V3",
+                    },
+                    {
+                        name: "Qwen/Qwen2.5-Coder-32B-Instruct (Code-focused)",
+                        value: "Qwen/Qwen2.5-Coder-32B-Instruct",
                     },
                     {
                         name: "Custom model name",
@@ -134,7 +142,7 @@ export const handleHuggingFaceSetup = async (argv) => {
                             return "Model name is required";
                         }
                         if (!trimmed.includes("/")) {
-                            return "Model name should include organization (e.g., 'microsoft/DialoGPT-large')";
+                            return "Model name should include organization (e.g., 'Qwen/Qwen2.5-72B-Instruct')";
                         }
                         return true;
                     },
@@ -173,7 +181,7 @@ export const handleHuggingFaceSetup = async (argv) => {
         }
         else {
             logger.always(chalk.gray('   neurolink generate "Tell me a story" --provider huggingface'));
-            logger.always(chalk.gray('   neurolink generate "Explain AI" --provider huggingface --model microsoft/DialoGPT-medium'));
+            logger.always(chalk.gray('   neurolink generate "Explain AI" --provider huggingface --model Qwen/Qwen2.5-72B-Instruct'));
         }
         logger.always(chalk.blue("\n🔗 Resources:"));
         logger.always(chalk.gray("   • Hugging Face Models: https://huggingface.co/models"));

package/dist/core/modelConfiguration.d.ts

@@ -51,9 +51,9 @@ export declare const MODEL_NAMES: {
         readonly QUALITY: "llama3.1:70b";
     };
     readonly HUGGINGFACE: {
-        readonly FAST: "microsoft/DialoGPT-medium";
-        readonly BALANCED: "microsoft/DialoGPT-large";
-        readonly QUALITY: "meta-llama/Llama-2-7b-chat-hf";
+        readonly FAST: "meta-llama/Llama-3.1-8B-Instruct";
+        readonly BALANCED: "Qwen/Qwen2.5-72B-Instruct";
+        readonly QUALITY: "deepseek-ai/DeepSeek-V3";
     };
     readonly MISTRAL: {
         readonly FAST: "mistral-small-latest";

package/dist/core/modelConfiguration.js

@@ -58,11 +58,13 @@ export const MODEL_NAMES = {
         BALANCED: "llama3.1:8b",
         QUALITY: "llama3.1:70b",
     },
-    // HuggingFace Models
+    // HuggingFace Models — router-served ids. The previous three
+    // (DialoGPT-medium / -large, Llama-2-7b-chat-hf) are absent from the
+    // router catalogue and answer 400 on every request.
     HUGGINGFACE: {
-        FAST: "microsoft/DialoGPT-medium",
-        BALANCED: "microsoft/DialoGPT-large",
-        QUALITY: "meta-llama/Llama-2-7b-chat-hf",
+        FAST: "meta-llama/Llama-3.1-8B-Instruct",
+        BALANCED: "Qwen/Qwen2.5-72B-Instruct",
+        QUALITY: "deepseek-ai/DeepSeek-V3",
     },
     // Mistral Models
     MISTRAL: {

package/dist/models/manifests/huggingface.d.ts

@@ -1,8 +1,11 @@
 import type { ProviderModelManifest } from "../../types/index.js";
 /**
  * Minimal manifest: huggingface has no MODEL_REGISTRY entries today, so
- * only the provider-wide fallback is known. Named models can be added here
- * incrementally without touching any consumer — see Task 5 of the model
- * metadata consolidation plan.
+ * only the provider-wide fallback is known. `functionCalling` is true
+ * because the router serves tool-calling models and returns `tool_calls` for
+ * them; it previously claimed false, contradicting the wire.
+ *
+ * Named models can be added here incrementally without touching any
+ * consumer — see Task 5 of the model metadata consolidation plan.
  */
 export declare const huggingfaceManifest: ProviderModelManifest;

package/dist/models/manifests/huggingface.js

@@ -1,8 +1,11 @@
 /**
  * Minimal manifest: huggingface has no MODEL_REGISTRY entries today, so
- * only the provider-wide fallback is known. Named models can be added here
- * incrementally without touching any consumer — see Task 5 of the model
- * metadata consolidation plan.
+ * only the provider-wide fallback is known. `functionCalling` is true
+ * because the router serves tool-calling models and returns `tool_calls` for
+ * them; it previously claimed false, contradicting the wire.
+ *
+ * Named models can be added here incrementally without touching any
+ * consumer — see Task 5 of the model metadata consolidation plan.
  */
 export const huggingfaceManifest = {
     defaultContextWindow: 32000,
@@ -12,7 +15,7 @@ export const huggingfaceManifest = {
             contextWindow: 32000,
             maxOutputTokens: 32000,
             vision: false,
-            functionCalling: false,
+            functionCalling: true,
         },
     },
 };

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

@@ -19,13 +19,36 @@ export declare class HuggingFaceProvider extends OpenAIChatCompletionsProvider {
     protected getDefaultModel(): string;
     protected getFallbackModelName(): string;
     /**
-     * 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.
+     * Tool support is resolved by the base through MODEL_REGISTRY's
+     * `modelSupports()` facade, which defaults unknown models to supported.
+     *
+     * This class used to override that with a 13-entry allowlist of model-name
+     * substrings, written for the old per-model Inference API where many
+     * endpoints rejected the OpenAI `tools` field. The router replaced that, and
+     * the list did not keep up. Measured against the 142 models
+     * `GET router.huggingface.co/v1/models` served on 2026-09-13 (the
+     * catalogue drifts, so treat the totals as a snapshot; the ratio is the
+     * point):
+     *
+     *   allowlist entries matching ANY served model   1 of 13
+     *   served models the allowlist ADMITS            1 of 142
+     *   served models the allowlist BLOCKS          141
+     *
+     * So it suppressed tool calling for 99% of the catalogue, including
+     * Qwen/Qwen2.5-72B-Instruct, which the router demonstrably tool-calls. A
+     * spread of the blocked models — GLM-5.3-Flash, gemma-4-31B-it,
+     * DeepSeek-V4-Flash — all accept `tools` with HTTP 200 and return
+     * `tool_calls`; none rejected the field.
+     *
+     * The allowlist's own comment named `microsoft/DialoGPT-medium` — which
+     * `getDefaultModel()`, not `getFallbackModelName()`, used to return — as a
+     * model that rejects `tools`. The router does not serve it under any
+     * request shape, so that path was never reachable; the default now points
+     * at a served model (see `getDefaultHuggingFaceModel`).
+     *
+     * A model that genuinely cannot use tools simply does not emit `tool_calls`,
+     * which the loop already handles, so the optimistic default costs nothing
+     * that the allowlist was protecting.
      */
-    supportsTools(): boolean;
     protected formatProviderError(error: unknown): Error;
 }

package/dist/providers/huggingFace/client.js

@@ -1,3 +1,4 @@
+import { HuggingFaceModels } from "../../constants/enums.js";
 import { AuthenticationError, InvalidModelError, ProviderError, RateLimitError, } from "../../types/index.js";
 import { classifyProviderError, DEFAULT_ERROR_RULES, } from "../../utils/errorClassifier.js";
 import { logger } from "../../utils/logger.js";
@@ -6,7 +7,19 @@ import { createHuggingFaceConfig, getProviderModel, validateApiKey, } from "../.
 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");
+/**
+ * Must stay the same id `providerRegistry` passes as the registration default
+ * (`process.env.HUGGINGFACE_MODEL || HuggingFaceModels.QWEN_2_5_72B_INSTRUCT`),
+ * hence the shared constant rather than a second literal.
+ *
+ * The registry supplies `modelName` on every factory-created provider, so this
+ * only decides a directly-constructed `new HuggingFaceProvider()`. It still
+ * matters: it previously read `microsoft/DialoGPT-medium`, which the router
+ * does not serve under any request shape — it answers 400 "not supported by
+ * any provider you have enabled" with or without `tools` — and that stale id
+ * was what the docs and setup guidance advertised as the default.
+ */
+const getDefaultHuggingFaceModel = () => getProviderModel("HUGGINGFACE_MODEL", HuggingFaceModels.QWEN_2_5_72B_INSTRUCT);
 /**
  * HuggingFace Provider — direct HTTP, no AI SDK.
  *
@@ -47,32 +60,37 @@ export class HuggingFaceProvider extends OpenAIChatCompletionsProvider {
         return "meta-llama/Llama-3.1-8B-Instruct";
     }
     /**
-     * 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.
+     * Tool support is resolved by the base through MODEL_REGISTRY's
+     * `modelSupports()` facade, which defaults unknown models to supported.
+     *
+     * This class used to override that with a 13-entry allowlist of model-name
+     * substrings, written for the old per-model Inference API where many
+     * endpoints rejected the OpenAI `tools` field. The router replaced that, and
+     * the list did not keep up. Measured against the 142 models
+     * `GET router.huggingface.co/v1/models` served on 2026-09-13 (the
+     * catalogue drifts, so treat the totals as a snapshot; the ratio is the
+     * point):
+     *
+     *   allowlist entries matching ANY served model   1 of 13
+     *   served models the allowlist ADMITS            1 of 142
+     *   served models the allowlist BLOCKS          141
+     *
+     * So it suppressed tool calling for 99% of the catalogue, including
+     * Qwen/Qwen2.5-72B-Instruct, which the router demonstrably tool-calls. A
+     * spread of the blocked models — GLM-5.3-Flash, gemma-4-31B-it,
+     * DeepSeek-V4-Flash — all accept `tools` with HTTP 200 and return
+     * `tool_calls`; none rejected the field.
+     *
+     * The allowlist's own comment named `microsoft/DialoGPT-medium` — which
+     * `getDefaultModel()`, not `getFallbackModelName()`, used to return — as a
+     * model that rejects `tools`. The router does not serve it under any
+     * request shape, so that path was never reachable; the default now points
+     * at a served model (see `getDefaultHuggingFaceModel`).
+     *
+     * A model that genuinely cannot use tools simply does not emit `tool_calls`,
+     * which the loop already handles, so the optimistic default costs nothing
+     * that the allowlist was protecting.
      */
-    supportsTools() {
-        const modelName = this.modelName.toLowerCase();
-        const toolCapableModels = [
-            "llama-3.1-8b-instruct",
-            "llama-3.1-70b-instruct",
-            "llama-3.1-405b-instruct",
-            "llama-3.1-nemotron-ultra",
-            "hermes-3-llama-3.2",
-            "hermes-2-pro",
-            "codellama-34b-instruct",
-            "codellama-13b-instruct",
-            "mistral-7b-instruct-v0.3",
-            "mistral-8x7b-instruct",
-            "nous-hermes",
-            "openchat",
-            "wizardcoder",
-        ];
-        return toolCapableModels.some((capable) => modelName.includes(capable));
-    }
     formatProviderError(error) {
         const rules = [
             {

package/dist/providers/openaiChatCompletionsBase.d.ts

@@ -244,6 +244,27 @@ export declare abstract class OpenAIChatCompletionsProvider extends BaseProvider
      */
     generate(optionsOrPrompt: TextGenerationOptions | string, analysisSchema?: ValidationSchema): Promise<EnhancedGenerateResult | null>;
     private executeNativeGenerate;
+    /**
+     * Recover structured output on a turn where tools forced `response_format`
+     * off the wire.
+     *
+     * Runs only after the tool turn is complete, and only when its answer does
+     * not already satisfy the schema. The re-ask drops the tools, which is what
+     * makes native `response_format` legal again on a vendor that refuses the two
+     * together, so this is a formatting pass over an answer the model has already
+     * produced rather than a second attempt at the question.
+     *
+     * The schema is deliberately NOT spelled into the tool turn's system prompt
+     * instead. That was tried and reverted: a model handed a JSON Schema next to a
+     * tool list can read it as another tool — Groq's llama-3.3 answers by calling
+     * one named "json", which the server rejects — turning a wrong answer into a
+     * hard error.
+     *
+     * Returns the merged result, or the untouched `toolPhase` whenever the re-ask
+     * cannot be run or cannot be trusted, so it can improve an outcome but never
+     * degrade one.
+     */
+    private reformatWithoutTools;
     protected executeStream(options: StreamOptions, _analysisSchema?: ZodUnknownSchema | Schema<unknown>): Promise<StreamResult>;
     private runStreamLoop;
     private streamOneStep;