@juspay/neurolink 12.23.3 → 12.24.0

package/dist/providers/openaiCompatCatalog.js

@@ -1,24 +1,28 @@
 import { buildCatalogEntries } from "./catalog/loader.js";
 /**
- * Config-driven catalog of the 9 zero-quirk OpenAI-compatible providers.
- * Each entry fully replaces what used to be a hand-written
- * OpenAIChatCompletionsProvider subclass — see ConfiguredOpenAICompatProvider
- * for the class that reads these entries, and providerRegistry.ts for the
- * registration loop that consumes this array.
+ * Config-driven catalog of the 10 OpenAI-compatible providers whose only
+ * differences from each other are expressible as named, closed catalog
+ * quirks (messageContentFormat, timeoutErrorClass, responseFormatDowngrade)
+ * rather than real code. Each entry fully replaces what used to be a
+ * hand-written OpenAIChatCompletionsProvider subclass — see
+ * ConfiguredOpenAICompatProvider for the class that reads these entries,
+ * and providerRegistry.ts for the registration loop that consumes this
+ * array.
  *
  * The entries themselves are derived from the JSON catalog
  * (src/lib/providers/catalog/<id>.json) via buildCatalogEntries() — that
- * JSON is now the single source of truth for these 9 providers' identity,
+ * JSON is now the single source of truth for these 10 providers' identity,
  * models, error rules and wire config. See
  * src/lib/providers/catalog/loader.ts for the derivation logic and
  * docs/superpowers/plans/2026-08-28-provider-json-catalog-spec.md for the
  * authoring format.
  *
- * To add a new zero-quirk OpenAI-compatible provider: add one JSON file
- * under src/lib/providers/catalog/ (see provider-catalog.schema.json) and
- * run `pnpm run codegen:catalog`. Do NOT add a provider here if it needs
- * any hook override beyond the 3 mandatory ones
- * (getProviderName/getDefaultModel/formatProviderError) — write a
- * dedicated subclass instead (see deepseek.ts, azureOpenai.ts).
+ * To add a new OpenAI-compatible provider: add one JSON file under
+ * src/lib/providers/catalog/ (see provider-catalog.schema.json) and run
+ * `pnpm run codegen:catalog`. Do NOT add a provider here if it needs a real
+ * hook override beyond the 3 mandatory ones
+ * (getProviderName/getDefaultModel/formatProviderError) and the named
+ * catalog quirks above — write a dedicated subclass instead (see
+ * azureOpenai.ts).
  */
 export const OPENAI_COMPAT_CATALOG = buildCatalogEntries();

package/dist/providers/voyage.js

@@ -1,6 +1,7 @@
 import { VoyageModels } from "../constants/enums.js";
 import { BaseProvider } from "../core/baseProvider.js";
 import { isNeuroLink } from "../neurolink.js";
+import { parseIndexedEmbeddingsResponse } from "./embeddingResponseParsing.js";
 import { createProxyFetch } from "../proxy/proxyFetch.js";
 import { AuthenticationError, InvalidModelError, ProviderError, RateLimitError, } from "../types/index.js";
 import { withTimeout } from "../utils/errorHandling.js";
@@ -146,24 +147,10 @@ export class VoyageProvider extends BaseProvider {
             throw this.formatProviderError(new Error(`Voyage embeddings failed: ${response.status} — ${text}`));
         }
         const data = (await response.json());
-        if (!data.data || data.data.length === 0) {
-            throw new ProviderError("Voyage embeddings response missing data", "voyage");
-        }
-        // Validate that the response covers all requested inputs.
-        // Voyage may return partial results in edge-case error scenarios.
-        if (data.data.length !== inputs.length) {
-            throw new ProviderError(`Voyage embeddings response count mismatch: expected ${inputs.length}, got ${data.data.length}`, "voyage");
-        }
-        // Sort by index (Voyage returns out-of-order under some conditions)
-        // and drop to a flat number[][] result.
-        const sorted = data.data.slice().sort((a, b) => a.index - b.index);
-        // Verify that index coverage is complete (0 … n-1).
-        for (let i = 0; i < sorted.length; i++) {
-            if (sorted[i].index !== i) {
-                throw new ProviderError(`Voyage embeddings response has unexpected index ordering: position ${i} has index ${sorted[i].index}`, "voyage");
-            }
-        }
-        return sorted.map((d) => d.embedding);
+        // Sorts by index (Voyage returns out-of-order under some conditions) and
+        // verifies complete 0..n-1 coverage (Voyage may return partial results
+        // in edge-case error scenarios).
+        return parseIndexedEmbeddingsResponse(data, inputs.length, "Voyage", (message) => new ProviderError(message, "voyage"));
     }
     async validateConfiguration() {
         return typeof this.apiKey === "string" && this.apiKey.trim().length > 0;

package/dist/types/providerCatalog.d.ts

@@ -10,6 +10,18 @@ export type CatalogPricingPerMTok = {
     cachedInput?: number;
 };
 export type CatalogModelStatus = "production" | "preview" | "retired";
+/**
+ * Descriptor-only per-provider turn-budget override — mirrors
+ * ProviderDescriptor.timeouts field-for-field (src/lib/types/providers.ts).
+ * Consumed only by buildCatalogDescriptor() (providerDescriptors.ts), never
+ * by the OpenAI-compat runtime path (loader.ts's buildCatalogEntries()).
+ * Optional and absent for every catalog provider except mistral, whose
+ * value preserves its pre-migration hand-typed descriptor exactly.
+ */
+export type CatalogTimeouts = {
+    generateMs?: number;
+    streamMs?: number;
+};
 export type CatalogModelSpec = {
     contextWindow?: number;
     maxOutputTokens?: number;
@@ -34,6 +46,11 @@ export type CatalogWire = {
         baseURL?: string;
         model?: string;
     };
+    /** Additional env vars validateApiKey() tries, in order, when the primary
+     *  apiKeyEnvVar is unset (e.g. HuggingFace's HF_TOKEN alongside
+     *  HUGGINGFACE_API_KEY). Consumed by buildCatalogConfigOptions() via
+     *  ProviderConfigOptions.fallbackEnvVars. */
+    apiKeyFallbackEnvVars?: string[];
 };
 export type CatalogErrorRuleClass = "authentication" | "rate-limit" | "invalid-model" | "network" | "provider";
 export type CatalogErrorRuleJson = {
@@ -51,6 +68,11 @@ export type CatalogQuirks = {
      *  ConfiguredOpenAICompatProvider so tool round-trips work. */
     messageContentFormat?: "string";
     registryDefaultIgnoresModelEnvVar?: boolean;
+    /** Vendor rejects `response_format: { type: "json_schema" }` outright but
+     *  accepts `{ type: "json_object" }`. Normalized by
+     *  ConfiguredOpenAICompatProvider so `generate({ schema })` keeps working
+     *  (mirrors the pre-catalog `supportsStructuredOutputs: false` behavior). */
+    responseFormatDowngrade?: "json-schema-to-json-object";
 };
 export type CatalogBillingPolicy = "free-tier" | "free-with-card" | "no-free-tier";
 export type CatalogSetup = {
@@ -85,7 +107,14 @@ export type CatalogEvidence = {
 export type CatalogCapabilities = {
     text: boolean;
     streaming: boolean;
-    tools: boolean;
+    /** "model-dependent" when tool support varies per served model and the
+     *  vendor doesn't reject `tools` for unsupported ones (the model just
+     *  never emits tool_calls) — e.g. HuggingFace's router. Maps to
+     *  ProviderDescriptor.toolSupport's own "model-dependent" member and
+     *  leaves OpenAICompatCatalogEntry.supportsTools unset so
+     *  ConfiguredOpenAICompatProvider falls through to the model-registry
+     *  default, exactly like an entry that never set supportsTools at all. */
+    tools: boolean | "model-dependent";
     toolsWithStreaming: boolean;
     structuredOutput: boolean;
     structuredOutputWithTools: boolean;
@@ -150,6 +179,22 @@ export type ProviderCatalogJson = {
     capabilities: CatalogCapabilities;
     errorRules: CatalogErrorRuleJson[];
     quirks?: CatalogQuirks;
+    /**
+     * Descriptor-only generate/stream turn-budget override — see
+     * CatalogTimeouts. Absent = buildCatalogDescriptor() omits
+     * ProviderDescriptor.timeouts entirely, matching every catalog provider's
+     * behavior before this field existed.
+     */
+    timeouts?: CatalogTimeouts;
+    /**
+     * Descriptor-only ascending auto-select priority override — mirrors
+     * ProviderDescriptor.autoSelectPriority field-for-field (lower = tried
+     * first in getBestProvider()'s fallback chain). Absent = not part of the
+     * auto-select chain, matching every catalog provider's behavior before
+     * this field existed (only mistral sets it today, preserving its
+     * pre-migration hand-typed value).
+     */
+    autoSelectPriority?: number;
     setup: CatalogSetup;
     evidence: CatalogEvidence;
 };

package/dist/types/providerCatalog.generated.d.ts

@@ -1,2 +1,2 @@
-export type CatalogProviderName = "api-route" | "baseten" | "cerebras" | "cloudflare" | "fireworks" | "friendli" | "gmicloud" | "groq" | "inception-labs" | "io-intelligence" | "mancer" | "mistral" | "perplexity" | "sambanova" | "together-ai" | "upstage" | "xai";
-export type CatalogCredentialKey = "apiRoute" | "baseten" | "cerebras" | "cloudflare" | "fireworks" | "friendli" | "gmicloud" | "groq" | "inceptionLabs" | "ioIntelligence" | "mancer" | "mistral" | "perplexity" | "sambanova" | "together" | "upstage" | "xai";
+export type CatalogProviderName = "api-route" | "baseten" | "cerebras" | "cloudflare" | "deepseek" | "fireworks" | "friendli" | "gmicloud" | "groq" | "huggingface" | "inception-labs" | "io-intelligence" | "mancer" | "mistral" | "perplexity" | "sambanova" | "together-ai" | "upstage" | "xai";
+export type CatalogCredentialKey = "apiRoute" | "baseten" | "cerebras" | "cloudflare" | "deepseek" | "fireworks" | "friendli" | "gmicloud" | "groq" | "huggingFace" | "inceptionLabs" | "ioIntelligence" | "mancer" | "mistral" | "perplexity" | "sambanova" | "together" | "upstage" | "xai";

package/dist/types/providers.d.ts

@@ -143,10 +143,6 @@ export type NeurolinkCredentials = {
         apiVersion?: string;
         useMaxCompletionTokens?: boolean;
     };
-    huggingFace?: {
-        apiKey?: string;
-        baseURL?: string;
-    };
     openrouter?: {
         apiKey?: string;
         baseURL?: string;
@@ -163,10 +159,6 @@ export type NeurolinkCredentials = {
         baseURL?: string;
         apiKey?: string;
     };
-    deepseek?: {
-        apiKey?: string;
-        baseURL?: string;
-    };
     nvidiaNim?: {
         apiKey?: string;
         baseURL?: string;
@@ -196,6 +188,10 @@ export type NeurolinkCredentials = {
         baseURL?: string;
         accountId?: string;
     };
+    deepseek?: {
+        apiKey?: string;
+        baseURL?: string;
+    };
     fireworks?: {
         apiKey?: string;
         baseURL?: string;
@@ -212,6 +208,10 @@ export type NeurolinkCredentials = {
         apiKey?: string;
         baseURL?: string;
     };
+    huggingFace?: {
+        apiKey?: string;
+        baseURL?: string;
+    };
     inceptionLabs?: {
         apiKey?: string;
         baseURL?: string;
@@ -771,6 +771,10 @@ export type OpenAICompatCatalogEntry = {
     /** See CatalogQuirks.messageContentFormat — a vendor that accepts
      *  `messages[].content` only as a plain string. */
     messageContentFormat?: "string";
+    /** See CatalogQuirks.responseFormatDowngrade — a vendor that rejects
+     *  `response_format: { type: "json_schema" }` but accepts
+     *  `{ type: "json_object" }`. */
+    responseFormatDowngrade?: "json-schema-to-json-object";
 };
 /** The subset of OpenAICompatCatalogEntry that resolveOpenAICompatConfig()
  *  needs — lets call sites pass a minimal object without the full catalog

package/dist/utils/modelChoices.d.ts

@@ -7,9 +7,9 @@ import type { CatalogProviderName, ModelChoice } from "../types/index.js";
 /**
  * Default models per provider (first choice/recommended).
  *
- * Covers every provider EXCEPT the 9 JSON-catalog providers — those default
+ * Covers every provider EXCEPT the 11 JSON-catalog providers — those default
  * models come from `models.default` in the catalog JSON, read directly by
- * `getDefaultModel()` below. (Two of the 9 — cerebras, sambanova — never had
+ * `getDefaultModel()` below. (Two of the 11 — cerebras, sambanova — never had
  * a hand entry here at all; `getDefaultModel()` now resolves them too.)
  *
  * AUTO is also excluded — it never had an entry here either (matches

package/dist/utils/modelChoices.js

@@ -2,10 +2,10 @@
  * Centralized model choices for CLI commands
  * Derives choices from model enums to ensure consistency
  */
-import { AIProviderName, OpenAIModels, AnthropicModels, GoogleAIModels, BedrockModels, VertexModels, OllamaModels, AzureOpenAIModels, LiteLLMModels, HuggingFaceModels, SageMakerModels, OpenRouterModels, DeepSeekModels, NvidiaNimModels, CohereModels, VoyageModels, JinaModels, StabilityModels, IdeogramModels, RecraftModels, TypeSafeModels, ReplicateModels, } from "../constants/enums.js";
+import { AIProviderName, OpenAIModels, AnthropicModels, GoogleAIModels, BedrockModels, VertexModels, OllamaModels, AzureOpenAIModels, LiteLLMModels, SageMakerModels, OpenRouterModels, NvidiaNimModels, CohereModels, VoyageModels, JinaModels, StabilityModels, IdeogramModels, RecraftModels, TypeSafeModels, ReplicateModels, } from "../constants/enums.js";
 import { getCatalogJsonEntries } from "../providers/catalog/loader.js";
 /**
- * Looks up the JSON catalog entry for a provider, if it is one of the 9
+ * Looks up the JSON catalog entry for a provider, if it is one of the 11
  * JSON-catalog providers. Used to short-circuit the hand tables below so a
  * catalog provider's models are always sourced from its catalog JSON, never
  * from stale hand-written duplicates.
@@ -36,11 +36,11 @@ function catalogTopModels(entry) {
  * Top models per provider with descriptions for CLI prompts
  * These are curated lists of the most commonly used/recommended models
  *
- * Covers every provider EXCEPT the 9 JSON-catalog providers (cerebras,
- * cloudflare, fireworks, groq, mistral, perplexity, sambanova, together-ai,
- * xai) — their entries are derived from `models.catalog[*].description` by
- * `catalogTopModels()` and merged in at the accessor-function level below,
- * never hand-duplicated here.
+ * Covers every provider EXCEPT the 11 JSON-catalog providers (cerebras,
+ * cloudflare, deepseek, fireworks, groq, huggingface, mistral, perplexity,
+ * sambanova, together-ai, xai) — their entries are derived from
+ * `models.catalog[*].description` by `catalogTopModels()` and merged in at
+ * the accessor-function level below, never hand-duplicated here.
  */
 const TOP_MODELS_CONFIG = {
     [AIProviderName.OPENAI]: [
@@ -194,23 +194,6 @@ const TOP_MODELS_CONFIG = {
             description: "Vertex via LiteLLM",
         },
     ],
-    [AIProviderName.HUGGINGFACE]: [
-        {
-            model: HuggingFaceModels.LLAMA_3_3_70B_INSTRUCT,
-            description: "Recommended - Latest Llama",
-        },
-        {
-            model: HuggingFaceModels.MISTRAL_LARGE_3_675B,
-            description: "Mistral Large",
-        },
-        { model: HuggingFaceModels.DEEPSEEK_R1, description: "Advanced reasoning" },
-        {
-            model: HuggingFaceModels.QWEN_2_5_72B_INSTRUCT,
-            description: "Qwen flagship",
-        },
-        { model: HuggingFaceModels.PHI_4, description: "Microsoft Phi-4" },
-        { model: HuggingFaceModels.GEMMA_3_27B_IT, description: "Google Gemma 3" },
-    ],
     [AIProviderName.SAGEMAKER]: [
         {
             model: SageMakerModels.LLAMA_4_MAVERICK_17B_128E,
@@ -245,13 +228,6 @@ const TOP_MODELS_CONFIG = {
         { model: "gpt-4-turbo", description: "Turbo compatible model" },
         { model: "gpt-3.5-turbo", description: "Legacy compatible model" },
     ],
-    [AIProviderName.DEEPSEEK]: [
-        { model: "deepseek-chat", description: "DeepSeek-V3 general chat" },
-        {
-            model: "deepseek-reasoner",
-            description: "DeepSeek-R1 reasoning (slower, deeper)",
-        },
-    ],
     [AIProviderName.NVIDIA_NIM]: [
         {
             model: "meta/llama-3.3-70b-instruct",
@@ -408,9 +384,9 @@ const TOP_MODELS_CONFIG = {
 /**
  * Default models per provider (first choice/recommended).
  *
- * Covers every provider EXCEPT the 9 JSON-catalog providers — those default
+ * Covers every provider EXCEPT the 11 JSON-catalog providers — those default
  * models come from `models.default` in the catalog JSON, read directly by
- * `getDefaultModel()` below. (Two of the 9 — cerebras, sambanova — never had
+ * `getDefaultModel()` below. (Two of the 11 — cerebras, sambanova — never had
  * a hand entry here at all; `getDefaultModel()` now resolves them too.)
  *
  * AUTO is also excluded — it never had an entry here either (matches
@@ -425,11 +401,9 @@ export const DEFAULT_MODELS = {
     [AIProviderName.AZURE]: AzureOpenAIModels.GPT_4O,
     [AIProviderName.OLLAMA]: OllamaModels.LLAMA4_LATEST,
     [AIProviderName.LITELLM]: LiteLLMModels.OPENAI_GPT_4O,
-    [AIProviderName.HUGGINGFACE]: HuggingFaceModels.LLAMA_3_3_70B_INSTRUCT,
     [AIProviderName.SAGEMAKER]: SageMakerModels.LLAMA_4_MAVERICK_17B_128E,
     [AIProviderName.OPENROUTER]: OpenRouterModels.CLAUDE_SONNET_4_5,
     [AIProviderName.OPENAI_COMPATIBLE]: "gpt-4o",
-    [AIProviderName.DEEPSEEK]: DeepSeekModels.DEEPSEEK_CHAT,
     [AIProviderName.NVIDIA_NIM]: NvidiaNimModels.GPT_OSS_20B,
     // LM Studio + llama.cpp auto-discover their loaded model from /v1/models;
     // an empty default is the documented signal to use that path.
@@ -447,7 +421,7 @@ export const DEFAULT_MODELS = {
 /**
  * Model enum mappings for getAllModels.
  *
- * Covers every provider EXCEPT the 9 JSON-catalog providers — those model
+ * Covers every provider EXCEPT the 11 JSON-catalog providers — those model
  * lists come from `Object.keys(models.catalog)` in the catalog JSON, read
  * directly by `getAllModels()` below.
  */
@@ -460,11 +434,9 @@ const MODEL_ENUMS = {
     [AIProviderName.AZURE]: AzureOpenAIModels,
     [AIProviderName.OLLAMA]: OllamaModels,
     [AIProviderName.LITELLM]: LiteLLMModels,
-    [AIProviderName.HUGGINGFACE]: HuggingFaceModels,
     [AIProviderName.SAGEMAKER]: SageMakerModels,
     [AIProviderName.OPENROUTER]: OpenRouterModels,
     [AIProviderName.OPENAI_COMPATIBLE]: null,
-    [AIProviderName.DEEPSEEK]: DeepSeekModels,
     [AIProviderName.NVIDIA_NIM]: NvidiaNimModels,
     [AIProviderName.LM_STUDIO]: null,
     [AIProviderName.LLAMACPP]: null,

package/dist/utils/sizeGuard.d.ts

@@ -2,9 +2,9 @@
  * Size Guard Utility
  *
  * Provides bounded binary downloads to prevent OOM when fetching generated
- * media from external providers. Applies a Content-Length pre-check and a
- * post-buffer guard so multi-GB responses are rejected before they fully
- * materialise in process memory.
+ * media from external providers. Applies a Content-Length pre-check and
+ * counts bytes while streaming, so multi-GB responses are rejected before
+ * they fully materialise in process memory.
  *
  * @module utils/sizeGuard
  */
@@ -21,8 +21,11 @@ export declare const MAX_IMAGE_BYTES: number;
  * Two checks are performed:
  * 1. If the response includes a `Content-Length` header that exceeds
  *    `maxBytes`, the download is rejected immediately (no data is read).
- * 2. After buffering, the actual buffer size is verified against `maxBytes`.
- *    This catches chunked transfers where no `Content-Length` was provided.
+ * 2. The body is read as a stream and the reader is cancelled as soon as
+ *    more than `maxBytes` arrive. This catches chunked transfers where no
+ *    `Content-Length` was provided (or it understated the body) without
+ *    buffering the excess. A response without a readable stream falls back
+ *    to buffering and checking afterwards.
  *
  * @param response  The fetch {@link Response} to drain.
  * @param maxBytes  Maximum number of bytes allowed.

package/dist/utils/sizeGuard.js

@@ -2,9 +2,9 @@
  * Size Guard Utility
  *
  * Provides bounded binary downloads to prevent OOM when fetching generated
- * media from external providers. Applies a Content-Length pre-check and a
- * post-buffer guard so multi-GB responses are rejected before they fully
- * materialise in process memory.
+ * media from external providers. Applies a Content-Length pre-check and
+ * counts bytes while streaming, so multi-GB responses are rejected before
+ * they fully materialise in process memory.
  *
  * @module utils/sizeGuard
  */
@@ -21,8 +21,11 @@ export const MAX_IMAGE_BYTES = 25 * 1024 * 1024;
  * Two checks are performed:
  * 1. If the response includes a `Content-Length` header that exceeds
  *    `maxBytes`, the download is rejected immediately (no data is read).
- * 2. After buffering, the actual buffer size is verified against `maxBytes`.
- *    This catches chunked transfers where no `Content-Length` was provided.
+ * 2. The body is read as a stream and the reader is cancelled as soon as
+ *    more than `maxBytes` arrive. This catches chunked transfers where no
+ *    `Content-Length` was provided (or it understated the body) without
+ *    buffering the excess. A response without a readable stream falls back
+ *    to buffering and checking afterwards.
  *
  * @param response  The fetch {@link Response} to drain.
  * @param maxBytes  Maximum number of bytes allowed.
@@ -36,9 +39,29 @@ export async function readBoundedBuffer(response, maxBytes, label) {
     if (contentLength > 0 && contentLength > maxBytes) {
         throw new Error(`${label} download too large: ${contentLength} bytes (max ${maxBytes})`);
     }
-    const buffer = Buffer.from(await response.arrayBuffer());
-    if (buffer.length > maxBytes) {
-        throw new Error(`${label} download exceeded size cap after fetch: ${buffer.length} bytes (max ${maxBytes})`);
+    const exceeded = (bytes) => new Error(`${label} download exceeded size cap after fetch: ${bytes} bytes (max ${maxBytes})`);
+    const body = response.body;
+    if (!body || typeof body.getReader !== "function") {
+        const buffer = Buffer.from(await response.arrayBuffer());
+        if (buffer.length > maxBytes) {
+            throw exceeded(buffer.length);
+        }
+        return buffer;
     }
-    return buffer;
+    const reader = body.getReader();
+    const chunks = [];
+    let received = 0;
+    for (;;) {
+        const { done, value } = await reader.read();
+        if (done) {
+            break;
+        }
+        received += value.byteLength;
+        if (received > maxBytes) {
+            await reader.cancel().catch(() => undefined);
+            throw exceeded(received);
+        }
+        chunks.push(value);
+    }
+    return Buffer.concat(chunks, received);
 }