@superlinked/sie-sdk 0.2.0 → 0.3.1

package/dist/index.d.cts

@@ -15,6 +15,19 @@ type DType = "float32" | "float16" | "bfloat16" | "int8" | "uint8" | "binary" |
  * Output type options for encode operation.
  */
 type OutputType = "dense" | "sparse" | "multivector";
+/**
+ * Document input for composite-document extractors (PDF, DOCX, HTML, ...).
+ *
+ * The wire format is the document bytes plus an optional format hint. The
+ * hint is advisory — adapters may sniff the bytes when it is missing or
+ * unrecognized.
+ */
+interface DocumentInput {
+    /** Document bytes (raw file content) */
+    data: Uint8Array;
+    /** Document format hint: "pdf", "docx", "html", etc. */
+    format?: string;
+}
 /**
  * A single item to encode, score, or extract from.
  *
@@ -30,6 +43,9 @@ type OutputType = "dense" | "sparse" | "multivector";
  * // With images for multimodal models (ColPali, CLIP)
  * { text: "Description", images: [imageBytes] }
  *
+ * // With a document for composite-document extractors (Docling, ...)
+ * { document: { data: pdfBytes, format: "pdf" } }
+ *
  * // Pre-encoded multivector (for use with maxsim utility)
  * { multivector: [tokenEmbedding1, tokenEmbedding2, ...] }
  */
@@ -40,6 +56,8 @@ interface Item {
     text?: string;
     /** Images as byte arrays (JPEG/PNG) for multimodal models */
     images?: Uint8Array[];
+    /** Document for composite-document extractors (PDF, DOCX, HTML, ...) */
+    document?: DocumentInput;
     /** Pre-encoded multivector (for use with maxsim utility) */
     multivector?: Float32Array[];
     /** Arbitrary metadata (passed through to results) */
@@ -98,7 +116,7 @@ interface ModelInfo {
     name: string;
     /** Whether the model is currently loaded in memory */
     loaded: boolean;
-    /** Supported input types: ["text"], ["text", "image"], etc. */
+    /** Supported input types: ["text"], ["text", "image"], ["text", "document"], etc. */
     inputs: string[];
     /** Supported output types: ["dense"], ["dense", "sparse"], etc. */
     outputs: string[];
@@ -448,7 +466,7 @@ declare function toFloat32Array(arr: number[]): Float32Array;
  *
  * @example Resource pool usage
  * ```typescript
- * const client = new SIEClient("http://router:8080");
+ * const client = new SIEClient("http://gateway:8080");
  *
  * // Create a dedicated pool
  * await client.createPool("eval-bench", { l4: 2 });
@@ -523,9 +541,9 @@ declare class SIEClient {
      */
     getModel(name: string): Promise<ModelInfo>;
     /**
-     * Stream real-time status updates from a worker or router.
+     * Stream real-time status updates from a worker or gateway.
      *
-     * @param mode - "cluster" uses router /ws/cluster-status, "worker" uses /ws/status.
+     * @param mode - "cluster" uses gateway /ws/cluster-status, "worker" uses /ws/status.
      *               "auto" detects the endpoint via /health.
      */
     watch(mode?: "auto" | "cluster" | "worker"): AsyncGenerator<StatusMessage>;
@@ -576,7 +594,7 @@ declare class SIEClient {
      * Close the client and cleanup resources.
      *
      * Stops pool lease renewal timers. Note that pools are not deleted
-     * automatically - they are garbage collected by the router after inactivity.
+     * automatically - they are garbage collected by the gateway after inactivity.
      * This allows pool reuse if the client reconnects.
      */
     close(): Promise<void>;
@@ -643,7 +661,7 @@ declare class SIEClient {
     /**
      * Get current cluster capacity information.
      *
-     * Queries the router's /health endpoint for cluster state. Useful for
+     * Queries the gateway's /health endpoint for cluster state. Useful for
      * checking if specific GPU types are available before sending requests.
      *
      * @param gpu - Optional filter to check specific GPU type availability
@@ -666,7 +684,7 @@ declare class SIEClient {
     /**
      * Wait for GPU capacity to become available.
      *
-     * Polls the router until workers with the specified GPU type are online.
+     * Polls the gateway until workers with the specified GPU type are online.
      * This is useful for pre-warming the cluster before running benchmarks.
      *
      * @param gpu - GPU type to wait for (e.g., "l4", "a100-80gb")
@@ -689,7 +707,15 @@ declare class SIEClient {
         pollInterval?: number;
     }): Promise<CapacityInfo>;
     /**
-     * Make a msgpack HTTP request with retry logic for 202 and LoRA loading.
+     * Make a msgpack HTTP request with retry logic.
+     *
+     * Retried (only when `waitForCapacity: true`, capped by `provisionTimeout`):
+     *  - 202 Accepted (provisioning)
+     *  - 503 `MODEL_LOADING` / `LORA_LOADING` / no error code (scale-from-zero)
+     *  - `SIEConnectionError` with `kind === "connect"` (issue #95)
+     *
+     * `kind === "timeout"` is NOT retried — would extend the user-visible
+     * timeout from `timeout` to `provisionTimeout`.
      */
     private requestWithRetry;
     /**
@@ -706,7 +732,7 @@ declare class SIEClient {
     private detectEndpointType;
 }
 
-declare const SDK_VERSION = "0.2.0";
+declare const SDK_VERSION = "0.3.1";
 
 /**
  * Helpers for converting SIE encode results to plain JavaScript types.
@@ -806,6 +832,11 @@ declare function multivectorEmbedding(raw: Float32Array[]): number[][];
 declare class SIEError extends Error {
     constructor(message: string);
 }
+/**
+ * `SIEConnectionError` failure category. Only `"connect"` is auto-retried
+ * under `waitForCapacity: true`; `"timeout"` and `"other"` fail fast.
+ */
+type SIEConnectionErrorKind = "connect" | "timeout" | "other";
 /**
  * Error connecting to the SIE server.
  *
@@ -816,7 +847,8 @@ declare class SIEError extends Error {
  * - Server refuses connection
  */
 declare class SIEConnectionError extends SIEError {
-    constructor(message: string);
+    readonly kind: SIEConnectionErrorKind;
+    constructor(message: string, kind?: SIEConnectionErrorKind);
 }
 /**
  * Error in the request (4xx responses).
@@ -916,6 +948,39 @@ declare class ModelLoadingError extends SIEError {
     readonly model: string | undefined;
     constructor(message: string, model?: string);
 }
+/**
+ * Error when the server reports a *terminal* model-load failure.
+ *
+ * Distinct from {@link ModelLoadingError} — this is thrown on the first
+ * response (no retry budget consumed) when the server returns HTTP
+ * `502 MODEL_LOAD_FAILED`. The server uses this code for permanent-class
+ * failures (gated repos, missing dependencies, unrecognised model
+ * architectures) where retrying would waste time. See sie-test#85.
+ *
+ * Permanent failures will not auto-clear; an operator must fix the
+ * underlying cause (e.g. set `HF_TOKEN`, accept the model license on
+ * HuggingFace, upgrade `transformers`).
+ */
+declare class ModelLoadFailedError extends ServerError {
+    /** The model that was requested */
+    readonly model: string | undefined;
+    /**
+     * Server-side classification: one of `GATED`, `OOM`, `DEPENDENCY`,
+     * `NOT_FOUND`, `NETWORK`, `UNKNOWN`. Use this to route to specific
+     * remediation paths (e.g. surface a "set HF_TOKEN" hint for `GATED`).
+     */
+    readonly errorClass: string | undefined;
+    /** Whether the failure is non-retryable per server policy. */
+    readonly permanent: boolean;
+    /** How many load attempts the server has logged. */
+    readonly attempts: number;
+    constructor(message: string, options?: {
+        model?: string;
+        errorClass?: string;
+        permanent?: boolean;
+        attempts?: number;
+    });
+}
 
 /**
  * MessagePack serialization with msgpack-numpy compatibility.
@@ -1021,4 +1086,4 @@ declare function toImageWireFormat(input: ImageInput, format?: "jpeg" | "png" |
  */
 declare function detectImageFormat(bytes: Uint8Array): "jpeg" | "png" | "webp" | "unknown";
 
-export { type CapacityInfo, type Classification, type ClusterStatusMessage, type ClusterSummary, type ClusterWorkerInfo, type DType, type DetectedObject, type EncodeOptions, type EncodeResult, type Entity, type ExtractOptions, type ExtractResult, type GPUMetrics, type ImageInput, type ImageWireFormat, type Item, LoraLoadingError, type ModelConfig, type ModelDims, type ModelInfo, ModelLoadingError, type ModelState, type ModelStatus, type ModelSummary, type OutputType, PoolError, type PoolInfo, type PoolSpec, type PoolStatus, ProvisioningError, type Relation, RequestError, SDK_VERSION, SIEClient, type SIEClientOptions, SIEConnectionError, SIEError, type ScoreEntry, type ScoreOptions, type ScoreResult, ServerError, type ServerInfo, type SparseResult, type SparseVector, type StatusMessage, type TimingInfo, type WorkerInfo, type WorkerStatusMessage, denseEmbedding, detectImageFormat, multivectorEmbedding, normalizeSparseVector, packMessage, sparseEmbedding, sparseEmbeddingMap, toFloat32Array, toImageBytes, toImageWireFormat, toNumberArray, unpackMessage };
+export { type CapacityInfo, type Classification, type ClusterStatusMessage, type ClusterSummary, type ClusterWorkerInfo, type DType, type DetectedObject, type EncodeOptions, type EncodeResult, type Entity, type ExtractOptions, type ExtractResult, type GPUMetrics, type ImageInput, type ImageWireFormat, type Item, LoraLoadingError, type ModelConfig, type ModelDims, type ModelInfo, ModelLoadFailedError, ModelLoadingError, type ModelState, type ModelStatus, type ModelSummary, type OutputType, PoolError, type PoolInfo, type PoolSpec, type PoolStatus, ProvisioningError, type Relation, RequestError, SDK_VERSION, SIEClient, type SIEClientOptions, SIEConnectionError, SIEError, type ScoreEntry, type ScoreOptions, type ScoreResult, ServerError, type ServerInfo, type SparseResult, type SparseVector, type StatusMessage, type TimingInfo, type WorkerInfo, type WorkerStatusMessage, denseEmbedding, detectImageFormat, multivectorEmbedding, normalizeSparseVector, packMessage, sparseEmbedding, sparseEmbeddingMap, toFloat32Array, toImageBytes, toImageWireFormat, toNumberArray, unpackMessage };

package/dist/index.d.ts

@@ -15,6 +15,19 @@ type DType = "float32" | "float16" | "bfloat16" | "int8" | "uint8" | "binary" |
  * Output type options for encode operation.
  */
 type OutputType = "dense" | "sparse" | "multivector";
+/**
+ * Document input for composite-document extractors (PDF, DOCX, HTML, ...).
+ *
+ * The wire format is the document bytes plus an optional format hint. The
+ * hint is advisory — adapters may sniff the bytes when it is missing or
+ * unrecognized.
+ */
+interface DocumentInput {
+    /** Document bytes (raw file content) */
+    data: Uint8Array;
+    /** Document format hint: "pdf", "docx", "html", etc. */
+    format?: string;
+}
 /**
  * A single item to encode, score, or extract from.
  *
@@ -30,6 +43,9 @@ type OutputType = "dense" | "sparse" | "multivector";
  * // With images for multimodal models (ColPali, CLIP)
  * { text: "Description", images: [imageBytes] }
  *
+ * // With a document for composite-document extractors (Docling, ...)
+ * { document: { data: pdfBytes, format: "pdf" } }
+ *
  * // Pre-encoded multivector (for use with maxsim utility)
  * { multivector: [tokenEmbedding1, tokenEmbedding2, ...] }
  */
@@ -40,6 +56,8 @@ interface Item {
     text?: string;
     /** Images as byte arrays (JPEG/PNG) for multimodal models */
     images?: Uint8Array[];
+    /** Document for composite-document extractors (PDF, DOCX, HTML, ...) */
+    document?: DocumentInput;
     /** Pre-encoded multivector (for use with maxsim utility) */
     multivector?: Float32Array[];
     /** Arbitrary metadata (passed through to results) */
@@ -98,7 +116,7 @@ interface ModelInfo {
     name: string;
     /** Whether the model is currently loaded in memory */
     loaded: boolean;
-    /** Supported input types: ["text"], ["text", "image"], etc. */
+    /** Supported input types: ["text"], ["text", "image"], ["text", "document"], etc. */
     inputs: string[];
     /** Supported output types: ["dense"], ["dense", "sparse"], etc. */
     outputs: string[];
@@ -448,7 +466,7 @@ declare function toFloat32Array(arr: number[]): Float32Array;
  *
  * @example Resource pool usage
  * ```typescript
- * const client = new SIEClient("http://router:8080");
+ * const client = new SIEClient("http://gateway:8080");
  *
  * // Create a dedicated pool
  * await client.createPool("eval-bench", { l4: 2 });
@@ -523,9 +541,9 @@ declare class SIEClient {
      */
     getModel(name: string): Promise<ModelInfo>;
     /**
-     * Stream real-time status updates from a worker or router.
+     * Stream real-time status updates from a worker or gateway.
      *
-     * @param mode - "cluster" uses router /ws/cluster-status, "worker" uses /ws/status.
+     * @param mode - "cluster" uses gateway /ws/cluster-status, "worker" uses /ws/status.
      *               "auto" detects the endpoint via /health.
      */
     watch(mode?: "auto" | "cluster" | "worker"): AsyncGenerator<StatusMessage>;
@@ -576,7 +594,7 @@ declare class SIEClient {
      * Close the client and cleanup resources.
      *
      * Stops pool lease renewal timers. Note that pools are not deleted
-     * automatically - they are garbage collected by the router after inactivity.
+     * automatically - they are garbage collected by the gateway after inactivity.
      * This allows pool reuse if the client reconnects.
      */
     close(): Promise<void>;
@@ -643,7 +661,7 @@ declare class SIEClient {
     /**
      * Get current cluster capacity information.
      *
-     * Queries the router's /health endpoint for cluster state. Useful for
+     * Queries the gateway's /health endpoint for cluster state. Useful for
      * checking if specific GPU types are available before sending requests.
      *
      * @param gpu - Optional filter to check specific GPU type availability
@@ -666,7 +684,7 @@ declare class SIEClient {
     /**
      * Wait for GPU capacity to become available.
      *
-     * Polls the router until workers with the specified GPU type are online.
+     * Polls the gateway until workers with the specified GPU type are online.
      * This is useful for pre-warming the cluster before running benchmarks.
      *
      * @param gpu - GPU type to wait for (e.g., "l4", "a100-80gb")
@@ -689,7 +707,15 @@ declare class SIEClient {
         pollInterval?: number;
     }): Promise<CapacityInfo>;
     /**
-     * Make a msgpack HTTP request with retry logic for 202 and LoRA loading.
+     * Make a msgpack HTTP request with retry logic.
+     *
+     * Retried (only when `waitForCapacity: true`, capped by `provisionTimeout`):
+     *  - 202 Accepted (provisioning)
+     *  - 503 `MODEL_LOADING` / `LORA_LOADING` / no error code (scale-from-zero)
+     *  - `SIEConnectionError` with `kind === "connect"` (issue #95)
+     *
+     * `kind === "timeout"` is NOT retried — would extend the user-visible
+     * timeout from `timeout` to `provisionTimeout`.
      */
     private requestWithRetry;
     /**
@@ -706,7 +732,7 @@ declare class SIEClient {
     private detectEndpointType;
 }
 
-declare const SDK_VERSION = "0.2.0";
+declare const SDK_VERSION = "0.3.1";
 
 /**
  * Helpers for converting SIE encode results to plain JavaScript types.
@@ -806,6 +832,11 @@ declare function multivectorEmbedding(raw: Float32Array[]): number[][];
 declare class SIEError extends Error {
     constructor(message: string);
 }
+/**
+ * `SIEConnectionError` failure category. Only `"connect"` is auto-retried
+ * under `waitForCapacity: true`; `"timeout"` and `"other"` fail fast.
+ */
+type SIEConnectionErrorKind = "connect" | "timeout" | "other";
 /**
  * Error connecting to the SIE server.
  *
@@ -816,7 +847,8 @@ declare class SIEError extends Error {
  * - Server refuses connection
  */
 declare class SIEConnectionError extends SIEError {
-    constructor(message: string);
+    readonly kind: SIEConnectionErrorKind;
+    constructor(message: string, kind?: SIEConnectionErrorKind);
 }
 /**
  * Error in the request (4xx responses).
@@ -916,6 +948,39 @@ declare class ModelLoadingError extends SIEError {
     readonly model: string | undefined;
     constructor(message: string, model?: string);
 }
+/**
+ * Error when the server reports a *terminal* model-load failure.
+ *
+ * Distinct from {@link ModelLoadingError} — this is thrown on the first
+ * response (no retry budget consumed) when the server returns HTTP
+ * `502 MODEL_LOAD_FAILED`. The server uses this code for permanent-class
+ * failures (gated repos, missing dependencies, unrecognised model
+ * architectures) where retrying would waste time. See sie-test#85.
+ *
+ * Permanent failures will not auto-clear; an operator must fix the
+ * underlying cause (e.g. set `HF_TOKEN`, accept the model license on
+ * HuggingFace, upgrade `transformers`).
+ */
+declare class ModelLoadFailedError extends ServerError {
+    /** The model that was requested */
+    readonly model: string | undefined;
+    /**
+     * Server-side classification: one of `GATED`, `OOM`, `DEPENDENCY`,
+     * `NOT_FOUND`, `NETWORK`, `UNKNOWN`. Use this to route to specific
+     * remediation paths (e.g. surface a "set HF_TOKEN" hint for `GATED`).
+     */
+    readonly errorClass: string | undefined;
+    /** Whether the failure is non-retryable per server policy. */
+    readonly permanent: boolean;
+    /** How many load attempts the server has logged. */
+    readonly attempts: number;
+    constructor(message: string, options?: {
+        model?: string;
+        errorClass?: string;
+        permanent?: boolean;
+        attempts?: number;
+    });
+}
 
 /**
  * MessagePack serialization with msgpack-numpy compatibility.
@@ -1021,4 +1086,4 @@ declare function toImageWireFormat(input: ImageInput, format?: "jpeg" | "png" |
  */
 declare function detectImageFormat(bytes: Uint8Array): "jpeg" | "png" | "webp" | "unknown";
 
-export { type CapacityInfo, type Classification, type ClusterStatusMessage, type ClusterSummary, type ClusterWorkerInfo, type DType, type DetectedObject, type EncodeOptions, type EncodeResult, type Entity, type ExtractOptions, type ExtractResult, type GPUMetrics, type ImageInput, type ImageWireFormat, type Item, LoraLoadingError, type ModelConfig, type ModelDims, type ModelInfo, ModelLoadingError, type ModelState, type ModelStatus, type ModelSummary, type OutputType, PoolError, type PoolInfo, type PoolSpec, type PoolStatus, ProvisioningError, type Relation, RequestError, SDK_VERSION, SIEClient, type SIEClientOptions, SIEConnectionError, SIEError, type ScoreEntry, type ScoreOptions, type ScoreResult, ServerError, type ServerInfo, type SparseResult, type SparseVector, type StatusMessage, type TimingInfo, type WorkerInfo, type WorkerStatusMessage, denseEmbedding, detectImageFormat, multivectorEmbedding, normalizeSparseVector, packMessage, sparseEmbedding, sparseEmbeddingMap, toFloat32Array, toImageBytes, toImageWireFormat, toNumberArray, unpackMessage };
+export { type CapacityInfo, type Classification, type ClusterStatusMessage, type ClusterSummary, type ClusterWorkerInfo, type DType, type DetectedObject, type EncodeOptions, type EncodeResult, type Entity, type ExtractOptions, type ExtractResult, type GPUMetrics, type ImageInput, type ImageWireFormat, type Item, LoraLoadingError, type ModelConfig, type ModelDims, type ModelInfo, ModelLoadFailedError, ModelLoadingError, type ModelState, type ModelStatus, type ModelSummary, type OutputType, PoolError, type PoolInfo, type PoolSpec, type PoolStatus, ProvisioningError, type Relation, RequestError, SDK_VERSION, SIEClient, type SIEClientOptions, SIEConnectionError, SIEError, type ScoreEntry, type ScoreOptions, type ScoreResult, ServerError, type ServerInfo, type SparseResult, type SparseVector, type StatusMessage, type TimingInfo, type WorkerInfo, type WorkerStatusMessage, denseEmbedding, detectImageFormat, multivectorEmbedding, normalizeSparseVector, packMessage, sparseEmbedding, sparseEmbeddingMap, toFloat32Array, toImageBytes, toImageWireFormat, toNumberArray, unpackMessage };

package/dist/index.js

@@ -9,9 +9,11 @@ var SIEError = class extends Error {
   }
 };
 var SIEConnectionError = class extends SIEError {
-  constructor(message) {
+  kind;
+  constructor(message, kind = "other") {
     super(message);
     this.name = "SIEConnectionError";
+    this.kind = kind;
   }
 };
 var RequestError = class extends SIEError {
@@ -83,6 +85,28 @@ var ModelLoadingError = class extends SIEError {
     this.model = model;
   }
 };
+var ModelLoadFailedError = class extends ServerError {
+  /** The model that was requested */
+  model;
+  /**
+   * Server-side classification: one of `GATED`, `OOM`, `DEPENDENCY`,
+   * `NOT_FOUND`, `NETWORK`, `UNKNOWN`. Use this to route to specific
+   * remediation paths (e.g. surface a "set HF_TOKEN" hint for `GATED`).
+   */
+  errorClass;
+  /** Whether the failure is non-retryable per server policy. */
+  permanent;
+  /** How many load attempts the server has logged. */
+  attempts;
+  constructor(message, options) {
+    super(message, "MODEL_LOAD_FAILED", 502);
+    this.name = "ModelLoadFailedError";
+    this.model = options?.model;
+    this.errorClass = options?.errorClass;
+    this.permanent = options?.permanent ?? true;
+    this.attempts = options?.attempts ?? 1;
+  }
+};
 
 // src/internal/constants.ts
 var MSGPACK_CONTENT_TYPE = "application/msgpack";
@@ -103,9 +127,6 @@ var MODEL_LOADING_DEFAULT_DELAY = 5e3;
 var MODEL_LOADING_ERROR_CODE = "MODEL_LOADING";
 var SDK_VERSION_HEADER = "X-SIE-SDK-Version";
 var SERVER_VERSION_HEADER = "X-SIE-Server-Version";
-
-// src/version.ts
-var SDK_VERSION = "0.2.0";
 var EXT_TYPE_NUMPY = 78;
 function parseDtype(dtype) {
   const typeChar = dtype.slice(-2, -1);
@@ -310,7 +331,7 @@ function getRetryAfter2(response) {
   const header = response.headers.get("Retry-After");
   return getRetryAfter(header);
 }
-async function getErrorCode(response) {
+async function getErrorDetail(response) {
   try {
     const contentType = response.headers.get("content-type") ?? "";
     let data;
@@ -321,24 +342,42 @@ async function getErrorCode(response) {
       data = await response.json();
     }
     if (data.error && typeof data.error === "object") {
-      const error = data.error;
-      if (typeof error.code === "string") {
-        return error.code;
-      }
+      return data.error;
     }
     if (data.detail && typeof data.detail === "object") {
-      const detail = data.detail;
-      if (typeof detail.code === "string") {
-        return detail.code;
-      }
+      return data.detail;
     }
     if (typeof data.code === "string") {
-      return data.code;
+      return data;
     }
   } catch {
   }
   return void 0;
 }
+async function getErrorCode(response) {
+  const detail = await getErrorDetail(response);
+  if (!detail) return void 0;
+  const code = detail.code;
+  return typeof code === "string" ? code : void 0;
+}
+async function throwIfModelLoadFailed(response, model) {
+  if (response.status !== 502) return;
+  const detail = await getErrorDetail(response.clone());
+  if (!detail) return;
+  if (detail.code !== "MODEL_LOAD_FAILED") return;
+  const errorClass = typeof detail.error_class === "string" ? detail.error_class : void 0;
+  const permanent = typeof detail.permanent === "boolean" ? detail.permanent : true;
+  const attemptsRaw = detail.attempts;
+  const parsedAttempts = typeof attemptsRaw === "number" ? attemptsRaw : typeof attemptsRaw === "string" ? Number.parseInt(attemptsRaw, 10) : Number.NaN;
+  const attempts = Number.isFinite(parsedAttempts) ? parsedAttempts : 1;
+  const message = typeof detail.message === "string" ? detail.message : `Model '${model ?? "?"}' failed to load`;
+  throw new ModelLoadFailedError(message, {
+    model,
+    errorClass,
+    permanent,
+    attempts
+  });
+}
 async function handleError(response, gpu) {
   const { status } = response;
   let errorBody = {};
@@ -474,6 +513,9 @@ function parseCapacityInfo(data, gpuFilter) {
   };
 }
 
+// src/version.ts
+var SDK_VERSION = "0.3.1";
+
 // src/client.ts
 function sleep(ms) {
   return new Promise((resolve) => setTimeout(resolve, ms));
@@ -614,9 +656,9 @@ var SIEClient = class {
     };
   }
   /**
-   * Stream real-time status updates from a worker or router.
+   * Stream real-time status updates from a worker or gateway.
    *
-   * @param mode - "cluster" uses router /ws/cluster-status, "worker" uses /ws/status.
+   * @param mode - "cluster" uses gateway /ws/cluster-status, "worker" uses /ws/status.
    *               "auto" detects the endpoint via /health.
    */
   async *watch(mode = "auto") {
@@ -787,7 +829,7 @@ var SIEClient = class {
    * Close the client and cleanup resources.
    *
    * Stops pool lease renewal timers. Note that pools are not deleted
-   * automatically - they are garbage collected by the router after inactivity.
+   * automatically - they are garbage collected by the gateway after inactivity.
    * This allows pool reuse if the client reconnects.
    */
   async close() {
@@ -1044,7 +1086,7 @@ var SIEClient = class {
   /**
    * Get current cluster capacity information.
    *
-   * Queries the router's /health endpoint for cluster state. Useful for
+   * Queries the gateway's /health endpoint for cluster state. Useful for
    * checking if specific GPU types are available before sending requests.
    *
    * @param gpu - Optional filter to check specific GPU type availability
@@ -1066,10 +1108,10 @@ var SIEClient = class {
   async getCapacity(gpu) {
     const response = await this.requestJson("/health");
     const data = await response.json();
-    if (data.type !== "router") {
+    if (data.type !== "gateway") {
       throw new RequestError(
-        "getCapacity() requires a router endpoint. This appears to be a worker.",
-        "not_router",
+        "getCapacity() requires a gateway endpoint. This appears to be a worker.",
+        "not_gateway",
         400
       );
     }
@@ -1078,7 +1120,7 @@ var SIEClient = class {
   /**
    * Wait for GPU capacity to become available.
    *
-   * Polls the router until workers with the specified GPU type are online.
+   * Polls the gateway until workers with the specified GPU type are online.
    * This is useful for pre-warming the cluster before running benchmarks.
    *
    * @param gpu - GPU type to wait for (e.g., "l4", "a100-80gb")
@@ -1124,13 +1166,35 @@ var SIEClient = class {
     }
   }
   /**
-   * Make a msgpack HTTP request with retry logic for 202 and LoRA loading.
+   * Make a msgpack HTTP request with retry logic.
+   *
+   * Retried (only when `waitForCapacity: true`, capped by `provisionTimeout`):
+   *  - 202 Accepted (provisioning)
+   *  - 503 `MODEL_LOADING` / `LORA_LOADING` / no error code (scale-from-zero)
+   *  - `SIEConnectionError` with `kind === "connect"` (issue #95)
+   *
+   * `kind === "timeout"` is NOT retried — would extend the user-visible
+   * timeout from `timeout` to `provisionTimeout`.
    */
   async requestWithRetry(path, body, pool, gpu, waitForCapacity, model) {
     const startTime = Date.now();
     let loraRetries = 0;
     while (true) {
-      const response = await this.request(path, body, pool, gpu);
+      let response;
+      try {
+        response = await this.request(path, body, pool, gpu);
+      } catch (err) {
+        if (waitForCapacity && err instanceof SIEConnectionError && err.kind === "connect") {
+          const elapsed = Date.now() - startTime;
+          if (elapsed < this.provisionTimeout) {
+            const remaining = this.provisionTimeout - elapsed;
+            const delay = Math.min(DEFAULT_RETRY_DELAY, remaining);
+            await sleep(delay);
+            continue;
+          }
+        }
+        throw err;
+      }
       if (response.status === HTTP_ACCEPTED) {
         const retryAfter = getRetryAfter2(response);
         if (!waitForCapacity) {
@@ -1154,6 +1218,7 @@ var SIEClient = class {
         await sleep(actualDelay);
         continue;
       }
+      await throwIfModelLoadFailed(response, model);
       if (response.status === 503) {
         const clonedResponse = response.clone();
         const errorCode = await getErrorCode(clonedResponse);
@@ -1187,6 +1252,17 @@ var SIEClient = class {
           await sleep(actualDelay);
           continue;
         }
+        if (waitForCapacity) {
+          const elapsed = Date.now() - startTime;
+          if (elapsed < this.provisionTimeout) {
+            const retryAfter = getRetryAfter2(response);
+            const delay = retryAfter ?? DEFAULT_RETRY_DELAY;
+            const remaining = this.provisionTimeout - elapsed;
+            const actualDelay = Math.min(delay, remaining);
+            await sleep(actualDelay);
+            continue;
+          }
+        }
       }
       if (!response.ok) {
         await handleError(response, gpu);
@@ -1228,10 +1304,10 @@ var SIEClient = class {
       return response;
     } catch (error) {
       if (error instanceof Error && error.name === "AbortError") {
-        throw new SIEConnectionError(`Request timeout after ${this.timeout}ms`);
+        throw new SIEConnectionError(`Request timeout after ${this.timeout}ms`, "timeout");
       }
       if (error instanceof TypeError) {
-        throw new SIEConnectionError(`Connection failed: ${error.message}`);
+        throw new SIEConnectionError(`Connection failed: ${error.message}`, "connect");
       }
       throw error;
     } finally {
@@ -1265,10 +1341,10 @@ var SIEClient = class {
       return response;
     } catch (error) {
       if (error instanceof Error && error.name === "AbortError") {
-        throw new SIEConnectionError(`Request timeout after ${this.timeout}ms`);
+        throw new SIEConnectionError(`Request timeout after ${this.timeout}ms`, "timeout");
       }
       if (error instanceof TypeError) {
-        throw new SIEConnectionError(`Connection failed: ${error.message}`);
+        throw new SIEConnectionError(`Connection failed: ${error.message}`, "connect");
       }
       throw error;
     } finally {
@@ -1316,7 +1392,7 @@ var SIEClient = class {
         return "worker";
       }
       const data = await response.json();
-      return data.type === "router" ? "cluster" : "worker";
+      return data.type === "gateway" ? "cluster" : "worker";
     } catch {
       return "worker";
     } finally {
@@ -1464,6 +1540,6 @@ function detectImageFormat(bytes) {
   return "unknown";
 }
 
-export { LoraLoadingError, ModelLoadingError, PoolError, ProvisioningError, RequestError, SDK_VERSION, SIEClient, SIEConnectionError, SIEError, ServerError, denseEmbedding, detectImageFormat, maxsim, maxsimBatch, maxsimDocuments, multivectorEmbedding, normalizeSparseVector, packMessage, sparseEmbedding, sparseEmbeddingMap, toFloat32Array, toImageBytes, toImageWireFormat, toNumberArray, unpackMessage };
+export { LoraLoadingError, ModelLoadFailedError, ModelLoadingError, PoolError, ProvisioningError, RequestError, SDK_VERSION, SIEClient, SIEConnectionError, SIEError, ServerError, denseEmbedding, detectImageFormat, maxsim, maxsimBatch, maxsimDocuments, multivectorEmbedding, normalizeSparseVector, packMessage, sparseEmbedding, sparseEmbeddingMap, toFloat32Array, toImageBytes, toImageWireFormat, toNumberArray, unpackMessage };
 //# sourceMappingURL=index.js.map
 //# sourceMappingURL=index.js.map