@superlinked/sie-sdk 0.6.22 → 0.6.24

package/dist/index.d.cts

@@ -879,10 +879,38 @@ interface GenerationUsage {
     completionTokens: number;
     totalTokens: number;
 }
+interface GrammarMetadata {
+    /** Optional schema/grammar label used by structured-output backends. */
+    label?: string | null;
+    /** Optional strictness hint for structured-output backends. */
+    strict?: boolean | null;
+}
+/** Constrain native generation to a JSON Schema. */
+type JsonSchemaGrammar = GrammarMetadata & {
+    json_schema: Record<string, unknown>;
+    regex?: never;
+    ebnf?: never;
+};
+/** Constrain native generation to a regular expression. */
+type RegexGrammar = GrammarMetadata & {
+    json_schema?: never;
+    regex: string;
+    ebnf?: never;
+};
+/** Constrain native generation to an EBNF grammar. */
+type EbnfGrammar = GrammarMetadata & {
+    json_schema?: never;
+    regex?: never;
+    ebnf: string;
+};
+/** Native structured-output grammar. Exactly one grammar variant is set. */
+type GenerateGrammar = JsonSchemaGrammar | RegexGrammar | EbnfGrammar;
 /** Options for the generate operation. */
 interface GenerateOptions {
     /** Hard cap on output tokens. Required. */
     maxNewTokens: number;
+    /** Optional native image inputs rendered with the prompt for vision-capable models. */
+    images?: (ImageInput | ImageWireFormat)[];
     /** Sampling temperature. */
     temperature?: number;
     /** Nucleus sampling cutoff. */
@@ -893,8 +921,14 @@ interface GenerateOptions {
     frequencyPenalty?: number;
     /** OpenAI-compatible presence penalty in [-2, 2]. */
     presencePenalty?: number;
-    /** Native structured-output grammar. */
-    grammar?: Record<string, unknown>;
+    /**
+     * Native structured-output grammar.
+     *
+     * The broad record arm preserves the pre-existing SDK input contract for
+     * callers that keep valid grammar objects in `Record<string, unknown>`
+     * variables. The SDK still validates the exact three-arm shape at runtime.
+     */
+    grammar?: GenerateGrammar | Record<string, unknown>;
     /**
      * Optional per-request sampling seed. Must be a JavaScript safe integer
      * (-(2^53 - 1) through 2^53 - 1) so JSON serialization preserves it exactly.
@@ -1887,7 +1921,7 @@ declare class SIEClient {
     private detectEndpointType;
 }
 
-declare const SDK_VERSION = "0.6.22";
+declare const SDK_VERSION = "0.6.24";
 
 /**
  * Helpers for converting SIE encode results to plain JavaScript types.
@@ -2232,4 +2266,4 @@ declare function packMessage(data: unknown): Uint8Array;
  */
 declare function unpackMessage<T = unknown>(data: Uint8Array): T;
 
-export { type AudioInput, type Batch, type BatchList, type BatchRequestCounts, type BatchesNamespace, type CapacityInfo, type ChatChoice, type ChatChunkChoice, type ChatCompletion, type ChatCompletionChunk, type ChatCompletionRequest, type ChatDelta, type ChatFinishReason, type ChatMessage, type ChatUsage, type Classification, type ClusterStatusMessage, type ClusterSummary, type ClusterWorkerInfo, type Connection, type ConnectionCreated, type ConnectionRevoked, type ConnectionsNamespace, type CreatePoolOptions, type DType, type DetectedObject, type EncodeOptions, type EncodeResult, type Entity, type ExtractItem, type ExtractItemError, type ExtractOptions, type ExtractResult, type FileDeleted, type FileUploadInput, type FilesNamespace, type FinishReason, type GPUMetrics, type GenerateChunk, type GenerateOptions, type GenerateResult, type GenerationUsage, type ImageInput, type ImageWireFormat, InputTooLongError, type Item, type JobChunk, type JobFieldMap, type JobItem, type JobPreflight, type JobResultItem, type JobResults, type JobSource, type JobState, type JobStatus, type JobSubmitResult, type JobsNamespace, LoraLoadingError, type ModelCapabilities, 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, ResourceExhaustedError, type ResponseFormat, SDK_VERSION, SIEClient, type SIEClientOptions, SIEConnectionError, SIEError, type File as SIEFile, SIEStreamError, type ScoreEntry, type ScoreOptions, type ScoreResult, type ScoreUsage, ServerError, type ServerInfo, type SparseResult, type SparseVector, type StatusMessage, type StreamGenerateOptions, type SubmitJobOptions, TERMINAL_JOB_STATES, type TimingInfo, type ToolCall, type ToolCallDelta, type ToolChoice, type ToolSpec, type WorkerInfo, type WorkerStatusMessage, buildJobBody, connectionName, denseEmbedding, detectImageFormat, multivectorEmbedding, normalizeSparseVector, packMessage, sparseEmbedding, sparseEmbeddingMap, toFloat32Array, toImageBytes, toImageWireFormat, toNumberArray, unpackMessage };
+export { type AudioInput, type Batch, type BatchList, type BatchRequestCounts, type BatchesNamespace, type CapacityInfo, type ChatChoice, type ChatChunkChoice, type ChatCompletion, type ChatCompletionChunk, type ChatCompletionRequest, type ChatDelta, type ChatFinishReason, type ChatMessage, type ChatUsage, type Classification, type ClusterStatusMessage, type ClusterSummary, type ClusterWorkerInfo, type Connection, type ConnectionCreated, type ConnectionRevoked, type ConnectionsNamespace, type CreatePoolOptions, type DType, type DetectedObject, type EbnfGrammar, type EncodeOptions, type EncodeResult, type Entity, type ExtractItem, type ExtractItemError, type ExtractOptions, type ExtractResult, type FileDeleted, type FileUploadInput, type FilesNamespace, type FinishReason, type GPUMetrics, type GenerateChunk, type GenerateGrammar, type GenerateOptions, type GenerateResult, type GenerationUsage, type ImageInput, type ImageWireFormat, InputTooLongError, type Item, type JobChunk, type JobFieldMap, type JobItem, type JobPreflight, type JobResultItem, type JobResults, type JobSource, type JobState, type JobStatus, type JobSubmitResult, type JobsNamespace, type JsonSchemaGrammar, LoraLoadingError, type ModelCapabilities, 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 RegexGrammar, type Relation, RequestError, ResourceExhaustedError, type ResponseFormat, SDK_VERSION, SIEClient, type SIEClientOptions, SIEConnectionError, SIEError, type File as SIEFile, SIEStreamError, type ScoreEntry, type ScoreOptions, type ScoreResult, type ScoreUsage, ServerError, type ServerInfo, type SparseResult, type SparseVector, type StatusMessage, type StreamGenerateOptions, type SubmitJobOptions, TERMINAL_JOB_STATES, type TimingInfo, type ToolCall, type ToolCallDelta, type ToolChoice, type ToolSpec, type WorkerInfo, type WorkerStatusMessage, buildJobBody, connectionName, denseEmbedding, detectImageFormat, multivectorEmbedding, normalizeSparseVector, packMessage, sparseEmbedding, sparseEmbeddingMap, toFloat32Array, toImageBytes, toImageWireFormat, toNumberArray, unpackMessage };

package/dist/index.d.ts

@@ -879,10 +879,38 @@ interface GenerationUsage {
     completionTokens: number;
     totalTokens: number;
 }
+interface GrammarMetadata {
+    /** Optional schema/grammar label used by structured-output backends. */
+    label?: string | null;
+    /** Optional strictness hint for structured-output backends. */
+    strict?: boolean | null;
+}
+/** Constrain native generation to a JSON Schema. */
+type JsonSchemaGrammar = GrammarMetadata & {
+    json_schema: Record<string, unknown>;
+    regex?: never;
+    ebnf?: never;
+};
+/** Constrain native generation to a regular expression. */
+type RegexGrammar = GrammarMetadata & {
+    json_schema?: never;
+    regex: string;
+    ebnf?: never;
+};
+/** Constrain native generation to an EBNF grammar. */
+type EbnfGrammar = GrammarMetadata & {
+    json_schema?: never;
+    regex?: never;
+    ebnf: string;
+};
+/** Native structured-output grammar. Exactly one grammar variant is set. */
+type GenerateGrammar = JsonSchemaGrammar | RegexGrammar | EbnfGrammar;
 /** Options for the generate operation. */
 interface GenerateOptions {
     /** Hard cap on output tokens. Required. */
     maxNewTokens: number;
+    /** Optional native image inputs rendered with the prompt for vision-capable models. */
+    images?: (ImageInput | ImageWireFormat)[];
     /** Sampling temperature. */
     temperature?: number;
     /** Nucleus sampling cutoff. */
@@ -893,8 +921,14 @@ interface GenerateOptions {
     frequencyPenalty?: number;
     /** OpenAI-compatible presence penalty in [-2, 2]. */
     presencePenalty?: number;
-    /** Native structured-output grammar. */
-    grammar?: Record<string, unknown>;
+    /**
+     * Native structured-output grammar.
+     *
+     * The broad record arm preserves the pre-existing SDK input contract for
+     * callers that keep valid grammar objects in `Record<string, unknown>`
+     * variables. The SDK still validates the exact three-arm shape at runtime.
+     */
+    grammar?: GenerateGrammar | Record<string, unknown>;
     /**
      * Optional per-request sampling seed. Must be a JavaScript safe integer
      * (-(2^53 - 1) through 2^53 - 1) so JSON serialization preserves it exactly.
@@ -1887,7 +1921,7 @@ declare class SIEClient {
     private detectEndpointType;
 }
 
-declare const SDK_VERSION = "0.6.22";
+declare const SDK_VERSION = "0.6.24";
 
 /**
  * Helpers for converting SIE encode results to plain JavaScript types.
@@ -2232,4 +2266,4 @@ declare function packMessage(data: unknown): Uint8Array;
  */
 declare function unpackMessage<T = unknown>(data: Uint8Array): T;
 
-export { type AudioInput, type Batch, type BatchList, type BatchRequestCounts, type BatchesNamespace, type CapacityInfo, type ChatChoice, type ChatChunkChoice, type ChatCompletion, type ChatCompletionChunk, type ChatCompletionRequest, type ChatDelta, type ChatFinishReason, type ChatMessage, type ChatUsage, type Classification, type ClusterStatusMessage, type ClusterSummary, type ClusterWorkerInfo, type Connection, type ConnectionCreated, type ConnectionRevoked, type ConnectionsNamespace, type CreatePoolOptions, type DType, type DetectedObject, type EncodeOptions, type EncodeResult, type Entity, type ExtractItem, type ExtractItemError, type ExtractOptions, type ExtractResult, type FileDeleted, type FileUploadInput, type FilesNamespace, type FinishReason, type GPUMetrics, type GenerateChunk, type GenerateOptions, type GenerateResult, type GenerationUsage, type ImageInput, type ImageWireFormat, InputTooLongError, type Item, type JobChunk, type JobFieldMap, type JobItem, type JobPreflight, type JobResultItem, type JobResults, type JobSource, type JobState, type JobStatus, type JobSubmitResult, type JobsNamespace, LoraLoadingError, type ModelCapabilities, 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, ResourceExhaustedError, type ResponseFormat, SDK_VERSION, SIEClient, type SIEClientOptions, SIEConnectionError, SIEError, type File as SIEFile, SIEStreamError, type ScoreEntry, type ScoreOptions, type ScoreResult, type ScoreUsage, ServerError, type ServerInfo, type SparseResult, type SparseVector, type StatusMessage, type StreamGenerateOptions, type SubmitJobOptions, TERMINAL_JOB_STATES, type TimingInfo, type ToolCall, type ToolCallDelta, type ToolChoice, type ToolSpec, type WorkerInfo, type WorkerStatusMessage, buildJobBody, connectionName, denseEmbedding, detectImageFormat, multivectorEmbedding, normalizeSparseVector, packMessage, sparseEmbedding, sparseEmbeddingMap, toFloat32Array, toImageBytes, toImageWireFormat, toNumberArray, unpackMessage };
+export { type AudioInput, type Batch, type BatchList, type BatchRequestCounts, type BatchesNamespace, type CapacityInfo, type ChatChoice, type ChatChunkChoice, type ChatCompletion, type ChatCompletionChunk, type ChatCompletionRequest, type ChatDelta, type ChatFinishReason, type ChatMessage, type ChatUsage, type Classification, type ClusterStatusMessage, type ClusterSummary, type ClusterWorkerInfo, type Connection, type ConnectionCreated, type ConnectionRevoked, type ConnectionsNamespace, type CreatePoolOptions, type DType, type DetectedObject, type EbnfGrammar, type EncodeOptions, type EncodeResult, type Entity, type ExtractItem, type ExtractItemError, type ExtractOptions, type ExtractResult, type FileDeleted, type FileUploadInput, type FilesNamespace, type FinishReason, type GPUMetrics, type GenerateChunk, type GenerateGrammar, type GenerateOptions, type GenerateResult, type GenerationUsage, type ImageInput, type ImageWireFormat, InputTooLongError, type Item, type JobChunk, type JobFieldMap, type JobItem, type JobPreflight, type JobResultItem, type JobResults, type JobSource, type JobState, type JobStatus, type JobSubmitResult, type JobsNamespace, type JsonSchemaGrammar, LoraLoadingError, type ModelCapabilities, 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 RegexGrammar, type Relation, RequestError, ResourceExhaustedError, type ResponseFormat, SDK_VERSION, SIEClient, type SIEClientOptions, SIEConnectionError, SIEError, type File as SIEFile, SIEStreamError, type ScoreEntry, type ScoreOptions, type ScoreResult, type ScoreUsage, ServerError, type ServerInfo, type SparseResult, type SparseVector, type StatusMessage, type StreamGenerateOptions, type SubmitJobOptions, TERMINAL_JOB_STATES, type TimingInfo, type ToolCall, type ToolCallDelta, type ToolChoice, type ToolSpec, type WorkerInfo, type WorkerStatusMessage, buildJobBody, connectionName, denseEmbedding, detectImageFormat, multivectorEmbedding, normalizeSparseVector, packMessage, sparseEmbedding, sparseEmbeddingMap, toFloat32Array, toImageBytes, toImageWireFormat, toNumberArray, unpackMessage };

package/dist/index.js

@@ -1189,7 +1189,7 @@ function extractDataPayload(block) {
 }
 
 // src/version.ts
-var SDK_VERSION = "0.6.22";
+var SDK_VERSION = "0.6.24";
 
 // src/client.ts
 function sleep2(ms) {
@@ -1243,6 +1243,33 @@ function validateGenerationSeed(seed) {
   }
   return seed;
 }
+function validateGenerateGrammar(grammar) {
+  if (typeof grammar !== "object" || grammar === null || Array.isArray(grammar)) {
+    throw new TypeError("grammar must be an object");
+  }
+  const allowed = /* @__PURE__ */ new Set(["json_schema", "regex", "ebnf", "label", "strict"]);
+  const unknown = Object.keys(grammar).filter((key) => !allowed.has(key));
+  if (unknown.length > 0) {
+    throw new TypeError(`grammar contains unsupported field(s): ${unknown.sort().join(", ")}`);
+  }
+  const variants = ["json_schema", "regex", "ebnf"].filter((key) => Object.hasOwn(grammar, key));
+  if (variants.length !== 1) {
+    throw new TypeError("grammar must contain exactly one of json_schema, regex, or ebnf");
+  }
+  const variant = variants[0];
+  const value = grammar[variant];
+  if (variant === "json_schema" ? typeof value !== "object" || value === null || Array.isArray(value) : typeof value !== "string") {
+    throw new TypeError(
+      variant === "json_schema" ? "grammar.json_schema must be an object" : `grammar.${variant} must be a string`
+    );
+  }
+  if (grammar.label !== void 0 && grammar.label !== null && typeof grammar.label !== "string") {
+    throw new TypeError("grammar.label must be a string");
+  }
+  if (grammar.strict !== void 0 && grammar.strict !== null && typeof grammar.strict !== "boolean") {
+    throw new TypeError("grammar.strict must be a boolean");
+  }
+}
 function applyGenerateOptions(body, options) {
   if (options.temperature !== void 0) body.temperature = options.temperature;
   if (options.topP !== void 0) body.top_p = options.topP;
@@ -1250,7 +1277,10 @@ function applyGenerateOptions(body, options) {
   if (options.stop !== void 0) body.stop = options.stop;
   if (options.frequencyPenalty !== void 0) body.frequency_penalty = options.frequencyPenalty;
   if (options.presencePenalty !== void 0) body.presence_penalty = options.presencePenalty;
-  if (options.grammar !== void 0) body.grammar = options.grammar;
+  if (options.grammar !== void 0) {
+    validateGenerateGrammar(options.grammar);
+    body.grammar = options.grammar;
+  }
   if (options.seed !== void 0) body.seed = validateGenerationSeed(options.seed);
   if (options.logitBias !== void 0) body.logit_bias = options.logitBias;
   if (options.routingKey !== void 0) body.routing_key = options.routingKey;
@@ -1290,6 +1320,25 @@ async function imageForWire(image) {
   }
   return toImageWireFormat(image);
 }
+function imageBytesToBase64(data) {
+  if (typeof Buffer !== "undefined") {
+    return Buffer.from(data.buffer, data.byteOffset, data.byteLength).toString("base64");
+  }
+  let binary = "";
+  const chunkSize = 32768;
+  for (let offset = 0; offset < data.length; offset += chunkSize) {
+    binary += String.fromCharCode(...data.subarray(offset, offset + chunkSize));
+  }
+  return btoa(binary);
+}
+async function generationImagesForWire(images) {
+  return Promise.all(
+    images.map(async (image) => {
+      const wire = await imageForWire(image);
+      return { data: imageBytesToBase64(wire.data), format: wire.format };
+    })
+  );
+}
 async function itemImagesForWire(item) {
   if (!item.images || item.images.length === 0) {
     return item;
@@ -1621,6 +1670,9 @@ var SIEClient = class {
       prompt,
       max_new_tokens: options.maxNewTokens
     };
+    if (options.images !== void 0) {
+      body.images = await generationImagesForWire(options.images);
+    }
     applyGenerateOptions(body, options);
     const { pool, gpu } = this.parseGpuParam(options.gpu);
     const headers = {
@@ -1841,6 +1893,9 @@ var SIEClient = class {
       max_new_tokens: options.maxNewTokens,
       stream: true
     };
+    if (options.images !== void 0) {
+      body.images = await generationImagesForWire(options.images);
+    }
     applyGenerateOptions(body, options);
     if (options.logprobs !== void 0) body.logprobs = options.logprobs;
     if (options.topLogprobs !== void 0) {