@superlinked/sie-sdk 0.6.20 → 0.6.21

package/dist/index.d.cts

@@ -256,6 +256,15 @@ interface DocumentInput {
     /** Document format hint: "pdf", "docx", "html", etc. */
     format?: string;
 }
+/** Encoded audio bytes and optional decoder metadata. */
+interface AudioInput {
+    /** Encoded audio bytes (WAV, MP3, FLAC, etc.) */
+    data: Uint8Array;
+    /** Audio container/codec format hint */
+    format?: string;
+    /** Source sample rate in Hz, when known */
+    sampleRate?: number;
+}
 /**
  * A single item to encode, score, or extract from.
  *
@@ -291,6 +300,11 @@ interface Item {
     /** Arbitrary metadata (passed through to results) */
     metadata?: Record<string, unknown>;
 }
+/** An item accepted by extract(), including optional encoded audio. */
+interface ExtractItem extends Item {
+    /** Encoded audio for speech extraction, optionally with decoder metadata. */
+    audio?: AudioInput | Uint8Array;
+}
 /**
  * Sparse vector result with non-zero indices and values.
  * Used by SPLADE-type models.
@@ -310,6 +324,23 @@ interface TimingInfo {
     tokenizationMs?: number;
     inferenceMs?: number;
 }
+/** Authoritative units settled for one completed HTTP request. */
+interface RequestUsage {
+    inputTokens?: number;
+    pairs?: number;
+    images?: number;
+    pages?: number;
+    outputTokens?: number;
+    audioMs?: number;
+}
+/** Optional gateway metadata from the successful terminal response. */
+interface RequestMetadata {
+    id?: string;
+    /** Worker-origin immutable release/runtime identity digest. */
+    executionIdentitySha256?: string;
+    usage?: RequestUsage;
+    creditsDebited?: number;
+}
 /**
  * Result of encoding a single item.
  *
@@ -327,6 +358,8 @@ interface EncodeResult {
     multivector?: Float32Array[];
     /** Server-side timing breakdown */
     timing?: TimingInfo;
+    /** Request-scoped usage and settled debit, when supplied by the gateway. */
+    request?: RequestMetadata;
 }
 /**
  * Model dimension information.
@@ -397,6 +430,12 @@ interface ScoreEntry {
     /** Position in sorted order (0 = most relevant) */
     rank: number;
 }
+interface ScoreUsage {
+    /** Post-truncation input tokens processed */
+    inputTokens: number;
+    /** Images processed across query-document pairs */
+    images?: number;
+}
 /**
  * Result of scoring items against a query.
  */
@@ -407,6 +446,10 @@ interface ScoreResult {
     queryId?: string;
     /** Score entries, sorted by relevance (descending) */
     scores: ScoreEntry[];
+    /** Authoritative usage when emitted by the score adapter */
+    usage?: ScoreUsage;
+    /** Request-scoped usage and settled debit, when supplied by the gateway. */
+    request?: RequestMetadata;
 }
 /**
  * A single extracted entity (NER span).
@@ -458,6 +501,13 @@ interface DetectedObject {
     /** Bounding box [x, y, width, height] */
     bbox: number[];
 }
+/** Stable per-item extraction failure. */
+interface ExtractItemError {
+    /** Stable extraction error code */
+    code: string;
+    /** Sanitized extraction error message */
+    message: string;
+}
 /**
  * Result of extraction for a single item.
  */
@@ -472,6 +522,12 @@ interface ExtractResult {
     classifications: Classification[];
     /** List of detected objects */
     objects: DetectedObject[];
+    /** Additional structured extraction data */
+    data?: Record<string, unknown>;
+    /** Stable per-item failure when extraction did not complete */
+    error?: ExtractItemError;
+    /** Request-scoped usage and settled debit, when supplied by the gateway. */
+    request?: RequestMetadata;
 }
 /**
  * Information about a worker in the cluster.
@@ -524,8 +580,9 @@ interface CreatePoolOptions {
     /** Optional bundle filter. When set, only workers running this bundle will be assigned to the pool. */
     bundle?: string;
     /**
-     * Per-pool warm floor (minimum machines kept warm). The gateway publishes
-     * it as `sie_gateway_pool_warm_floor` for KEDA. Defaults to 0 (scale to zero).
+     * Per-pool warm floor (minimum machines kept warm). The gateway emits
+     * canonical `sie.gateway.pool.warm_floor` telemetry; the collector exposes
+     * `sie_gateway_pool_warm_floor` to KEDA. Defaults to 0 (scale to zero).
      */
     minimumWorkerCount?: number;
     /**
@@ -650,6 +707,10 @@ interface ModelStatus {
     queue_depth: number;
     queue_pending_items: number;
 }
+/**
+ * Worker status snapshot. Request-rate and latency telemetry is exported
+ * through the configured OTel destination instead of this message.
+ */
 interface WorkerStatusMessage {
     timestamp: number;
     name: string;
@@ -661,8 +722,6 @@ interface WorkerStatusMessage {
     server: ServerInfo;
     gpus: GPUMetrics[];
     models: ModelStatus[];
-    counters: Record<string, Record<string, number>>;
-    histograms: Record<string, Record<string, Record<string, unknown>>>;
 }
 interface ClusterStatusMessage {
     timestamp: number;
@@ -830,11 +889,45 @@ interface GenerateOptions {
     topP?: number;
     /** Optional list of stop strings. */
     stop?: string[];
+    /** OpenAI-compatible frequency penalty in [-2, 2]. */
+    frequencyPenalty?: number;
+    /** OpenAI-compatible presence penalty in [-2, 2]. */
+    presencePenalty?: number;
+    /** Native structured-output grammar. */
+    grammar?: 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.
+     * Reproducibility depends on the active backend and deployment configuration.
+     */
+    seed?: number;
+    /** Token-id-to-bias map. */
+    logitBias?: Record<string, number>;
+    /** Stable request routing key. */
+    routingKey?: string;
+    /** Prompt cache affinity key. */
+    promptCacheKey?: string;
+    /** Opaque privacy-preserving safety identifier. */
+    safetyIdentifier?: string;
+    /** Served LoRA adapter name. */
+    loraAdapter?: string;
+    /**
+     * Governed generation runtime options forwarded as `options`. Typed native
+     * fields still win when both surfaces provide the same sampler control.
+     */
+    adapterOptions?: Record<string, unknown>;
     /** GPU type / pool spec, e.g. ``"l4"`` or ``"eval-bench/l4"``. */
     gpu?: string;
     /** Auto-retry under provisioning. */
     waitForCapacity?: boolean;
 }
+/** Options for streaming native generation. */
+interface StreamGenerateOptions extends GenerateOptions {
+    /** Include per-token log probabilities in streamed chunks. */
+    logprobs?: boolean;
+    /** Number of alternate token log probabilities to return, in [0, 20]. */
+    topLogprobs?: number;
+}
 /** Aggregated generation result. */
 interface GenerateResult {
     /** Model id the gateway dispatched to. */
@@ -851,6 +944,8 @@ interface GenerateResult {
     ttftMs?: number;
     /** Average time per output token in milliseconds. */
     tpotMs?: number;
+    /** Request-scoped usage and settled debit, when supplied by the gateway. */
+    request?: RequestMetadata;
 }
 /**
  * A single message in a chat completion request.
@@ -1009,6 +1104,12 @@ interface ChatCompletionRequest {
      * `[-100, 100]` and caps map size.
      */
     logit_bias?: Record<string, number>;
+    /**
+     * Optional per-request sampling seed. Must be a JavaScript safe integer
+     * (`-(2^53 - 1)` through `2^53 - 1`) so the caller's integer is represented
+     * exactly. The SDK throws `RangeError` before network I/O for invalid values.
+     * Reproducibility depends on the active backend and deployment configuration.
+     */
     seed?: number;
     /**
      * OpenAI's free-text end-user identifier. Accepted and logged at debug
@@ -1053,6 +1154,8 @@ interface ChatCompletion {
     system_fingerprint: string | null;
     choices: ChatChoice[];
     usage: ChatUsage;
+    /** Request-scoped usage and settled debit, when supplied by the gateway. */
+    request?: RequestMetadata;
 }
 /** Incremental delta emitted on each streaming chunk. */
 interface ChatDelta {
@@ -1131,6 +1234,8 @@ interface GenerateChunk {
     request_id: string;
     seq: number;
     text_delta: string;
+    /** Per-token log probabilities aligned with `text_delta`. */
+    logprobs?: Array<Record<string, unknown>>;
     done: boolean;
     finish_reason?: "stop" | "length" | "cancelled" | "error";
     usage?: ChatUsage;
@@ -1527,10 +1632,10 @@ declare class SIEClient {
      * chunk shape documented in
      * `packages/sie_gateway/src/handlers/sse.rs::build_generate_chunk_event`.
      *
-     * The first delta carries `seq: 0` and `text_delta` populated; the
-     * terminal chunk has `done: true`, `finish_reason`, and (typically)
-     * `usage` + `ttft_ms`. The generator completes on the `data: [DONE]`
-     * sentinel.
+     * Delta chunks may carry text, log probabilities, or both; callers must not
+     * assume every non-terminal event has a non-empty `text_delta`. The terminal
+     * chunk has `done: true`, `finish_reason`, and (typically) `usage` +
+     * `ttft_ms`. The generator completes on the `data: [DONE]` sentinel.
      *
      * Error semantics match {@link streamChatCompletions}: pre-stream HTTP
      * errors throw normally, mid-stream `error` chunks throw
@@ -1548,7 +1653,7 @@ declare class SIEClient {
      * }
      * ```
      */
-    streamGenerate(model: string, prompt: string, options: GenerateOptions, signal?: AbortSignal): AsyncGenerator<GenerateChunk, void, undefined>;
+    streamGenerate(model: string, prompt: string, options: StreamGenerateOptions, signal?: AbortSignal): AsyncGenerator<GenerateChunk, void, undefined>;
     /**
      * Shared SSE consumption helper for the streaming methods.
      *
@@ -1586,7 +1691,7 @@ declare class SIEClient {
      * @param options - Extract options with labels
      * @returns Extract result with entities
      */
-    extract(model: string, item: Item, options: ExtractOptions): Promise<ExtractResult>;
+    extract(model: string, item: ExtractItem, options: ExtractOptions): Promise<ExtractResult>;
     /**
      * Extract entities from multiple items.
      *
@@ -1595,7 +1700,7 @@ declare class SIEClient {
      * @param options - Extract options with labels
      * @returns Array of extract results in same order as input
      */
-    extract(model: string, items: Item[], options: ExtractOptions): Promise<ExtractResult[]>;
+    extract(model: string, items: ExtractItem[], options: ExtractOptions): Promise<ExtractResult[]>;
     /**
      * Close the client and cleanup resources.
      *
@@ -1782,7 +1887,7 @@ declare class SIEClient {
     private detectEndpointType;
 }
 
-declare const SDK_VERSION = "0.6.20";
+declare const SDK_VERSION = "0.6.21";
 
 /**
  * Helpers for converting SIE encode results to plain JavaScript types.
@@ -2127,4 +2232,4 @@ declare function packMessage(data: unknown): Uint8Array;
  */
 declare function unpackMessage<T = unknown>(data: Uint8Array): T;
 
-export { 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 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, ServerError, type ServerInfo, type SparseResult, type SparseVector, type StatusMessage, 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 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 };

package/dist/index.d.ts

@@ -256,6 +256,15 @@ interface DocumentInput {
     /** Document format hint: "pdf", "docx", "html", etc. */
     format?: string;
 }
+/** Encoded audio bytes and optional decoder metadata. */
+interface AudioInput {
+    /** Encoded audio bytes (WAV, MP3, FLAC, etc.) */
+    data: Uint8Array;
+    /** Audio container/codec format hint */
+    format?: string;
+    /** Source sample rate in Hz, when known */
+    sampleRate?: number;
+}
 /**
  * A single item to encode, score, or extract from.
  *
@@ -291,6 +300,11 @@ interface Item {
     /** Arbitrary metadata (passed through to results) */
     metadata?: Record<string, unknown>;
 }
+/** An item accepted by extract(), including optional encoded audio. */
+interface ExtractItem extends Item {
+    /** Encoded audio for speech extraction, optionally with decoder metadata. */
+    audio?: AudioInput | Uint8Array;
+}
 /**
  * Sparse vector result with non-zero indices and values.
  * Used by SPLADE-type models.
@@ -310,6 +324,23 @@ interface TimingInfo {
     tokenizationMs?: number;
     inferenceMs?: number;
 }
+/** Authoritative units settled for one completed HTTP request. */
+interface RequestUsage {
+    inputTokens?: number;
+    pairs?: number;
+    images?: number;
+    pages?: number;
+    outputTokens?: number;
+    audioMs?: number;
+}
+/** Optional gateway metadata from the successful terminal response. */
+interface RequestMetadata {
+    id?: string;
+    /** Worker-origin immutable release/runtime identity digest. */
+    executionIdentitySha256?: string;
+    usage?: RequestUsage;
+    creditsDebited?: number;
+}
 /**
  * Result of encoding a single item.
  *
@@ -327,6 +358,8 @@ interface EncodeResult {
     multivector?: Float32Array[];
     /** Server-side timing breakdown */
     timing?: TimingInfo;
+    /** Request-scoped usage and settled debit, when supplied by the gateway. */
+    request?: RequestMetadata;
 }
 /**
  * Model dimension information.
@@ -397,6 +430,12 @@ interface ScoreEntry {
     /** Position in sorted order (0 = most relevant) */
     rank: number;
 }
+interface ScoreUsage {
+    /** Post-truncation input tokens processed */
+    inputTokens: number;
+    /** Images processed across query-document pairs */
+    images?: number;
+}
 /**
  * Result of scoring items against a query.
  */
@@ -407,6 +446,10 @@ interface ScoreResult {
     queryId?: string;
     /** Score entries, sorted by relevance (descending) */
     scores: ScoreEntry[];
+    /** Authoritative usage when emitted by the score adapter */
+    usage?: ScoreUsage;
+    /** Request-scoped usage and settled debit, when supplied by the gateway. */
+    request?: RequestMetadata;
 }
 /**
  * A single extracted entity (NER span).
@@ -458,6 +501,13 @@ interface DetectedObject {
     /** Bounding box [x, y, width, height] */
     bbox: number[];
 }
+/** Stable per-item extraction failure. */
+interface ExtractItemError {
+    /** Stable extraction error code */
+    code: string;
+    /** Sanitized extraction error message */
+    message: string;
+}
 /**
  * Result of extraction for a single item.
  */
@@ -472,6 +522,12 @@ interface ExtractResult {
     classifications: Classification[];
     /** List of detected objects */
     objects: DetectedObject[];
+    /** Additional structured extraction data */
+    data?: Record<string, unknown>;
+    /** Stable per-item failure when extraction did not complete */
+    error?: ExtractItemError;
+    /** Request-scoped usage and settled debit, when supplied by the gateway. */
+    request?: RequestMetadata;
 }
 /**
  * Information about a worker in the cluster.
@@ -524,8 +580,9 @@ interface CreatePoolOptions {
     /** Optional bundle filter. When set, only workers running this bundle will be assigned to the pool. */
     bundle?: string;
     /**
-     * Per-pool warm floor (minimum machines kept warm). The gateway publishes
-     * it as `sie_gateway_pool_warm_floor` for KEDA. Defaults to 0 (scale to zero).
+     * Per-pool warm floor (minimum machines kept warm). The gateway emits
+     * canonical `sie.gateway.pool.warm_floor` telemetry; the collector exposes
+     * `sie_gateway_pool_warm_floor` to KEDA. Defaults to 0 (scale to zero).
      */
     minimumWorkerCount?: number;
     /**
@@ -650,6 +707,10 @@ interface ModelStatus {
     queue_depth: number;
     queue_pending_items: number;
 }
+/**
+ * Worker status snapshot. Request-rate and latency telemetry is exported
+ * through the configured OTel destination instead of this message.
+ */
 interface WorkerStatusMessage {
     timestamp: number;
     name: string;
@@ -661,8 +722,6 @@ interface WorkerStatusMessage {
     server: ServerInfo;
     gpus: GPUMetrics[];
     models: ModelStatus[];
-    counters: Record<string, Record<string, number>>;
-    histograms: Record<string, Record<string, Record<string, unknown>>>;
 }
 interface ClusterStatusMessage {
     timestamp: number;
@@ -830,11 +889,45 @@ interface GenerateOptions {
     topP?: number;
     /** Optional list of stop strings. */
     stop?: string[];
+    /** OpenAI-compatible frequency penalty in [-2, 2]. */
+    frequencyPenalty?: number;
+    /** OpenAI-compatible presence penalty in [-2, 2]. */
+    presencePenalty?: number;
+    /** Native structured-output grammar. */
+    grammar?: 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.
+     * Reproducibility depends on the active backend and deployment configuration.
+     */
+    seed?: number;
+    /** Token-id-to-bias map. */
+    logitBias?: Record<string, number>;
+    /** Stable request routing key. */
+    routingKey?: string;
+    /** Prompt cache affinity key. */
+    promptCacheKey?: string;
+    /** Opaque privacy-preserving safety identifier. */
+    safetyIdentifier?: string;
+    /** Served LoRA adapter name. */
+    loraAdapter?: string;
+    /**
+     * Governed generation runtime options forwarded as `options`. Typed native
+     * fields still win when both surfaces provide the same sampler control.
+     */
+    adapterOptions?: Record<string, unknown>;
     /** GPU type / pool spec, e.g. ``"l4"`` or ``"eval-bench/l4"``. */
     gpu?: string;
     /** Auto-retry under provisioning. */
     waitForCapacity?: boolean;
 }
+/** Options for streaming native generation. */
+interface StreamGenerateOptions extends GenerateOptions {
+    /** Include per-token log probabilities in streamed chunks. */
+    logprobs?: boolean;
+    /** Number of alternate token log probabilities to return, in [0, 20]. */
+    topLogprobs?: number;
+}
 /** Aggregated generation result. */
 interface GenerateResult {
     /** Model id the gateway dispatched to. */
@@ -851,6 +944,8 @@ interface GenerateResult {
     ttftMs?: number;
     /** Average time per output token in milliseconds. */
     tpotMs?: number;
+    /** Request-scoped usage and settled debit, when supplied by the gateway. */
+    request?: RequestMetadata;
 }
 /**
  * A single message in a chat completion request.
@@ -1009,6 +1104,12 @@ interface ChatCompletionRequest {
      * `[-100, 100]` and caps map size.
      */
     logit_bias?: Record<string, number>;
+    /**
+     * Optional per-request sampling seed. Must be a JavaScript safe integer
+     * (`-(2^53 - 1)` through `2^53 - 1`) so the caller's integer is represented
+     * exactly. The SDK throws `RangeError` before network I/O for invalid values.
+     * Reproducibility depends on the active backend and deployment configuration.
+     */
     seed?: number;
     /**
      * OpenAI's free-text end-user identifier. Accepted and logged at debug
@@ -1053,6 +1154,8 @@ interface ChatCompletion {
     system_fingerprint: string | null;
     choices: ChatChoice[];
     usage: ChatUsage;
+    /** Request-scoped usage and settled debit, when supplied by the gateway. */
+    request?: RequestMetadata;
 }
 /** Incremental delta emitted on each streaming chunk. */
 interface ChatDelta {
@@ -1131,6 +1234,8 @@ interface GenerateChunk {
     request_id: string;
     seq: number;
     text_delta: string;
+    /** Per-token log probabilities aligned with `text_delta`. */
+    logprobs?: Array<Record<string, unknown>>;
     done: boolean;
     finish_reason?: "stop" | "length" | "cancelled" | "error";
     usage?: ChatUsage;
@@ -1527,10 +1632,10 @@ declare class SIEClient {
      * chunk shape documented in
      * `packages/sie_gateway/src/handlers/sse.rs::build_generate_chunk_event`.
      *
-     * The first delta carries `seq: 0` and `text_delta` populated; the
-     * terminal chunk has `done: true`, `finish_reason`, and (typically)
-     * `usage` + `ttft_ms`. The generator completes on the `data: [DONE]`
-     * sentinel.
+     * Delta chunks may carry text, log probabilities, or both; callers must not
+     * assume every non-terminal event has a non-empty `text_delta`. The terminal
+     * chunk has `done: true`, `finish_reason`, and (typically) `usage` +
+     * `ttft_ms`. The generator completes on the `data: [DONE]` sentinel.
      *
      * Error semantics match {@link streamChatCompletions}: pre-stream HTTP
      * errors throw normally, mid-stream `error` chunks throw
@@ -1548,7 +1653,7 @@ declare class SIEClient {
      * }
      * ```
      */
-    streamGenerate(model: string, prompt: string, options: GenerateOptions, signal?: AbortSignal): AsyncGenerator<GenerateChunk, void, undefined>;
+    streamGenerate(model: string, prompt: string, options: StreamGenerateOptions, signal?: AbortSignal): AsyncGenerator<GenerateChunk, void, undefined>;
     /**
      * Shared SSE consumption helper for the streaming methods.
      *
@@ -1586,7 +1691,7 @@ declare class SIEClient {
      * @param options - Extract options with labels
      * @returns Extract result with entities
      */
-    extract(model: string, item: Item, options: ExtractOptions): Promise<ExtractResult>;
+    extract(model: string, item: ExtractItem, options: ExtractOptions): Promise<ExtractResult>;
     /**
      * Extract entities from multiple items.
      *
@@ -1595,7 +1700,7 @@ declare class SIEClient {
      * @param options - Extract options with labels
      * @returns Array of extract results in same order as input
      */
-    extract(model: string, items: Item[], options: ExtractOptions): Promise<ExtractResult[]>;
+    extract(model: string, items: ExtractItem[], options: ExtractOptions): Promise<ExtractResult[]>;
     /**
      * Close the client and cleanup resources.
      *
@@ -1782,7 +1887,7 @@ declare class SIEClient {
     private detectEndpointType;
 }
 
-declare const SDK_VERSION = "0.6.20";
+declare const SDK_VERSION = "0.6.21";
 
 /**
  * Helpers for converting SIE encode results to plain JavaScript types.
@@ -2127,4 +2232,4 @@ declare function packMessage(data: unknown): Uint8Array;
  */
 declare function unpackMessage<T = unknown>(data: Uint8Array): T;
 
-export { 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 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, ServerError, type ServerInfo, type SparseResult, type SparseVector, type StatusMessage, 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 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 };