@qvac/sdk 0.13.4 → 0.14.0

package/dist/schemas/completion-stream.d.ts

@@ -32,7 +32,8 @@ export declare const generationParamsSchema: z.ZodObject<{
     frequency_penalty: z.ZodOptional<z.ZodNumber>;
     presence_penalty: z.ZodOptional<z.ZodNumber>;
     repeat_penalty: z.ZodOptional<z.ZodNumber>;
-    reasoning_budget: z.ZodOptional<z.ZodUnion<readonly [z.ZodLiteral<-1>, z.ZodLiteral<0>]>>;
+    reasoning_budget: z.ZodOptional<z.ZodNumber>;
+    remove_thinking_from_context: z.ZodOptional<z.ZodBoolean>;
 }, z.core.$strict>;
 export declare const responseFormatSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
     type: z.ZodLiteral<"text">;
@@ -99,7 +100,8 @@ export declare const completionClientParamsSchema: z.ZodObject<{
         frequency_penalty: z.ZodOptional<z.ZodNumber>;
         presence_penalty: z.ZodOptional<z.ZodNumber>;
         repeat_penalty: z.ZodOptional<z.ZodNumber>;
-        reasoning_budget: z.ZodOptional<z.ZodUnion<readonly [z.ZodLiteral<-1>, z.ZodLiteral<0>]>>;
+        reasoning_budget: z.ZodOptional<z.ZodNumber>;
+        remove_thinking_from_context: z.ZodOptional<z.ZodBoolean>;
     }, z.core.$strict>>;
     captureThinking: z.ZodOptional<z.ZodBoolean>;
     emitRawDeltas: z.ZodOptional<z.ZodBoolean>;
@@ -167,7 +169,8 @@ export declare const completionStreamRequestSchema: z.ZodObject<{
         frequency_penalty: z.ZodOptional<z.ZodNumber>;
         presence_penalty: z.ZodOptional<z.ZodNumber>;
         repeat_penalty: z.ZodOptional<z.ZodNumber>;
-        reasoning_budget: z.ZodOptional<z.ZodUnion<readonly [z.ZodLiteral<-1>, z.ZodLiteral<0>]>>;
+        reasoning_budget: z.ZodOptional<z.ZodNumber>;
+        remove_thinking_from_context: z.ZodOptional<z.ZodBoolean>;
     }, z.core.$strict>>;
     captureThinking: z.ZodOptional<z.ZodBoolean>;
     emitRawDeltas: z.ZodOptional<z.ZodBoolean>;

package/dist/schemas/completion-stream.js

@@ -1,6 +1,7 @@
 import { z } from "zod";
 import { toolSchema } from "./tools.js";
 import { completionEventSchema } from "./completion-event.js";
+import { REASONING_BUDGET_MAX } from "./llamacpp-config.js";
 export { completionStatsSchema, } from "./completion-event.js";
 /**
  * Tool-call output dialect. Auto-detected from the model name; pass via
@@ -63,9 +64,16 @@ export const generationParamsSchema = z
         .optional()
         .describe("Penalty applied to repeated tokens."),
     reasoning_budget: z
-        .union([z.literal(-1), z.literal(0)])
+        .number()
+        .int()
+        .min(-1)
+        .max(REASONING_BUDGET_MAX)
+        .optional()
+        .describe("Per-request reasoning channel budget. `-1` keeps the model's reasoning channel on; `0` disables it for this request; any positive integer caps the reasoning channel at that many tokens. Equivalent to the load-time `reasoning_budget` config but scoped to a single `run()` call; the prior value is restored afterwards."),
+    remove_thinking_from_context: z
+        .boolean()
         .optional()
-        .describe("Per-request reasoning channel budget. `-1` keeps the model's reasoning channel on; `0` disables it for this request. Equivalent to the load-time `reasoning_budget` config but scoped to a single `run()` call; the prior value is restored afterwards."),
+        .describe("When the model emits a reasoning block during generation (e.g. `<think>...</think>` for the Qwen3 family, `<|channel>thought ... <channel|>` for Gemma 4), drop those tokens from the KV cache at end-of-generation so subsequent turns do not accumulate reasoning history. Defaults to `false`. No-op for models without a recognised reasoning channel. Throws on models with recurrent memory (SSM / hybrid SSM such as Qwen3.5), where the cache edit is unsupported."),
 })
     .strict();
 const jsonSchemaObjectSchema = z.record(z.string(), z.unknown());

package/dist/schemas/download-asset.d.ts

@@ -26,7 +26,7 @@ export declare const downloadAssetOptionsSchema: z.ZodPipe<z.ZodObject<{
             "onnx-tts": "onnx-tts";
             "tts-ggml": "tts-ggml";
             "parakeet-transcription": "parakeet-transcription";
-            "onnx-ocr": "onnx-ocr";
+            "ggml-ocr": "ggml-ocr";
             "sdcpp-generation": "sdcpp-generation";
             "ggml-vla": "ggml-vla";
             "ggml-classification": "ggml-classification";
@@ -52,7 +52,7 @@ export declare const downloadAssetOptionsSchema: z.ZodPipe<z.ZodObject<{
         engine?: string | undefined;
         expectedSize?: number | undefined;
         sha256Checksum?: string | undefined;
-        addon?: "embeddings" | "diffusion" | "tts" | "ocr" | "vla" | "llamacpp-completion" | "whispercpp-transcription" | "bci-whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "tts-ggml" | "parakeet-transcription" | "onnx-ocr" | "sdcpp-generation" | "ggml-vla" | "ggml-classification" | "llm" | "whisper" | "bci" | "nmt" | "parakeet" | "classification" | "vad" | undefined;
+        addon?: "embeddings" | "diffusion" | "tts" | "ocr" | "vla" | "llamacpp-completion" | "whispercpp-transcription" | "bci-whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "tts-ggml" | "parakeet-transcription" | "ggml-ocr" | "sdcpp-generation" | "ggml-vla" | "ggml-classification" | "llm" | "whisper" | "bci" | "nmt" | "parakeet" | "classification" | "vad" | undefined;
     };
 }, {
     assetSrc: string | {
@@ -66,7 +66,7 @@ export declare const downloadAssetOptionsSchema: z.ZodPipe<z.ZodObject<{
         engine?: string | undefined;
         expectedSize?: number | undefined;
         sha256Checksum?: string | undefined;
-        addon?: "embeddings" | "diffusion" | "tts" | "ocr" | "vla" | "llamacpp-completion" | "whispercpp-transcription" | "bci-whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "tts-ggml" | "parakeet-transcription" | "onnx-ocr" | "sdcpp-generation" | "ggml-vla" | "ggml-classification" | "llm" | "whisper" | "bci" | "nmt" | "parakeet" | "classification" | "vad" | undefined;
+        addon?: "embeddings" | "diffusion" | "tts" | "ocr" | "vla" | "llamacpp-completion" | "whispercpp-transcription" | "bci-whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "tts-ggml" | "parakeet-transcription" | "ggml-ocr" | "sdcpp-generation" | "ggml-vla" | "ggml-classification" | "llm" | "whisper" | "bci" | "nmt" | "parakeet" | "classification" | "vad" | undefined;
     };
     seed?: boolean | undefined;
 }>>;
@@ -96,7 +96,7 @@ export declare const downloadAssetOptionsToRequestSchema: z.ZodPipe<z.ZodObject<
             "onnx-tts": "onnx-tts";
             "tts-ggml": "tts-ggml";
             "parakeet-transcription": "parakeet-transcription";
-            "onnx-ocr": "onnx-ocr";
+            "ggml-ocr": "ggml-ocr";
             "sdcpp-generation": "sdcpp-generation";
             "ggml-vla": "ggml-vla";
             "ggml-classification": "ggml-classification";
@@ -130,7 +130,7 @@ export declare const downloadAssetOptionsToRequestSchema: z.ZodPipe<z.ZodObject<
         engine?: string | undefined;
         expectedSize?: number | undefined;
         sha256Checksum?: string | undefined;
-        addon?: "embeddings" | "diffusion" | "tts" | "ocr" | "vla" | "llamacpp-completion" | "whispercpp-transcription" | "bci-whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "tts-ggml" | "parakeet-transcription" | "onnx-ocr" | "sdcpp-generation" | "ggml-vla" | "ggml-classification" | "llm" | "whisper" | "bci" | "nmt" | "parakeet" | "classification" | "vad" | undefined;
+        addon?: "embeddings" | "diffusion" | "tts" | "ocr" | "vla" | "llamacpp-completion" | "whispercpp-transcription" | "bci-whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "tts-ggml" | "parakeet-transcription" | "ggml-ocr" | "sdcpp-generation" | "ggml-vla" | "ggml-classification" | "llm" | "whisper" | "bci" | "nmt" | "parakeet" | "classification" | "vad" | undefined;
     };
     seed?: boolean | undefined;
     onProgress?: unknown;

package/dist/schemas/engine-addon-map.d.ts

@@ -7,7 +7,7 @@ export declare const ENGINE_TO_ADDON: {
     readonly "nmtcpp-translation": "nmt";
     readonly "onnx-tts": "tts";
     readonly "tts-ggml": "tts";
-    readonly "onnx-ocr": "ocr";
+    readonly "ggml-ocr": "ocr";
     readonly "parakeet-transcription": "parakeet";
     readonly "sdcpp-generation": "diffusion";
     readonly "ggml-vla": "vla";

package/dist/schemas/engine-addon-map.js

@@ -11,7 +11,7 @@ export const ENGINE_TO_ADDON = {
     [ModelType.nmtcppTranslation]: "nmt",
     [ModelType.onnxTts]: "tts",
     [ModelType.ttsGgml]: "tts",
-    [ModelType.onnxOcr]: "ocr",
+    [ModelType.ggmlOcr]: "ocr",
     [ModelType.parakeetTranscription]: "parakeet",
     [ModelType.sdcppGeneration]: "diffusion",
     [ModelType.ggmlVla]: "vla",
@@ -27,7 +27,10 @@ const LEGACY_ENGINE_TO_CANONICAL = {
     [ADDON_EMBEDDING]: ModelType.llamacppEmbedding,
     [ADDON_NMT]: ModelType.nmtcppTranslation,
     [ADDON_TTS]: ModelType.ttsGgml,
-    [ADDON_OCR]: ModelType.onnxOcr,
+    [ADDON_OCR]: ModelType.ggmlOcr,
+    // Pre-GGML package / tag names from the ONNX era — resolve to the GGML engine.
+    "@qvac/ocr-onnx": ModelType.ggmlOcr,
+    "onnx-ocr": ModelType.ggmlOcr,
     [ADDON_PARAKEET]: ModelType.parakeetTranscription,
     "@qvac/translation-llamacpp": ModelType.nmtcppTranslation,
     "@qvac/vad-silero": "onnx-vad",
@@ -42,7 +45,7 @@ const LEGACY_ENGINE_TO_CANONICAL = {
     translation: ModelType.nmtcppTranslation,
     vad: "onnx-vad",
     tts: ModelType.ttsGgml,
-    ocr: ModelType.onnxOcr,
+    ocr: ModelType.ggmlOcr,
     [ADDON_DIFFUSION]: ModelType.sdcppGeneration,
     diffusion: ModelType.sdcppGeneration,
     [ADDON_VLA]: ModelType.ggmlVla,

package/dist/schemas/error.js

@@ -1,5 +1,6 @@
 import { z } from "zod";
 import { QvacErrorBase } from "@qvac/error";
+import { formatZodError } from "../utils/zod-error.js";
 /**
  * Wire shape for errors thrown across the RPC boundary. The fields are
  * the union of (a) the legacy `QvacErrorBase` serialisation (`name`,
@@ -52,7 +53,11 @@ export function createErrorResponse(error) {
         }
         return response;
     }
-    const message = error instanceof Error ? error.message : String(error);
+    const message = error instanceof z.ZodError
+        ? formatZodError(error)
+        : error instanceof Error
+            ? error.message
+            : String(error);
     const stack = error instanceof Error ? error.stack : undefined;
     return {
         type: "error",

package/dist/schemas/llamacpp-config.d.ts

@@ -1,4 +1,10 @@
 import { z } from "zod";
+/**
+ * Upper bound for `reasoning_budget`. Mirrors the llm-llamacpp addon, which
+ * stores the budget as a 32-bit `int` and rejects values above
+ * `std::numeric_limits<int>::max()`.
+ */
+export declare const REASONING_BUDGET_MAX = 2147483647;
 export declare const VERBOSITY: {
     readonly ERROR: 0;
     readonly WARN: 1;
@@ -41,7 +47,7 @@ export declare const llmConfigBaseSchema: z.ZodObject<{
     }>>;
     "tensor-split": z.ZodOptional<z.ZodString>;
     openclCacheDir: z.ZodOptional<z.ZodString>;
-    reasoning_budget: z.ZodOptional<z.ZodUnion<readonly [z.ZodLiteral<-1>, z.ZodLiteral<0>]>>;
+    reasoning_budget: z.ZodOptional<z.ZodNumber>;
     projectionModelSrc: z.ZodOptional<z.ZodUnion<readonly [z.ZodString, z.ZodObject<{
         src: z.ZodString;
         name: z.ZodOptional<z.ZodString>;
@@ -67,7 +73,7 @@ export declare const llmConfigBaseSchema: z.ZodObject<{
             "onnx-tts": "onnx-tts";
             "tts-ggml": "tts-ggml";
             "parakeet-transcription": "parakeet-transcription";
-            "onnx-ocr": "onnx-ocr";
+            "ggml-ocr": "ggml-ocr";
             "sdcpp-generation": "sdcpp-generation";
             "ggml-vla": "ggml-vla";
             "ggml-classification": "ggml-classification";
@@ -79,6 +85,11 @@ export declare const llmConfigBaseSchema: z.ZodObject<{
             classification: "classification";
         }>, z.ZodLiteral<"vad">]>>;
     }, z.core.$strip>]>>;
+    image_tile_mode: z.ZodOptional<z.ZodEnum<{
+        disabled: "disabled";
+        batched: "batched";
+        sequential: "sequential";
+    }>>;
 }, z.core.$strip>;
 export type LlmConfigInput = z.infer<typeof llmConfigBaseSchema>;
 export declare const LLM_CONFIG_DEFAULTS: {
@@ -86,6 +97,7 @@ export declare const LLM_CONFIG_DEFAULTS: {
     readonly gpu_layers: 99;
     readonly device: "gpu";
     readonly system_prompt: "You are a helpful assistant.";
+    readonly image_tile_mode: "sequential";
 };
 export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
     ctx_size: z.ZodOptional<z.ZodNumber>;
@@ -123,7 +135,7 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
     }>>;
     "tensor-split": z.ZodOptional<z.ZodString>;
     openclCacheDir: z.ZodOptional<z.ZodString>;
-    reasoning_budget: z.ZodOptional<z.ZodUnion<readonly [z.ZodLiteral<-1>, z.ZodLiteral<0>]>>;
+    reasoning_budget: z.ZodOptional<z.ZodNumber>;
     projectionModelSrc: z.ZodOptional<z.ZodUnion<readonly [z.ZodString, z.ZodObject<{
         src: z.ZodString;
         name: z.ZodOptional<z.ZodString>;
@@ -149,7 +161,7 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
             "onnx-tts": "onnx-tts";
             "tts-ggml": "tts-ggml";
             "parakeet-transcription": "parakeet-transcription";
-            "onnx-ocr": "onnx-ocr";
+            "ggml-ocr": "ggml-ocr";
             "sdcpp-generation": "sdcpp-generation";
             "ggml-vla": "ggml-vla";
             "ggml-classification": "ggml-classification";
@@ -161,6 +173,11 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
             classification: "classification";
         }>, z.ZodLiteral<"vad">]>>;
     }, z.core.$strip>]>>;
+    image_tile_mode: z.ZodOptional<z.ZodEnum<{
+        disabled: "disabled";
+        batched: "batched";
+        sequential: "sequential";
+    }>>;
 }, z.core.$strip>, z.ZodTransform<{
     ctx_size: number | undefined;
     temp?: number | undefined;
@@ -187,7 +204,7 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
     "split-mode"?: "none" | "layer" | "row" | undefined;
     "tensor-split"?: string | undefined;
     openclCacheDir?: string | undefined;
-    reasoning_budget?: 0 | -1 | undefined;
+    reasoning_budget?: number | undefined;
     projectionModelSrc?: string | {
         src: string;
         name?: string | undefined;
@@ -199,8 +216,9 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
         engine?: string | undefined;
         expectedSize?: number | undefined;
         sha256Checksum?: string | undefined;
-        addon?: "embeddings" | "diffusion" | "tts" | "ocr" | "vla" | "llamacpp-completion" | "whispercpp-transcription" | "bci-whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "tts-ggml" | "parakeet-transcription" | "onnx-ocr" | "sdcpp-generation" | "ggml-vla" | "ggml-classification" | "llm" | "whisper" | "bci" | "nmt" | "parakeet" | "classification" | "vad" | undefined;
+        addon?: "embeddings" | "diffusion" | "tts" | "ocr" | "vla" | "llamacpp-completion" | "whispercpp-transcription" | "bci-whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "tts-ggml" | "parakeet-transcription" | "ggml-ocr" | "sdcpp-generation" | "ggml-vla" | "ggml-classification" | "llm" | "whisper" | "bci" | "nmt" | "parakeet" | "classification" | "vad" | undefined;
     } | undefined;
+    image_tile_mode: "disabled" | "batched" | "sequential" | undefined;
 }, {
     ctx_size?: number | undefined;
     temp?: number | undefined;
@@ -227,7 +245,7 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
     "split-mode"?: "none" | "layer" | "row" | undefined;
     "tensor-split"?: string | undefined;
     openclCacheDir?: string | undefined;
-    reasoning_budget?: 0 | -1 | undefined;
+    reasoning_budget?: number | undefined;
     projectionModelSrc?: string | {
         src: string;
         name?: string | undefined;
@@ -239,8 +257,9 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
         engine?: string | undefined;
         expectedSize?: number | undefined;
         sha256Checksum?: string | undefined;
-        addon?: "embeddings" | "diffusion" | "tts" | "ocr" | "vla" | "llamacpp-completion" | "whispercpp-transcription" | "bci-whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "tts-ggml" | "parakeet-transcription" | "onnx-ocr" | "sdcpp-generation" | "ggml-vla" | "ggml-classification" | "llm" | "whisper" | "bci" | "nmt" | "parakeet" | "classification" | "vad" | undefined;
+        addon?: "embeddings" | "diffusion" | "tts" | "ocr" | "vla" | "llamacpp-completion" | "whispercpp-transcription" | "bci-whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "tts-ggml" | "parakeet-transcription" | "ggml-ocr" | "sdcpp-generation" | "ggml-vla" | "ggml-classification" | "llm" | "whisper" | "bci" | "nmt" | "parakeet" | "classification" | "vad" | undefined;
     } | undefined;
+    image_tile_mode?: "disabled" | "batched" | "sequential" | undefined;
 }>>;
 export type LlmConfig = z.infer<typeof llmConfigSchema>;
 export declare const embedConfigBaseSchema: z.ZodObject<{

package/dist/schemas/llamacpp-config.js

@@ -1,6 +1,12 @@
 import { z } from "zod";
 import { modelSrcInputSchema } from "./model-src-utils.js";
 import { TOOLS_MODE } from "./tools.js";
+/**
+ * Upper bound for `reasoning_budget`. Mirrors the llm-llamacpp addon, which
+ * stores the budget as a 32-bit `int` and rejects values above
+ * `std::numeric_limits<int>::max()`.
+ */
+export const REASONING_BUDGET_MAX = 2147483647;
 export const VERBOSITY = {
     ERROR: 0,
     WARN: 1,
@@ -57,10 +63,20 @@ export const llmConfigBaseSchema = z.object({
      */
     openclCacheDir: z.string().optional(),
     /**
-     * Reasoning channel token budget. `-1` = unrestricted, `0` = disabled.
+     * Reasoning channel token budget. `-1` = unrestricted, `0` = disabled, any
+     * positive integer caps the reasoning channel at that many tokens (the
+     * sampler force-emits the closing think tag once the budget is exhausted).
      */
-    reasoning_budget: z.union([z.literal(-1), z.literal(0)]).optional(),
+    reasoning_budget: z.number().int().min(-1).max(REASONING_BUDGET_MAX).optional(),
     projectionModelSrc: modelSrcInputSchema.optional(),
+    /**
+     * Qwen3.5-VL multi-tile image encoding mode (multimodal models only):
+     *   - `"sequential"` (default): encode image tiles one tile at a time.
+     *   - `"batched"`: encode all tiles in a single batched pass.
+     *   - `"disabled"`: no multi-tile encoding (single tile).
+     * Ignored by text-only models. Default is `"sequential"`.
+     */
+    image_tile_mode: z.enum(["disabled", "batched", "sequential"]).optional(),
 });
 // Default values - typed as partial of the config
 export const LLM_CONFIG_DEFAULTS = {
@@ -68,6 +84,7 @@ export const LLM_CONFIG_DEFAULTS = {
     gpu_layers: 99,
     device: "gpu",
     system_prompt: "You are a helpful assistant.",
+    image_tile_mode: "sequential",
 };
 // Full schema - applies defaults via transform (no duplication)
 export const llmConfigSchema = llmConfigBaseSchema.transform((data) => ({