@qvac/sdk 0.7.1 → 0.8.0

package/dist/schemas/common.js

@@ -1,4 +1,5 @@
 import { z } from "zod";
+import { perCallProfilingSchema } from "./profiling.js";
 import { pingRequestSchema, pingResponseSchema } from "./ping.js";
 import { completionStreamRequestSchema, completionStreamResponseSchema, } from "./completion-stream.js";
 import { loadModelRequestSchema, loadModelResponseSchema, modelProgressUpdateSchema, } from "./load-model.js";
@@ -73,4 +74,5 @@ export const responseSchema = z.discriminatedUnion("type", [
 export const rpcOptionsSchema = z.object({
     timeout: z.number().min(100).optional(),
     forceNewConnection: z.boolean().optional(),
+    profiling: perCallProfilingSchema.optional(),
 });

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

@@ -2,6 +2,16 @@ import { z } from "zod";
 export declare const attachmentSchema: z.ZodObject<{
     path: z.ZodString;
 }, z.core.$strip>;
+export declare const generationParamsSchema: z.ZodObject<{
+    temp: z.ZodOptional<z.ZodNumber>;
+    top_p: z.ZodOptional<z.ZodNumber>;
+    top_k: z.ZodOptional<z.ZodNumber>;
+    predict: z.ZodOptional<z.ZodNumber>;
+    seed: z.ZodOptional<z.ZodNumber>;
+    frequency_penalty: z.ZodOptional<z.ZodNumber>;
+    presence_penalty: z.ZodOptional<z.ZodNumber>;
+    repeat_penalty: z.ZodOptional<z.ZodNumber>;
+}, z.core.$strict>;
 export declare const completionParamsSchema: z.ZodObject<{
     history: z.ZodArray<z.ZodObject<{
         role: z.ZodString;
@@ -45,6 +55,16 @@ export declare const completionClientParamsSchema: z.ZodObject<{
     }, z.core.$strip>>>;
     stream: z.ZodBoolean;
     kvCache: z.ZodOptional<z.ZodUnion<readonly [z.ZodBoolean, z.ZodString]>>;
+    generationParams: z.ZodOptional<z.ZodObject<{
+        temp: z.ZodOptional<z.ZodNumber>;
+        top_p: z.ZodOptional<z.ZodNumber>;
+        top_k: z.ZodOptional<z.ZodNumber>;
+        predict: z.ZodOptional<z.ZodNumber>;
+        seed: z.ZodOptional<z.ZodNumber>;
+        frequency_penalty: z.ZodOptional<z.ZodNumber>;
+        presence_penalty: z.ZodOptional<z.ZodNumber>;
+        repeat_penalty: z.ZodOptional<z.ZodNumber>;
+    }, z.core.$strict>>;
 }, z.core.$strip>;
 export declare const completionStreamRequestSchema: z.ZodObject<{
     history: z.ZodArray<z.ZodObject<{
@@ -78,6 +98,16 @@ export declare const completionStreamRequestSchema: z.ZodObject<{
     }, z.core.$strip>>>;
     stream: z.ZodBoolean;
     kvCache: z.ZodOptional<z.ZodUnion<readonly [z.ZodBoolean, z.ZodString]>>;
+    generationParams: z.ZodOptional<z.ZodObject<{
+        temp: z.ZodOptional<z.ZodNumber>;
+        top_p: z.ZodOptional<z.ZodNumber>;
+        top_k: z.ZodOptional<z.ZodNumber>;
+        predict: z.ZodOptional<z.ZodNumber>;
+        seed: z.ZodOptional<z.ZodNumber>;
+        frequency_penalty: z.ZodOptional<z.ZodNumber>;
+        presence_penalty: z.ZodOptional<z.ZodNumber>;
+        repeat_penalty: z.ZodOptional<z.ZodNumber>;
+    }, z.core.$strict>>;
     type: z.ZodLiteral<"completionStream">;
 }, z.core.$strip>;
 export declare const completionStatsSchema: z.ZodObject<{
@@ -122,6 +152,7 @@ export declare const completionStreamResponseSchema: z.ZodObject<{
     }, z.core.$strip>>>;
     error: z.ZodOptional<z.ZodString>;
 }, z.core.$strip>;
+export type GenerationParams = z.infer<typeof generationParamsSchema>;
 export type CompletionParams = z.infer<typeof completionParamsSchema>;
 export type CompletionClientParams = z.input<typeof completionClientParamsSchema>;
 export type CompletionStreamRequest = z.infer<typeof completionStreamRequestSchema>;

package/dist/schemas/completion-stream.js

@@ -7,6 +7,18 @@ const kvCacheSchema = z.union([
     z.boolean(),
     z.string().min(1, "KV cache key cannot be empty string"),
 ]);
+export const generationParamsSchema = z
+    .object({
+    temp: z.number().optional(),
+    top_p: z.number().optional(),
+    top_k: z.number().optional(),
+    predict: z.number().optional(),
+    seed: z.number().optional(),
+    frequency_penalty: z.number().optional(),
+    presence_penalty: z.number().optional(),
+    repeat_penalty: z.number().optional(),
+})
+    .strict();
 export const completionParamsSchema = z.object({
     history: z.array(z.object({
         role: z.string(),
@@ -20,6 +32,7 @@ export const completionClientParamsSchema = completionParamsSchema.extend({
     tools: z.array(toolSchema).optional(),
     stream: z.boolean(),
     kvCache: kvCacheSchema.optional(),
+    generationParams: generationParamsSchema.optional(),
 });
 export const completionStreamRequestSchema = completionClientParamsSchema.extend({
     type: z.literal("completionStream"),

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

@@ -18,11 +18,13 @@ export declare const downloadAssetOptionsSchema: z.ZodPipe<z.ZodObject<{
             "llamacpp-embedding": "llamacpp-embedding";
             "nmtcpp-translation": "nmtcpp-translation";
             "onnx-tts": "onnx-tts";
+            "parakeet-transcription": "parakeet-transcription";
             "onnx-ocr": "onnx-ocr";
             llm: "llm";
             whisper: "whisper";
             embeddings: "embeddings";
             nmt: "nmt";
+            parakeet: "parakeet";
             tts: "tts";
             ocr: "ocr";
         }>, z.ZodLiteral<"vad">]>>;
@@ -41,7 +43,7 @@ export declare const downloadAssetOptionsSchema: z.ZodPipe<z.ZodObject<{
         engine?: string | undefined;
         expectedSize?: number | undefined;
         sha256Checksum?: string | undefined;
-        addon?: "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "onnx-ocr" | "llm" | "whisper" | "embeddings" | "nmt" | "tts" | "ocr" | "vad" | undefined;
+        addon?: "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "parakeet-transcription" | "onnx-ocr" | "llm" | "whisper" | "embeddings" | "nmt" | "parakeet" | "tts" | "ocr" | "vad" | undefined;
     };
 }, {
     assetSrc: string | {
@@ -55,7 +57,7 @@ export declare const downloadAssetOptionsSchema: z.ZodPipe<z.ZodObject<{
         engine?: string | undefined;
         expectedSize?: number | undefined;
         sha256Checksum?: string | undefined;
-        addon?: "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "onnx-ocr" | "llm" | "whisper" | "embeddings" | "nmt" | "tts" | "ocr" | "vad" | undefined;
+        addon?: "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "parakeet-transcription" | "onnx-ocr" | "llm" | "whisper" | "embeddings" | "nmt" | "parakeet" | "tts" | "ocr" | "vad" | undefined;
     };
     seed?: boolean | undefined;
 }>>;
@@ -77,11 +79,13 @@ export declare const downloadAssetOptionsToRequestSchema: z.ZodPipe<z.ZodObject<
             "llamacpp-embedding": "llamacpp-embedding";
             "nmtcpp-translation": "nmtcpp-translation";
             "onnx-tts": "onnx-tts";
+            "parakeet-transcription": "parakeet-transcription";
             "onnx-ocr": "onnx-ocr";
             llm: "llm";
             whisper: "whisper";
             embeddings: "embeddings";
             nmt: "nmt";
+            parakeet: "parakeet";
             tts: "tts";
             ocr: "ocr";
         }>, z.ZodLiteral<"vad">]>>;
@@ -106,7 +110,7 @@ export declare const downloadAssetOptionsToRequestSchema: z.ZodPipe<z.ZodObject<
         engine?: string | undefined;
         expectedSize?: number | undefined;
         sha256Checksum?: string | undefined;
-        addon?: "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "onnx-ocr" | "llm" | "whisper" | "embeddings" | "nmt" | "tts" | "ocr" | "vad" | undefined;
+        addon?: "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "parakeet-transcription" | "onnx-ocr" | "llm" | "whisper" | "embeddings" | "nmt" | "parakeet" | "tts" | "ocr" | "vad" | undefined;
     };
     seed?: boolean | undefined;
     onProgress?: unknown;
@@ -139,6 +143,12 @@ export declare const downloadTypeSchema: z.ZodEnum<{
     http: "http";
     registry: "registry";
 }>;
+export declare const sourceTypeSchema: z.ZodEnum<{
+    hyperdrive: "hyperdrive";
+    http: "http";
+    registry: "registry";
+    filesystem: "filesystem";
+}>;
 export declare const downloadMetadataSchema: z.ZodObject<{
     key: z.ZodString;
     startTime: z.ZodNumber;
@@ -174,6 +184,7 @@ export type DownloadAssetOptions = z.input<typeof downloadAssetOptionsSchema> &
 export type DownloadAssetRequest = z.infer<typeof downloadAssetRequestSchema>;
 export type DownloadAssetResponse = z.infer<typeof downloadAssetResponseSchema>;
 export type DownloadType = z.infer<typeof downloadTypeSchema>;
+export type SourceType = z.infer<typeof sourceTypeSchema>;
 export type BaseDownloadEntry = z.infer<typeof downloadMetadataSchema> & {
     promise: Promise<string>;
     abortController: AbortController;

package/dist/schemas/download-asset.js

@@ -36,7 +36,9 @@ export const downloadAssetResponseSchema = z.object({
     assetId: z.string().optional(),
     error: z.string().optional(),
 });
-export const downloadTypeSchema = z.enum(["hyperdrive", "http", "registry"]);
+const downloadTypes = ["hyperdrive", "http", "registry"];
+export const downloadTypeSchema = z.enum(downloadTypes);
+export const sourceTypeSchema = z.enum([...downloadTypes, "filesystem"]);
 export const downloadMetadataSchema = z.object({
     key: z.string(),
     startTime: z.number(),

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

@@ -1,4 +1,5 @@
 import { ModelType } from "./model-types.js";
+import { ADDON_EMBEDDING, ADDON_LLM, ADDON_NMT, ADDON_OCR, ADDON_PARAKEET, ADDON_TTS, ADDON_WHISPER, } from "./plugin.js";
 import { modelRegistryEngineSchema, } from "./registry.js";
 // Canonical engine → addon mapping (exhaustive).
 // TypeScript enforces that every ModelRegistryEngine has an entry.
@@ -9,20 +10,22 @@ export const ENGINE_TO_ADDON = {
     [ModelType.nmtcppTranslation]: "nmt",
     [ModelType.onnxTts]: "tts",
     [ModelType.onnxOcr]: "ocr",
+    [ModelType.parakeetTranscription]: "parakeet",
     "onnx-vad": "vad",
 };
 // Legacy engine names → canonical engine.
 // Used for backward compatibility with old registry data that uses @qvac/* package names.
 const LEGACY_ENGINE_TO_CANONICAL = {
-    "@qvac/llm-llamacpp": ModelType.llamacppCompletion,
-    "@qvac/transcription-whispercpp": ModelType.whispercppTranscription,
-    "@qvac/embed-llamacpp": ModelType.llamacppEmbedding,
-    "@qvac/translation-nmtcpp": ModelType.nmtcppTranslation,
+    [ADDON_LLM]: ModelType.llamacppCompletion,
+    [ADDON_WHISPER]: ModelType.whispercppTranscription,
+    [ADDON_EMBEDDING]: ModelType.llamacppEmbedding,
+    [ADDON_NMT]: ModelType.nmtcppTranslation,
+    [ADDON_TTS]: ModelType.onnxTts,
+    [ADDON_OCR]: ModelType.onnxOcr,
+    [ADDON_PARAKEET]: ModelType.parakeetTranscription,
     "@qvac/translation-llamacpp": ModelType.nmtcppTranslation,
     "@qvac/vad-silero": "onnx-vad",
-    "@qvac/tts-onnx": ModelType.onnxTts,
     "@qvac/tts": ModelType.onnxTts,
-    "@qvac/ocr-onnx": ModelType.onnxOcr,
     // Tag-style names (used by some older registry entries)
     generation: ModelType.llamacppCompletion,
     transcription: ModelType.whispercppTranscription,

package/dist/schemas/get-model-info.d.ts

@@ -39,6 +39,7 @@ export declare const modelInfoSchema: z.ZodObject<{
         whisper: "whisper";
         embeddings: "embeddings";
         nmt: "nmt";
+        parakeet: "parakeet";
         tts: "tts";
         ocr: "ocr";
         vad: "vad";
@@ -84,6 +85,7 @@ export declare const getModelInfoResponseSchema: z.ZodObject<{
             whisper: "whisper";
             embeddings: "embeddings";
             nmt: "nmt";
+            parakeet: "parakeet";
             tts: "tts";
             ocr: "ocr";
             vad: "vad";

package/dist/schemas/get-model-info.js

@@ -37,6 +37,7 @@ export const modelInfoSchema = z.object({
     addon: z.enum([
         "llm",
         "whisper",
+        "parakeet",
         "embeddings",
         "nmt",
         "vad",

package/dist/schemas/index.d.ts

@@ -18,7 +18,7 @@ export * from "./transcription";
 export * from "./translate";
 export * from "./translation-config";
 export * from "./llamacpp-config";
-export * from "./whispercpp-config";
+export * from "./transcription-config";
 export * from "./text-to-speech";
 export * from "./error";
 export * from "./rag";
@@ -28,6 +28,7 @@ export { SDK_CLIENT_ERROR_CODES } from "./sdk-errors-client";
 export { SDK_SERVER_ERROR_CODES } from "./sdk-errors-server";
 export { REGISTRY_ERROR_CODES } from "./sdk-errors-registry";
 export { qvacConfigSchema, deviceMatchSchema, deviceConfigDefaultsSchema, devicePatternSchema, type QvacConfig, type DeviceMatch, type DeviceConfigDefaults, type DevicePattern, } from "./sdk-config";
+export { PROFILING_KEY, PROFILING_TRAILER_KEY, DELEGATION_BREAKDOWN_KEY, OPERATION_EVENT_KEY, profilerModeSchema, serverBreakdownSchema, delegationBreakdownSchema, operationEventSchema, profilingRequestMetaSchema, profilingResponseMetaSchema, perCallProfilingSchema, type ProfilerMode, type ServerBreakdown, type DelegationBreakdown, type OperationEvent, type ProfilingRequestMeta, type ProfilingResponseMeta, type PerCallProfiling, } from "./profiling";
 export { runtimeContextSchema, type RuntimeContext } from "./runtime-context";
 export * from "./get-model-info";
 export * from "./model-src-utils";

package/dist/schemas/index.js

@@ -19,7 +19,7 @@ export * from "./transcription.js";
 export * from "./translate.js";
 export * from "./translation-config.js";
 export * from "./llamacpp-config.js";
-export * from "./whispercpp-config.js";
+export * from "./transcription-config.js";
 export * from "./text-to-speech.js";
 export * from "./error.js";
 export * from "./rag.js";
@@ -29,6 +29,7 @@ export { SDK_CLIENT_ERROR_CODES } from "./sdk-errors-client.js";
 export { SDK_SERVER_ERROR_CODES } from "./sdk-errors-server.js";
 export { REGISTRY_ERROR_CODES } from "./sdk-errors-registry.js";
 export { qvacConfigSchema, deviceMatchSchema, deviceConfigDefaultsSchema, devicePatternSchema, } from "./sdk-config.js";
+export { PROFILING_KEY, PROFILING_TRAILER_KEY, DELEGATION_BREAKDOWN_KEY, OPERATION_EVENT_KEY, profilerModeSchema, serverBreakdownSchema, delegationBreakdownSchema, operationEventSchema, profilingRequestMetaSchema, profilingResponseMetaSchema, perCallProfilingSchema, } from "./profiling.js";
 export { runtimeContextSchema } from "./runtime-context.js";
 export * from "./get-model-info.js";
 export * from "./model-src-utils.js";

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

@@ -24,6 +24,34 @@ export declare const llmConfigBaseSchema: z.ZodObject<{
     stop_sequences: z.ZodOptional<z.ZodArray<z.ZodString>>;
     n_discarded: z.ZodOptional<z.ZodNumber>;
     tools: z.ZodOptional<z.ZodBoolean>;
+    projectionModelSrc: z.ZodOptional<z.ZodUnion<readonly [z.ZodString, z.ZodObject<{
+        src: z.ZodString;
+        name: z.ZodOptional<z.ZodString>;
+        modelId: z.ZodOptional<z.ZodString>;
+        registryPath: z.ZodOptional<z.ZodString>;
+        registrySource: z.ZodOptional<z.ZodString>;
+        blobCoreKey: z.ZodOptional<z.ZodString>;
+        blobIndex: z.ZodOptional<z.ZodNumber>;
+        engine: z.ZodOptional<z.ZodString>;
+        expectedSize: z.ZodOptional<z.ZodNumber>;
+        sha256Checksum: z.ZodOptional<z.ZodString>;
+        addon: z.ZodOptional<z.ZodUnion<readonly [z.ZodEnum<{
+            "llamacpp-completion": "llamacpp-completion";
+            "whispercpp-transcription": "whispercpp-transcription";
+            "llamacpp-embedding": "llamacpp-embedding";
+            "nmtcpp-translation": "nmtcpp-translation";
+            "onnx-tts": "onnx-tts";
+            "parakeet-transcription": "parakeet-transcription";
+            "onnx-ocr": "onnx-ocr";
+            llm: "llm";
+            whisper: "whisper";
+            embeddings: "embeddings";
+            nmt: "nmt";
+            parakeet: "parakeet";
+            tts: "tts";
+            ocr: "ocr";
+        }>, z.ZodLiteral<"vad">]>>;
+    }, z.core.$strip>]>>;
 }, z.core.$strip>;
 export type LlmConfigInput = z.infer<typeof llmConfigBaseSchema>;
 export declare const LLM_CONFIG_DEFAULTS: {
@@ -51,6 +79,34 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
     stop_sequences: z.ZodOptional<z.ZodArray<z.ZodString>>;
     n_discarded: z.ZodOptional<z.ZodNumber>;
     tools: z.ZodOptional<z.ZodBoolean>;
+    projectionModelSrc: z.ZodOptional<z.ZodUnion<readonly [z.ZodString, z.ZodObject<{
+        src: z.ZodString;
+        name: z.ZodOptional<z.ZodString>;
+        modelId: z.ZodOptional<z.ZodString>;
+        registryPath: z.ZodOptional<z.ZodString>;
+        registrySource: z.ZodOptional<z.ZodString>;
+        blobCoreKey: z.ZodOptional<z.ZodString>;
+        blobIndex: z.ZodOptional<z.ZodNumber>;
+        engine: z.ZodOptional<z.ZodString>;
+        expectedSize: z.ZodOptional<z.ZodNumber>;
+        sha256Checksum: z.ZodOptional<z.ZodString>;
+        addon: z.ZodOptional<z.ZodUnion<readonly [z.ZodEnum<{
+            "llamacpp-completion": "llamacpp-completion";
+            "whispercpp-transcription": "whispercpp-transcription";
+            "llamacpp-embedding": "llamacpp-embedding";
+            "nmtcpp-translation": "nmtcpp-translation";
+            "onnx-tts": "onnx-tts";
+            "parakeet-transcription": "parakeet-transcription";
+            "onnx-ocr": "onnx-ocr";
+            llm: "llm";
+            whisper: "whisper";
+            embeddings: "embeddings";
+            nmt: "nmt";
+            parakeet: "parakeet";
+            tts: "tts";
+            ocr: "ocr";
+        }>, z.ZodLiteral<"vad">]>>;
+    }, z.core.$strip>]>>;
 }, z.core.$strip>, z.ZodTransform<{
     ctx_size: number | undefined;
     temp?: number | undefined;
@@ -70,6 +126,19 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
     stop_sequences?: string[] | undefined;
     n_discarded?: number | undefined;
     tools?: boolean | undefined;
+    projectionModelSrc?: string | {
+        src: string;
+        name?: string | undefined;
+        modelId?: string | undefined;
+        registryPath?: string | undefined;
+        registrySource?: string | undefined;
+        blobCoreKey?: string | undefined;
+        blobIndex?: number | undefined;
+        engine?: string | undefined;
+        expectedSize?: number | undefined;
+        sha256Checksum?: string | undefined;
+        addon?: "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "parakeet-transcription" | "onnx-ocr" | "llm" | "whisper" | "embeddings" | "nmt" | "parakeet" | "tts" | "ocr" | "vad" | undefined;
+    } | undefined;
 }, {
     ctx_size?: number | undefined;
     temp?: number | undefined;
@@ -89,18 +158,50 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
     stop_sequences?: string[] | undefined;
     n_discarded?: number | undefined;
     tools?: boolean | undefined;
+    projectionModelSrc?: string | {
+        src: string;
+        name?: string | undefined;
+        modelId?: string | undefined;
+        registryPath?: string | undefined;
+        registrySource?: string | undefined;
+        blobCoreKey?: string | undefined;
+        blobIndex?: number | undefined;
+        engine?: string | undefined;
+        expectedSize?: number | undefined;
+        sha256Checksum?: string | undefined;
+        addon?: "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "parakeet-transcription" | "onnx-ocr" | "llm" | "whisper" | "embeddings" | "nmt" | "parakeet" | "tts" | "ocr" | "vad" | undefined;
+    } | undefined;
 }>>;
 export type LlmConfig = z.infer<typeof llmConfigSchema>;
 export declare const embedConfigBaseSchema: z.ZodObject<{
     gpuLayers: z.ZodOptional<z.ZodNumber>;
-    device: z.ZodOptional<z.ZodString>;
+    device: z.ZodOptional<z.ZodEnum<{
+        gpu: "gpu";
+        cpu: "cpu";
+    }>>;
     batchSize: z.ZodOptional<z.ZodNumber>;
-    ctxSize: z.ZodOptional<z.ZodNumber>;
+    pooling: z.ZodOptional<z.ZodEnum<{
+        none: "none";
+        mean: "mean";
+        cls: "cls";
+        last: "last";
+        rank: "rank";
+    }>>;
+    attention: z.ZodOptional<z.ZodEnum<{
+        causal: "causal";
+        "non-causal": "non-causal";
+    }>>;
+    embdNormalize: z.ZodOptional<z.ZodNumber>;
     flashAttention: z.ZodOptional<z.ZodEnum<{
         on: "on";
         off: "off";
+        auto: "auto";
     }>>;
-    rawConfig: z.ZodOptional<z.ZodString>;
+    mainGpu: z.ZodOptional<z.ZodUnion<readonly [z.ZodNumber, z.ZodEnum<{
+        integrated: "integrated";
+        dedicated: "dedicated";
+    }>]>>;
+    verbosity: z.ZodOptional<z.ZodUnion<readonly [z.ZodLiteral<0>, z.ZodLiteral<1>, z.ZodLiteral<2>, z.ZodLiteral<3>]>>;
 }, z.core.$strip>;
 export type EmbedConfigInput = z.infer<typeof embedConfigBaseSchema>;
 export declare const EMBED_CONFIG_DEFAULTS: {
@@ -110,28 +211,53 @@ export declare const EMBED_CONFIG_DEFAULTS: {
 };
 export declare const embedConfigSchema: z.ZodPipe<z.ZodObject<{
     gpuLayers: z.ZodOptional<z.ZodNumber>;
-    device: z.ZodOptional<z.ZodString>;
+    device: z.ZodOptional<z.ZodEnum<{
+        gpu: "gpu";
+        cpu: "cpu";
+    }>>;
     batchSize: z.ZodOptional<z.ZodNumber>;
-    ctxSize: z.ZodOptional<z.ZodNumber>;
+    pooling: z.ZodOptional<z.ZodEnum<{
+        none: "none";
+        mean: "mean";
+        cls: "cls";
+        last: "last";
+        rank: "rank";
+    }>>;
+    attention: z.ZodOptional<z.ZodEnum<{
+        causal: "causal";
+        "non-causal": "non-causal";
+    }>>;
+    embdNormalize: z.ZodOptional<z.ZodNumber>;
     flashAttention: z.ZodOptional<z.ZodEnum<{
         on: "on";
         off: "off";
+        auto: "auto";
     }>>;
-    rawConfig: z.ZodOptional<z.ZodString>;
+    mainGpu: z.ZodOptional<z.ZodUnion<readonly [z.ZodNumber, z.ZodEnum<{
+        integrated: "integrated";
+        dedicated: "dedicated";
+    }>]>>;
+    verbosity: z.ZodOptional<z.ZodUnion<readonly [z.ZodLiteral<0>, z.ZodLiteral<1>, z.ZodLiteral<2>, z.ZodLiteral<3>]>>;
 }, z.core.$strip>, z.ZodTransform<{
     gpuLayers: number | undefined;
-    device: string | undefined;
+    device: "gpu" | "cpu" | undefined;
     batchSize: number | undefined;
-    ctxSize?: number | undefined;
-    flashAttention?: "on" | "off" | undefined;
-    rawConfig?: string | undefined;
+    pooling?: "none" | "mean" | "cls" | "last" | "rank" | undefined;
+    attention?: "causal" | "non-causal" | undefined;
+    embdNormalize?: number | undefined;
+    flashAttention?: "on" | "off" | "auto" | undefined;
+    mainGpu?: number | "integrated" | "dedicated" | undefined;
+    verbosity?: 0 | 3 | 2 | 1 | undefined;
 }, {
     gpuLayers?: number | undefined;
-    device?: string | undefined;
+    device?: "gpu" | "cpu" | undefined;
     batchSize?: number | undefined;
-    ctxSize?: number | undefined;
-    flashAttention?: "on" | "off" | undefined;
-    rawConfig?: string | undefined;
+    pooling?: "none" | "mean" | "cls" | "last" | "rank" | undefined;
+    attention?: "causal" | "non-causal" | undefined;
+    embdNormalize?: number | undefined;
+    flashAttention?: "on" | "off" | "auto" | undefined;
+    mainGpu?: number | "integrated" | "dedicated" | undefined;
+    verbosity?: 0 | 3 | 2 | 1 | undefined;
 }>>;
 export type EmbedConfig = z.infer<typeof embedConfigSchema>;
 //# sourceMappingURL=llamacpp-config.d.ts.map

package/dist/schemas/llamacpp-config.js

@@ -1,4 +1,5 @@
 import { z } from "zod";
+import { modelSrcInputSchema } from "./model-src-utils.js";
 export const VERBOSITY = {
     ERROR: 0,
     WARN: 1,
@@ -37,6 +38,7 @@ export const llmConfigBaseSchema = z.object({
     stop_sequences: z.array(z.string()).optional(),
     n_discarded: z.number().optional(),
     tools: z.boolean().optional(),
+    projectionModelSrc: modelSrcInputSchema.optional(),
 });
 // Default values - typed as partial of the config
 export const LLM_CONFIG_DEFAULTS = {
@@ -53,14 +55,16 @@ export const llmConfigSchema = llmConfigBaseSchema.transform((data) => ({
 // Base schema - validates types, all fields optional (for client-side validation)
 export const embedConfigBaseSchema = z.object({
     gpuLayers: z.number().int().optional(),
-    device: z.string().optional(),
+    device: z.enum(["gpu", "cpu"]).optional(),
     batchSize: z.number().int().min(1).optional(),
-    ctxSize: z.number().int().min(1).optional(),
-    flashAttention: z.enum(["on", "off"]).optional(),
-    // Raw CLI override (advanced, takes precedence if set)
-    // Format: "-ngl\t99\n-dev\tgpu\n--batch_size\t1024"
-    // See: https://github.com/tetherto/qvac-lib-infer-llamacpp-embed/tree/main?tab=readme-ov-file#4-create-config
-    rawConfig: z.string().optional(),
+    pooling: z.enum(["none", "mean", "cls", "last", "rank"]).optional(),
+    attention: z.enum(["causal", "non-causal"]).optional(),
+    embdNormalize: z.number().int().optional(),
+    flashAttention: z.enum(["on", "off", "auto"]).optional(),
+    mainGpu: z
+        .union([z.number().int().min(0), z.enum(["integrated", "dedicated"])])
+        .optional(),
+    verbosity: verbositySchema.optional(),
 });
 // Default values - typed as partial of the config
 export const EMBED_CONFIG_DEFAULTS = {