@qvac/sdk 0.4.1 → 0.6.0

package/dist/schemas/common.js

@@ -13,10 +13,10 @@ import { translateRequestSchema, translateResponseSchema } from "./translate.js"
 import { loggingStreamRequestSchema, loggingStreamResponseSchema, } from "./logging-stream.js";
 import { ttsRequestSchema, ttsResponseSchema } from "./text-to-speech.js";
 import { errorResponseSchema } from "./error.js";
-import { ragRequestSchema, ragResponseSchema } from "./rag.js";
+import { ragRequestSchema, ragResponseSchema, ragProgressUpdateSchema, } from "./rag.js";
 import { deleteCacheRequestSchema, deleteCacheResponseSchema, } from "./delete-cache.js";
-import { setConfigRequestSchema, setConfigResponseSchema } from "./set-config.js";
 import { getModelInfoRequestSchema, getModelInfoResponseSchema, } from "./get-model-info.js";
+import { ocrStreamRequestSchema, ocrStreamResponseSchema } from "./ocr.js";
 export const requestSchema = z.union([
     pingRequestSchema,
     loadModelRequestSchema,
@@ -33,8 +33,8 @@ export const requestSchema = z.union([
     stopProvideRequestSchema,
     ragRequestSchema,
     deleteCacheRequestSchema,
-    setConfigRequestSchema,
     getModelInfoRequestSchema,
+    ocrStreamRequestSchema,
 ]);
 export const responseSchema = z.discriminatedUnion("type", [
     pingResponseSchema,
@@ -53,10 +53,12 @@ export const responseSchema = z.discriminatedUnion("type", [
     stopProvideResponseSchema,
     errorResponseSchema,
     ragResponseSchema,
+    ragProgressUpdateSchema,
     deleteCacheResponseSchema,
-    setConfigResponseSchema,
     getModelInfoResponseSchema,
+    ocrStreamResponseSchema,
 ]);
 export const rpcOptionsSchema = z.object({
     timeout: z.number().min(100).optional(),
+    forceNewConnection: z.boolean().optional(),
 });

package/dist/schemas/delegate.d.ts

@@ -4,6 +4,7 @@ export declare const delegateSchema: z.ZodOptional<z.ZodObject<{
     providerPublicKey: z.ZodString;
     timeout: z.ZodOptional<z.ZodNumber>;
     fallbackToLocal: z.ZodDefault<z.ZodOptional<z.ZodBoolean>>;
+    forceNewConnection: z.ZodDefault<z.ZodOptional<z.ZodBoolean>>;
 }, z.core.$strip>>;
 export type Delegate = z.infer<typeof delegateSchema>;
 //# sourceMappingURL=delegate.d.ts.map

package/dist/schemas/delegate.js

@@ -5,5 +5,6 @@ export const delegateSchema = z
     providerPublicKey: z.string(),
     timeout: z.number().min(100).optional(),
     fallbackToLocal: z.boolean().optional().default(false),
+    forceNewConnection: z.boolean().optional().default(false),
 })
     .optional();

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

@@ -9,14 +9,20 @@ export declare const downloadAssetOptionsSchema: z.ZodPipe<z.ZodObject<{
         hyperbeeKey: z.ZodOptional<z.ZodString>;
         expectedSize: z.ZodOptional<z.ZodNumber>;
         sha256Checksum: z.ZodOptional<z.ZodString>;
-        addon: z.ZodOptional<z.ZodEnum<{
+        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";
+            "onnx-ocr": "onnx-ocr";
             llm: "llm";
             whisper: "whisper";
             embeddings: "embeddings";
             nmt: "nmt";
-            vad: "vad";
             tts: "tts";
-        }>>;
+            ocr: "ocr";
+        }>, z.ZodLiteral<"vad">]>>;
     }, z.core.$strip>]>;
     seed: z.ZodOptional<z.ZodBoolean>;
 }, z.core.$strip>, z.ZodTransform<{
@@ -29,7 +35,7 @@ export declare const downloadAssetOptionsSchema: z.ZodPipe<z.ZodObject<{
         hyperbeeKey?: string | undefined;
         expectedSize?: number | undefined;
         sha256Checksum?: string | undefined;
-        addon?: "llm" | "whisper" | "embeddings" | "nmt" | "vad" | "tts" | undefined;
+        addon?: "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "onnx-ocr" | "llm" | "whisper" | "embeddings" | "nmt" | "tts" | "ocr" | "vad" | undefined;
     };
 }, {
     assetSrc: string | {
@@ -40,7 +46,7 @@ export declare const downloadAssetOptionsSchema: z.ZodPipe<z.ZodObject<{
         hyperbeeKey?: string | undefined;
         expectedSize?: number | undefined;
         sha256Checksum?: string | undefined;
-        addon?: "llm" | "whisper" | "embeddings" | "nmt" | "vad" | "tts" | undefined;
+        addon?: "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "onnx-ocr" | "llm" | "whisper" | "embeddings" | "nmt" | "tts" | "ocr" | "vad" | undefined;
     };
     seed?: boolean | undefined;
 }>>;
@@ -53,14 +59,20 @@ export declare const downloadAssetOptionsToRequestSchema: z.ZodPipe<z.ZodObject<
         hyperbeeKey: z.ZodOptional<z.ZodString>;
         expectedSize: z.ZodOptional<z.ZodNumber>;
         sha256Checksum: z.ZodOptional<z.ZodString>;
-        addon: z.ZodOptional<z.ZodEnum<{
+        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";
+            "onnx-ocr": "onnx-ocr";
             llm: "llm";
             whisper: "whisper";
             embeddings: "embeddings";
             nmt: "nmt";
-            vad: "vad";
             tts: "tts";
-        }>>;
+            ocr: "ocr";
+        }>, z.ZodLiteral<"vad">]>>;
     }, z.core.$strip>]>;
     seed: z.ZodOptional<z.ZodBoolean>;
     onProgress: z.ZodOptional<z.ZodUnknown>;
@@ -79,7 +91,7 @@ export declare const downloadAssetOptionsToRequestSchema: z.ZodPipe<z.ZodObject<
         hyperbeeKey?: string | undefined;
         expectedSize?: number | undefined;
         sha256Checksum?: string | undefined;
-        addon?: "llm" | "whisper" | "embeddings" | "nmt" | "vad" | "tts" | undefined;
+        addon?: "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "onnx-ocr" | "llm" | "whisper" | "embeddings" | "nmt" | "tts" | "ocr" | "vad" | undefined;
     };
     seed?: boolean | undefined;
     onProgress?: unknown;

package/dist/schemas/embed.d.ts

@@ -1,17 +1,17 @@
 import { z } from "zod";
 export declare const embedParamsSchema: z.ZodObject<{
     modelId: z.ZodString;
-    text: z.ZodString;
+    text: z.ZodUnion<readonly [z.ZodString, z.ZodArray<z.ZodString>]>;
 }, z.core.$strip>;
 export declare const embedRequestSchema: z.ZodObject<{
     modelId: z.ZodString;
-    text: z.ZodString;
+    text: z.ZodUnion<readonly [z.ZodString, z.ZodArray<z.ZodString>]>;
     type: z.ZodLiteral<"embed">;
 }, z.core.$strip>;
 export declare const embedResponseSchema: z.ZodObject<{
     type: z.ZodLiteral<"embed">;
     success: z.ZodBoolean;
-    embedding: z.ZodDefault<z.ZodArray<z.ZodNumber>>;
+    embedding: z.ZodDefault<z.ZodUnion<readonly [z.ZodArray<z.ZodNumber>, z.ZodArray<z.ZodArray<z.ZodNumber>>]>>;
     error: z.ZodOptional<z.ZodString>;
 }, z.core.$strip>;
 export type EmbedParams = z.infer<typeof embedParamsSchema>;

package/dist/schemas/embed.js

@@ -1,7 +1,12 @@
 import { z } from "zod";
 export const embedParamsSchema = z.object({
     modelId: z.string(),
-    text: z.string().min(1, "Text cannot be empty"),
+    text: z.union([
+        z.string().min(1, "Text cannot be empty"),
+        z
+            .array(z.string().min(1, "Text cannot be empty"))
+            .min(1, "Text array cannot be empty"),
+    ]),
 });
 export const embedRequestSchema = embedParamsSchema.extend({
     type: z.literal("embed"),
@@ -9,6 +14,8 @@ export const embedRequestSchema = embedParamsSchema.extend({
 export const embedResponseSchema = z.object({
     type: z.literal("embed"),
     success: z.boolean(),
-    embedding: z.array(z.number()).default([]),
+    embedding: z
+        .union([z.array(z.number()), z.array(z.array(z.number()))])
+        .default([]),
     error: z.string().optional(),
 });

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

@@ -28,12 +28,13 @@ export declare const modelInfoSchema: z.ZodObject<{
     expectedSize: z.ZodNumber;
     sha256Checksum: z.ZodString;
     addon: z.ZodEnum<{
-        llm: "llm";
-        whisper: "whisper";
-        embeddings: "embeddings";
-        nmt: "nmt";
+        "llamacpp-completion": "llamacpp-completion";
+        "whispercpp-transcription": "whispercpp-transcription";
+        "llamacpp-embedding": "llamacpp-embedding";
+        "nmtcpp-translation": "nmtcpp-translation";
+        "onnx-tts": "onnx-tts";
+        "onnx-ocr": "onnx-ocr";
         vad: "vad";
-        tts: "tts";
     }>;
     isCached: z.ZodBoolean;
     isLoaded: z.ZodBoolean;
@@ -64,12 +65,13 @@ export declare const getModelInfoResponseSchema: z.ZodObject<{
         expectedSize: z.ZodNumber;
         sha256Checksum: z.ZodString;
         addon: z.ZodEnum<{
-            llm: "llm";
-            whisper: "whisper";
-            embeddings: "embeddings";
-            nmt: "nmt";
+            "llamacpp-completion": "llamacpp-completion";
+            "whispercpp-transcription": "whispercpp-transcription";
+            "llamacpp-embedding": "llamacpp-embedding";
+            "nmtcpp-translation": "nmtcpp-translation";
+            "onnx-tts": "onnx-tts";
+            "onnx-ocr": "onnx-ocr";
             vad: "vad";
-            tts: "tts";
         }>;
         isCached: z.ZodBoolean;
         isLoaded: z.ZodBoolean;

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

@@ -26,7 +26,15 @@ export const modelInfoSchema = z.object({
     hyperbeeKey: z.string(),
     expectedSize: z.number(),
     sha256Checksum: z.string(),
-    addon: z.enum(["llm", "whisper", "embeddings", "nmt", "vad", "tts"]),
+    addon: z.enum([
+        "llamacpp-completion",
+        "whispercpp-transcription",
+        "llamacpp-embedding",
+        "nmtcpp-translation",
+        "vad",
+        "onnx-tts",
+        "onnx-ocr",
+    ]),
     isCached: z.boolean(),
     isLoaded: z.boolean(),
     cacheFiles: z.array(cacheFileInfoSchema),

package/dist/schemas/index.d.ts

@@ -1,3 +1,4 @@
+export * from "./archive";
 export * from "./cancel";
 export * from "./completion-stream";
 export * from "./tools";
@@ -6,6 +7,7 @@ export * from "./delete-cache";
 export * from "./download-asset";
 export * from "./embed";
 export * from "./load-model";
+export * from "./reload-config";
 export * from "./logging-stream";
 export * from "./provide";
 export * from "./stop-provide";
@@ -20,9 +22,15 @@ export * from "./whispercpp-config";
 export * from "./text-to-speech";
 export * from "./error";
 export * from "./rag";
+export * from "./ocr";
+export * from "./shard";
 export { SDK_CLIENT_ERROR_CODES } from "./sdk-errors-client";
 export { SDK_SERVER_ERROR_CODES } from "./sdk-errors-server";
-export * from "./set-config";
+export { qvacConfigSchema, deviceMatchSchema, deviceConfigDefaultsSchema, devicePatternSchema, type QvacConfig, type DeviceMatch, type DeviceConfigDefaults, type DevicePattern, } from "./sdk-config";
+export { runtimeContextSchema, type RuntimeContext } from "./runtime-context";
 export * from "./get-model-info";
 export * from "./model-src-utils";
+export * from "./json-schema";
+export { type McpClient, type McpClientInput } from "./mcp-adapter";
+export { ModelType, ModelTypeAliases, PUBLIC_MODEL_TYPES, modelTypeInputSchema, modelTypeSchema, normalizeModelType, isModelTypeAlias, llmModelTypeSchema, whisperModelTypeSchema, embeddingsModelTypeSchema, nmtModelTypeSchema, ttsModelTypeSchema, ocrModelTypeSchema, type CanonicalModelType, type AliasKey, type ModelTypeInput, type LlmModelTypeInput, type WhisperModelTypeInput, type EmbeddingsModelTypeInput, type NmtModelTypeInput, type TtsModelTypeInput, type OcrModelTypeInput, } from "./model-types";
 //# sourceMappingURL=index.d.ts.map

package/dist/schemas/index.js

@@ -1,4 +1,5 @@
 // Re-export all schemas and types
+export * from "./archive.js";
 export * from "./cancel.js";
 export * from "./completion-stream.js";
 export * from "./tools.js";
@@ -7,6 +8,7 @@ export * from "./delete-cache.js";
 export * from "./download-asset.js";
 export * from "./embed.js";
 export * from "./load-model.js";
+export * from "./reload-config.js";
 export * from "./logging-stream.js";
 export * from "./provide.js";
 export * from "./stop-provide.js";
@@ -21,8 +23,17 @@ export * from "./whispercpp-config.js";
 export * from "./text-to-speech.js";
 export * from "./error.js";
 export * from "./rag.js";
+export * from "./ocr.js";
+export * from "./shard.js";
 export { SDK_CLIENT_ERROR_CODES } from "./sdk-errors-client.js";
 export { SDK_SERVER_ERROR_CODES } from "./sdk-errors-server.js";
-export * from "./set-config.js";
+export { qvacConfigSchema, deviceMatchSchema, deviceConfigDefaultsSchema, devicePatternSchema, } from "./sdk-config.js";
+export { runtimeContextSchema } from "./runtime-context.js";
 export * from "./get-model-info.js";
 export * from "./model-src-utils.js";
+export * from "./json-schema.js";
+export {} from "./mcp-adapter.js";
+// Model types - new canonical naming system
+export { ModelType, ModelTypeAliases, PUBLIC_MODEL_TYPES, modelTypeInputSchema, modelTypeSchema, normalizeModelType, isModelTypeAlias, 
+// Per-model-type schemas for discriminated unions
+llmModelTypeSchema, whisperModelTypeSchema, embeddingsModelTypeSchema, nmtModelTypeSchema, ttsModelTypeSchema, ocrModelTypeSchema, } from "./model-types.js";

package/dist/schemas/json-schema.d.ts

@@ -0,0 +1,16 @@
+export type JsonSchemaType = "string" | "number" | "integer" | "boolean" | "object" | "array" | "null";
+export type JsonSchemaProperty = {
+    type?: JsonSchemaType | JsonSchemaType[];
+    description?: string;
+    enum?: (string | number | boolean | null)[];
+    items?: JsonSchema;
+    properties?: Record<string, JsonSchemaProperty>;
+    required?: string[];
+    default?: unknown;
+    additionalProperties?: boolean | JsonSchemaProperty;
+};
+export type JsonSchema = JsonSchemaProperty & {
+    $schema?: string;
+    title?: string;
+};
+//# sourceMappingURL=json-schema.d.ts.map

package/dist/schemas/json-schema.js

@@ -0,0 +1,2 @@
+// ============== JSON Schema Types ==============
+export {};

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

@@ -1,22 +1,23 @@
 import { z } from "zod";
 export declare const VERBOSITY: {
-    readonly SILENT: -1;
-    readonly INFO: 0;
-    readonly DEBUG: 1;
+    readonly ERROR: 0;
+    readonly WARN: 1;
+    readonly INFO: 2;
+    readonly DEBUG: 3;
 };
-export declare const llmConfigSchema: z.ZodObject<{
-    ctx_size: z.ZodDefault<z.ZodNumber>;
+export declare const llmConfigBaseSchema: z.ZodObject<{
+    ctx_size: z.ZodOptional<z.ZodNumber>;
     temp: z.ZodOptional<z.ZodNumber>;
     top_p: z.ZodOptional<z.ZodNumber>;
     top_k: z.ZodOptional<z.ZodNumber>;
     seed: z.ZodOptional<z.ZodNumber>;
-    gpu_layers: z.ZodDefault<z.ZodNumber>;
+    gpu_layers: z.ZodOptional<z.ZodNumber>;
     lora: z.ZodOptional<z.ZodString>;
-    device: z.ZodDefault<z.ZodString>;
+    device: z.ZodOptional<z.ZodString>;
     predict: z.ZodOptional<z.ZodUnion<readonly [z.ZodLiteral<-1>, z.ZodLiteral<-2>, z.ZodNumber]>>;
-    system_prompt: z.ZodDefault<z.ZodString>;
+    system_prompt: z.ZodOptional<z.ZodString>;
     no_mmap: z.ZodOptional<z.ZodBoolean>;
-    verbosity: z.ZodOptional<z.ZodUnion<readonly [z.ZodLiteral<-1>, z.ZodLiteral<0>, z.ZodLiteral<1>]>>;
+    verbosity: z.ZodOptional<z.ZodUnion<readonly [z.ZodLiteral<0>, z.ZodLiteral<1>, z.ZodLiteral<2>, z.ZodLiteral<3>]>>;
     presence_penalty: z.ZodOptional<z.ZodNumber>;
     frequency_penalty: z.ZodOptional<z.ZodNumber>;
     repeat_penalty: z.ZodOptional<z.ZodNumber>;
@@ -24,9 +25,113 @@ export declare const llmConfigSchema: z.ZodObject<{
     n_discarded: z.ZodOptional<z.ZodNumber>;
     tools: z.ZodOptional<z.ZodBoolean>;
 }, z.core.$strip>;
-export declare const embedConfigSchema: z.ZodObject<{
-    config: z.ZodDefault<z.ZodString>;
-}, z.core.$strip>;
+export type LlmConfigInput = z.infer<typeof llmConfigBaseSchema>;
+export declare const LLM_CONFIG_DEFAULTS: {
+    readonly ctx_size: 1024;
+    readonly gpu_layers: 99;
+    readonly device: "gpu";
+    readonly system_prompt: "You are a helpful assistant.";
+};
+export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
+    ctx_size: z.ZodOptional<z.ZodNumber>;
+    temp: z.ZodOptional<z.ZodNumber>;
+    top_p: z.ZodOptional<z.ZodNumber>;
+    top_k: z.ZodOptional<z.ZodNumber>;
+    seed: z.ZodOptional<z.ZodNumber>;
+    gpu_layers: z.ZodOptional<z.ZodNumber>;
+    lora: z.ZodOptional<z.ZodString>;
+    device: z.ZodOptional<z.ZodString>;
+    predict: z.ZodOptional<z.ZodUnion<readonly [z.ZodLiteral<-1>, z.ZodLiteral<-2>, z.ZodNumber]>>;
+    system_prompt: z.ZodOptional<z.ZodString>;
+    no_mmap: z.ZodOptional<z.ZodBoolean>;
+    verbosity: z.ZodOptional<z.ZodUnion<readonly [z.ZodLiteral<0>, z.ZodLiteral<1>, z.ZodLiteral<2>, z.ZodLiteral<3>]>>;
+    presence_penalty: z.ZodOptional<z.ZodNumber>;
+    frequency_penalty: z.ZodOptional<z.ZodNumber>;
+    repeat_penalty: z.ZodOptional<z.ZodNumber>;
+    stop_sequences: z.ZodOptional<z.ZodArray<z.ZodString>>;
+    n_discarded: z.ZodOptional<z.ZodNumber>;
+    tools: z.ZodOptional<z.ZodBoolean>;
+}, z.core.$strip>, z.ZodTransform<{
+    ctx_size: number | undefined;
+    temp?: number | undefined;
+    top_p?: number | undefined;
+    top_k?: number | undefined;
+    seed?: number | undefined;
+    gpu_layers: number | undefined;
+    lora?: string | undefined;
+    device: string | undefined;
+    predict?: number | undefined;
+    system_prompt: string | undefined;
+    no_mmap?: boolean | undefined;
+    verbosity?: 0 | 3 | 2 | 1 | undefined;
+    presence_penalty?: number | undefined;
+    frequency_penalty?: number | undefined;
+    repeat_penalty?: number | undefined;
+    stop_sequences?: string[] | undefined;
+    n_discarded?: number | undefined;
+    tools?: boolean | undefined;
+}, {
+    ctx_size?: number | undefined;
+    temp?: number | undefined;
+    top_p?: number | undefined;
+    top_k?: number | undefined;
+    seed?: number | undefined;
+    gpu_layers?: number | undefined;
+    lora?: string | undefined;
+    device?: string | undefined;
+    predict?: number | undefined;
+    system_prompt?: string | undefined;
+    no_mmap?: boolean | undefined;
+    verbosity?: 0 | 3 | 2 | 1 | undefined;
+    presence_penalty?: number | undefined;
+    frequency_penalty?: number | undefined;
+    repeat_penalty?: number | undefined;
+    stop_sequences?: string[] | undefined;
+    n_discarded?: number | undefined;
+    tools?: boolean | undefined;
+}>>;
 export type LlmConfig = z.infer<typeof llmConfigSchema>;
+export declare const embedConfigBaseSchema: z.ZodObject<{
+    gpuLayers: z.ZodOptional<z.ZodNumber>;
+    device: z.ZodOptional<z.ZodString>;
+    batchSize: z.ZodOptional<z.ZodNumber>;
+    ctxSize: z.ZodOptional<z.ZodNumber>;
+    flashAttention: z.ZodOptional<z.ZodEnum<{
+        on: "on";
+        off: "off";
+    }>>;
+    rawConfig: z.ZodOptional<z.ZodString>;
+}, z.core.$strip>;
+export type EmbedConfigInput = z.infer<typeof embedConfigBaseSchema>;
+export declare const EMBED_CONFIG_DEFAULTS: {
+    readonly gpuLayers: 99;
+    readonly device: "gpu";
+    readonly batchSize: 1024;
+};
+export declare const embedConfigSchema: z.ZodPipe<z.ZodObject<{
+    gpuLayers: z.ZodOptional<z.ZodNumber>;
+    device: z.ZodOptional<z.ZodString>;
+    batchSize: z.ZodOptional<z.ZodNumber>;
+    ctxSize: z.ZodOptional<z.ZodNumber>;
+    flashAttention: z.ZodOptional<z.ZodEnum<{
+        on: "on";
+        off: "off";
+    }>>;
+    rawConfig: z.ZodOptional<z.ZodString>;
+}, z.core.$strip>, z.ZodTransform<{
+    gpuLayers: number | undefined;
+    device: string | undefined;
+    batchSize: number | undefined;
+    ctxSize?: number | undefined;
+    flashAttention?: "on" | "off" | undefined;
+    rawConfig?: string | undefined;
+}, {
+    gpuLayers?: number | undefined;
+    device?: string | undefined;
+    batchSize?: number | undefined;
+    ctxSize?: number | undefined;
+    flashAttention?: "on" | "off" | undefined;
+    rawConfig?: string | undefined;
+}>>;
 export type EmbedConfig = z.infer<typeof embedConfigSchema>;
 //# sourceMappingURL=llamacpp-config.d.ts.map

package/dist/schemas/llamacpp-config.js

@@ -1,23 +1,26 @@
 import { z } from "zod";
 export const VERBOSITY = {
-    SILENT: -1,
-    INFO: 0,
-    DEBUG: 1,
+    ERROR: 0,
+    WARN: 1,
+    INFO: 2,
+    DEBUG: 3,
 };
 const verbositySchema = z.union([
-    z.literal(VERBOSITY.SILENT),
+    z.literal(VERBOSITY.ERROR),
+    z.literal(VERBOSITY.WARN),
     z.literal(VERBOSITY.INFO),
     z.literal(VERBOSITY.DEBUG),
 ]);
-export const llmConfigSchema = z.object({
-    ctx_size: z.number().default(1024),
+// Base schema - validates types, all fields optional (for client-side validation)
+export const llmConfigBaseSchema = z.object({
+    ctx_size: z.number().optional(),
     temp: z.number().min(0).max(2).optional(),
     top_p: z.number().min(0).max(1).optional(),
     top_k: z.number().int().min(0).max(128).optional(),
     seed: z.number().optional(),
-    gpu_layers: z.number().default(99),
+    gpu_layers: z.number().optional(),
     lora: z.string().optional(),
-    device: z.string().default("gpu"),
+    device: z.string().optional(),
     predict: z
         .union([
         z.literal(-1), // special: until stop token
@@ -25,7 +28,7 @@ export const llmConfigSchema = z.object({
         z.number().int().min(1), // positive integer: fixed token count
     ])
         .optional(),
-    system_prompt: z.string().default("You are a helpful assistant."),
+    system_prompt: z.string().optional(),
     no_mmap: z.boolean().optional(),
     verbosity: verbositySchema.optional(),
     presence_penalty: z.number().optional(),
@@ -35,7 +38,38 @@ export const llmConfigSchema = z.object({
     n_discarded: z.number().optional(),
     tools: z.boolean().optional(),
 });
-export const embedConfigSchema = z.object({
-    // https://github.com/tetherto/qvac-lib-infer-llamacpp-embed/tree/main?tab=readme-ov-file#4-create-config
-    config: z.string().default("-ngl\t99\n-dev\tgpu"),
+// Default values - typed as partial of the config
+export const LLM_CONFIG_DEFAULTS = {
+    ctx_size: 1024,
+    gpu_layers: 99,
+    device: "gpu",
+    system_prompt: "You are a helpful assistant.",
+};
+// Full schema - applies defaults via transform (no duplication)
+export const llmConfigSchema = llmConfigBaseSchema.transform((data) => ({
+    ...LLM_CONFIG_DEFAULTS,
+    ...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(),
+    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(),
 });
+// Default values - typed as partial of the config
+export const EMBED_CONFIG_DEFAULTS = {
+    gpuLayers: 99,
+    device: "gpu",
+    batchSize: 1024,
+};
+// Full schema - validates then applies defaults via transform
+export const embedConfigSchema = embedConfigBaseSchema.transform((data) => ({
+    ...EMBED_CONFIG_DEFAULTS,
+    ...data,
+}));