@qvac/sdk 0.2.1-dev.1760851161.68dc306 → 0.2.1

package/README.md

@@ -79,7 +79,7 @@ npm i expo-file-system react-native-bare-kit
 ```js
 export default {
   expo: {
-    plugins: ["qvac-sdk/expo-plugin"],
+    plugins: ["@qvac/sdk/expo-plugin"],
   },
 };
 ```
@@ -141,6 +141,7 @@ const cachedResponse = completion({
 
 // Unload model to free up system resources
 await unloadModel({ modelId });
+process.exit();
 ```
 
 **For more on how to use QVAC SDK, see [SDK at QVAC documentation](https://tether-2.gitbook.io/qvac-by-tether-or-reference/y2AsIh94H9DJ1pIBCNj9/sdk)**.
@@ -161,10 +162,10 @@ bun run build  # or `watch` for hotreload
 bun run build:pack
 ```
 
-This outputs a tarball under `dist/tetherto-qvac-sdk-{version}.tgz` that you can install in your project, e.g.:
+This outputs a tarball under `dist/qvac-sdk-{version}.tgz` that you can install in your project, e.g.:
 
 ```bash
-npm i path/to/tetherto-qvac-sdk-0.0.1.tgz
+npm i path/to/qvac-sdk-0.1.0.tgz
 ```
 
 ## Examples

package/dist/examples/llamacpp-multimodal.js

@@ -1,4 +1,4 @@
-import { completion, loadModel, unloadModel } from "@qvac/sdk";
+import { completion, loadModel, SMOLVLM2_2_500M_MULTIMODAL_Q8_0, MMPROJ_SMOLVLM2_2_500M_MULTIMODAL_Q8_0, unloadModel, } from "@qvac/sdk";
 if (process.argv.length < 3) {
     console.error(`Specify an image file path as the first argument and a second image file path as the second (optional) argument`);
     process.exit(1);
@@ -8,9 +8,9 @@ try {
     const imageFilePath = process.argv[2];
     // Load the main model with projection in a single step
     const modelId = await loadModel({
-        modelSrc: "https://huggingface.co/ggml-org/SmolVLM2-500M-Video-Instruct-GGUF/resolve/main/SmolVLM2-500M-Video-Instruct-Q8_0.gguf",
+        modelSrc: SMOLVLM2_2_500M_MULTIMODAL_Q8_0,
         modelType: "llm",
-        projectionModelSrc: "https://huggingface.co/ggml-org/SmolVLM2-500M-Video-Instruct-GGUF/resolve/main/mmproj-SmolVLM2-500M-Video-Instruct-Q8_0.gguf",
+        projectionModelSrc: MMPROJ_SMOLVLM2_2_500M_MULTIMODAL_Q8_0,
         modelConfig: {
             ctx_size: 1024,
         },

package/dist/examples/rag/rag-hyperdb-workspaces.js

@@ -41,9 +41,10 @@ try {
         documents: medicalArticle,
         chunk: true,
         chunkOpts: {
-            chunkSize: 30,
+            chunkSize: 100,
             chunkOverlap: 10,
             chunkStrategy: "paragraph",
+            splitStrategy: "token",
         },
     });
     console.log(`✅ Created ${medicalResult.processed.length} chunks from medical article`);

package/dist/models/hyperdrive/models.d.ts

@@ -62,6 +62,20 @@ export declare const models: readonly [{
     readonly addon: "llm";
     readonly expectedSize: 1056782912;
     readonly sha256Checksum: "c876f159707a4e4f70e045106c69db15bfc935a4981706fd4f65c6e7ea1e81c5";
+}, {
+    readonly hyperdriveKey: "583f04c31d151b29e02aaf98f80d16aff585714ef9e2806a5067651c0f64ae31";
+    readonly hyperbeeKey: "generation:qwen2.5-omni:multimodal::3B::1.0.0";
+    readonly modelId: "Qwen2.5-Omni-3B-Q4_K_M.gguf";
+    readonly addon: "llm";
+    readonly expectedSize: 2104931648;
+    readonly sha256Checksum: "4b0bd358c1e9ec55dd3055ef6d71c958c821533d85916a10cfa89c4552a86e29";
+}, {
+    readonly hyperdriveKey: "583f04c31d151b29e02aaf98f80d16aff585714ef9e2806a5067651c0f64ae31";
+    readonly hyperbeeKey: "generation:qwen2.5-omni:multimodal::3B::1.0.0";
+    readonly modelId: "mmproj-Qwen2.5-Omni-3B-Q8_0.gguf";
+    readonly addon: "llm";
+    readonly expectedSize: 1538031328;
+    readonly sha256Checksum: "4e6c816cd33f7298d07cb780c136a396631e50e62f6501660271f8c6e302e565";
 }, {
     readonly hyperdriveKey: "d331ed49c444f90b3b2a70aa5b820752685142bf2d769a0592ed666140477578";
     readonly hyperbeeKey: "generation:qwen3:instruct::0.6B::1.0.0";
@@ -97,6 +111,34 @@ export declare const models: readonly [{
     readonly addon: "llm";
     readonly expectedSize: 2401075552;
     readonly sha256Checksum: "9cfebb63cd7e6b6d1755ea8ef001fcdadfa17831da1db96eafe3fc554913d5e1";
+}, {
+    readonly hyperdriveKey: "b99c6f5dcadaad827cdcb1a129a3b34aff6ad28056738e0bd559c7ffb807e30f";
+    readonly hyperbeeKey: "generation:smolVLM2:multimodal:2:500M:f16:1.0.0:video-instruct";
+    readonly modelId: "SmolVLM2-500M-Video-Instruct-f16.gguf";
+    readonly addon: "llm";
+    readonly expectedSize: 820424704;
+    readonly sha256Checksum: "80f7e3f04bc2d3324ac1a9f52f5776fe13a69912adf74f8e7edacf773d140d77";
+}, {
+    readonly hyperdriveKey: "b99c6f5dcadaad827cdcb1a129a3b34aff6ad28056738e0bd559c7ffb807e30f";
+    readonly hyperbeeKey: "generation:smolVLM2:multimodal:2:500M:f16:1.0.0:video-instruct";
+    readonly modelId: "mmproj-SmolVLM2-500M-Video-Instruct-f16.gguf";
+    readonly addon: "llm";
+    readonly expectedSize: 199470624;
+    readonly sha256Checksum: "b5dc8ebe7cbeab66a5369693960a52515d7824f13d4063ceca78431f2a6b59b0";
+}, {
+    readonly hyperdriveKey: "73b1bc01d01e25fa27be7d7f434337d14f054b0315e8463766ca31e778ac6576";
+    readonly hyperbeeKey: "generation:smolVLM2:multimodal:2:500M:q8_0:1.0.0:video-instruct";
+    readonly modelId: "SmolVLM2-500M-Video-Instruct-Q8_0.gguf";
+    readonly addon: "llm";
+    readonly expectedSize: 436808704;
+    readonly sha256Checksum: "6f67b8036b2469fcd71728702720c6b51aebd759b78137a8120733b4d66438bc";
+}, {
+    readonly hyperdriveKey: "73b1bc01d01e25fa27be7d7f434337d14f054b0315e8463766ca31e778ac6576";
+    readonly hyperbeeKey: "generation:smolVLM2:multimodal:2:500M:q8_0:1.0.0:video-instruct";
+    readonly modelId: "mmproj-SmolVLM2-500M-Video-Instruct-Q8_0.gguf";
+    readonly addon: "llm";
+    readonly expectedSize: 108785184;
+    readonly sha256Checksum: "921dc7e259f308e5b027111fa185efcbf33db13f6e35749ddf7f5cdb60ef520b";
 }, {
     readonly hyperdriveKey: "268c2e9b2a3420632e4b6649e32822f42d5dfbda4c7e96daec5b629ed20f99f7";
     readonly hyperbeeKey: "translation:ggml-indictrans:dist:2:200M:q0f32:1.0.0:en-hi";
@@ -421,11 +463,17 @@ export declare const GTE_LARGE_FP16: "pear://d1896d9259692818df95bd2480e90c2d057
 export declare const LASER_DOLPHIN_2_6_2X7B_INST_Q2_K: "pear://bd40e4bfae48b5bf436d1b0b447ae26e4c384aebe101a5e7df6ae71c919dadf1/dolphin-mixtral-2x7b-dop-Q2_K.gguf";
 export declare const LLAMA_3_2_1B_INST_Q4_0: "pear://afa79ee07c0a138bb9f11bfaee771fb1bdfca8c82d961cff0474e49827bd1de3/Llama-3.2-1B-Instruct-Q4_0.gguf";
 export declare const QWEN_3_1_7B_INST_Q4: "pear://05d3d7ad9cd650f53c28f85e312ef09a645dd487845897958b3be8a19cb3aab9/Qwen3-1.7B-Q4_0.gguf";
+export declare const QWEN2_5_OMNI_3B_MULTIMODAL: "pear://583f04c31d151b29e02aaf98f80d16aff585714ef9e2806a5067651c0f64ae31/Qwen2.5-Omni-3B-Q4_K_M.gguf";
+export declare const MMPROJ_QWEN2_5_OMNI_3B_MULTIMODAL: "pear://583f04c31d151b29e02aaf98f80d16aff585714ef9e2806a5067651c0f64ae31/mmproj-Qwen2.5-Omni-3B-Q8_0.gguf";
 export declare const QWEN3_0_6B_INST: "pear://d331ed49c444f90b3b2a70aa5b820752685142bf2d769a0592ed666140477578/Qwen3-0.6B-UD-IQ1_S.gguf";
 export declare const QWEN3_4B_INST_Q4: "pear://19ffb75463149955ee24d786dfddd84d41fc872ea813cd4465f5f7299d165adc/model.gguf";
 export declare const QWEN3_600M_INST_Q4: "pear://211874c9885f6b88b9926904420e365f5e74e1b6ac47207b7536408539bef4b7/Qwen3-0.6B-Q4_0.gguf";
 export declare const SALAMANDRATA_2B_INST_Q4: "pear://1610d81772a9e7c37660666dbdfdcef915b6b83c522ea1ad31c19cab0075811d/salamandrata_2b_inst_q4.gguf";
 export declare const SALAMANDRATA_2B_INST_Q8: "pear://96860337b0bdffdbb8ef0df4c8b2c3ab5e78568f5f7e15815a0a6b392512c9b5/salamandrata_2b_inst_q8.gguf";
+export declare const SMOLVLM2_2_500M_MULTIMODAL_F16: "pear://b99c6f5dcadaad827cdcb1a129a3b34aff6ad28056738e0bd559c7ffb807e30f/SmolVLM2-500M-Video-Instruct-f16.gguf";
+export declare const MMPROJ_SMOLVLM2_2_500M_MULTIMODAL_F16: "pear://b99c6f5dcadaad827cdcb1a129a3b34aff6ad28056738e0bd559c7ffb807e30f/mmproj-SmolVLM2-500M-Video-Instruct-f16.gguf";
+export declare const SMOLVLM2_2_500M_MULTIMODAL_Q8_0: "pear://73b1bc01d01e25fa27be7d7f434337d14f054b0315e8463766ca31e778ac6576/SmolVLM2-500M-Video-Instruct-Q8_0.gguf";
+export declare const MMPROJ_SMOLVLM2_2_500M_MULTIMODAL_Q8_0: "pear://73b1bc01d01e25fa27be7d7f434337d14f054b0315e8463766ca31e778ac6576/mmproj-SmolVLM2-500M-Video-Instruct-Q8_0.gguf";
 export declare const MARIAN_EN_HI_INDIC_200M_Q0F32: "pear://268c2e9b2a3420632e4b6649e32822f42d5dfbda4c7e96daec5b629ed20f99f7/ggml-indictrans2-en-indic-dist-200M.bin";
 export declare const MARIAN_HI_EN_INDIC_200M_Q0F32: "pear://dfea395ffe1cb6d259029e5c81c290f8828981987855e23e7b66fd7c705901eb/ggml-indictrans2-indic-en-dist-200M.bin";
 export declare const MARIAN_HI_HI_INDIC_320M_Q0F32: "pear://9a66da960931c449b41e3315d9aea4d3925f2cda024e8a4c139b64d0a539ed41/ggml-indictrans2-indic-indic-dist-320M.bin";
@@ -471,5 +519,5 @@ export declare const WHISPER_LARGE_3: "pear://76fb36d7f16a430618962d3e78b0e22e61
 export declare const WHISPER_SMALL: "pear://c6a987fbb4ad8fba1526e446501f88ee45dfafbf1685410a7aeadb589fe18608/ggml-small.en-q8_0.bin";
 export declare const WHISPER_TINY: "pear://9281d61503593fdf3eaa9d5e5d44746c271ef930aed3b7f8ee9ce153c76399d6/ggml-tiny.bin";
 export declare const WHISPER_TINY_SILERO: "pear://ebfb94b378276da139554668f1ff737644eadff529c2ea0f2662d7df61fd86ca/ggml-tiny.bin";
-export declare function getModelBySrc(modelId: string, hyperdriveKey?: string): HyperdriveItem | undefined;
+export declare function getModelBySrc(modelId: string, hyperdriveKey: string): HyperdriveItem | undefined;
 //# sourceMappingURL=models.d.ts.map

package/dist/models/hyperdrive/models.js

@@ -65,6 +65,22 @@ export const models = [
         expectedSize: 1056782912,
         sha256Checksum: "c876f159707a4e4f70e045106c69db15bfc935a4981706fd4f65c6e7ea1e81c5",
     },
+    {
+        hyperdriveKey: "583f04c31d151b29e02aaf98f80d16aff585714ef9e2806a5067651c0f64ae31",
+        hyperbeeKey: "generation:qwen2.5-omni:multimodal::3B::1.0.0",
+        modelId: "Qwen2.5-Omni-3B-Q4_K_M.gguf",
+        addon: "llm",
+        expectedSize: 2104931648,
+        sha256Checksum: "4b0bd358c1e9ec55dd3055ef6d71c958c821533d85916a10cfa89c4552a86e29",
+    },
+    {
+        hyperdriveKey: "583f04c31d151b29e02aaf98f80d16aff585714ef9e2806a5067651c0f64ae31",
+        hyperbeeKey: "generation:qwen2.5-omni:multimodal::3B::1.0.0",
+        modelId: "mmproj-Qwen2.5-Omni-3B-Q8_0.gguf",
+        addon: "llm",
+        expectedSize: 1538031328,
+        sha256Checksum: "4e6c816cd33f7298d07cb780c136a396631e50e62f6501660271f8c6e302e565",
+    },
     {
         hyperdriveKey: "d331ed49c444f90b3b2a70aa5b820752685142bf2d769a0592ed666140477578",
         hyperbeeKey: "generation:qwen3:instruct::0.6B::1.0.0",
@@ -105,6 +121,38 @@ export const models = [
         expectedSize: 2401075552,
         sha256Checksum: "9cfebb63cd7e6b6d1755ea8ef001fcdadfa17831da1db96eafe3fc554913d5e1",
     },
+    {
+        hyperdriveKey: "b99c6f5dcadaad827cdcb1a129a3b34aff6ad28056738e0bd559c7ffb807e30f",
+        hyperbeeKey: "generation:smolVLM2:multimodal:2:500M:f16:1.0.0:video-instruct",
+        modelId: "SmolVLM2-500M-Video-Instruct-f16.gguf",
+        addon: "llm",
+        expectedSize: 820424704,
+        sha256Checksum: "80f7e3f04bc2d3324ac1a9f52f5776fe13a69912adf74f8e7edacf773d140d77",
+    },
+    {
+        hyperdriveKey: "b99c6f5dcadaad827cdcb1a129a3b34aff6ad28056738e0bd559c7ffb807e30f",
+        hyperbeeKey: "generation:smolVLM2:multimodal:2:500M:f16:1.0.0:video-instruct",
+        modelId: "mmproj-SmolVLM2-500M-Video-Instruct-f16.gguf",
+        addon: "llm",
+        expectedSize: 199470624,
+        sha256Checksum: "b5dc8ebe7cbeab66a5369693960a52515d7824f13d4063ceca78431f2a6b59b0",
+    },
+    {
+        hyperdriveKey: "73b1bc01d01e25fa27be7d7f434337d14f054b0315e8463766ca31e778ac6576",
+        hyperbeeKey: "generation:smolVLM2:multimodal:2:500M:q8_0:1.0.0:video-instruct",
+        modelId: "SmolVLM2-500M-Video-Instruct-Q8_0.gguf",
+        addon: "llm",
+        expectedSize: 436808704,
+        sha256Checksum: "6f67b8036b2469fcd71728702720c6b51aebd759b78137a8120733b4d66438bc",
+    },
+    {
+        hyperdriveKey: "73b1bc01d01e25fa27be7d7f434337d14f054b0315e8463766ca31e778ac6576",
+        hyperbeeKey: "generation:smolVLM2:multimodal:2:500M:q8_0:1.0.0:video-instruct",
+        modelId: "mmproj-SmolVLM2-500M-Video-Instruct-Q8_0.gguf",
+        addon: "llm",
+        expectedSize: 108785184,
+        sha256Checksum: "921dc7e259f308e5b027111fa185efcbf33db13f6e35749ddf7f5cdb60ef520b",
+    },
     {
         hyperdriveKey: "268c2e9b2a3420632e4b6649e32822f42d5dfbda4c7e96daec5b629ed20f99f7",
         hyperbeeKey: "translation:ggml-indictrans:dist:2:200M:q0f32:1.0.0:en-hi",
@@ -475,60 +523,63 @@ export const GTE_LARGE_FP16 = `pear://${models[4].hyperdriveKey}/${models[4].mod
 export const LASER_DOLPHIN_2_6_2X7B_INST_Q2_K = `pear://${models[5].hyperdriveKey}/${models[5].modelId}`;
 export const LLAMA_3_2_1B_INST_Q4_0 = `pear://${models[6].hyperdriveKey}/${models[6].modelId}`;
 export const QWEN_3_1_7B_INST_Q4 = `pear://${models[7].hyperdriveKey}/${models[7].modelId}`;
-export const QWEN3_0_6B_INST = `pear://${models[8].hyperdriveKey}/${models[8].modelId}`;
-export const QWEN3_4B_INST_Q4 = `pear://${models[9].hyperdriveKey}/${models[9].modelId}`;
-export const QWEN3_600M_INST_Q4 = `pear://${models[10].hyperdriveKey}/${models[10].modelId}`;
-export const SALAMANDRATA_2B_INST_Q4 = `pear://${models[11].hyperdriveKey}/${models[11].modelId}`;
-export const SALAMANDRATA_2B_INST_Q8 = `pear://${models[12].hyperdriveKey}/${models[12].modelId}`;
-export const MARIAN_EN_HI_INDIC_200M_Q0F32 = `pear://${models[13].hyperdriveKey}/${models[13].modelId}`;
-export const MARIAN_HI_EN_INDIC_200M_Q0F32 = `pear://${models[14].hyperdriveKey}/${models[14].modelId}`;
-export const MARIAN_HI_HI_INDIC_320M_Q0F32 = `pear://${models[15].hyperdriveKey}/${models[15].modelId}`;
-export const MARIAN_EN_HI_INDIC_1B_Q0F16 = `pear://${models[16].hyperdriveKey}/${models[16].modelId}`;
-export const MARIAN_HI_EN_INDIC_1B_Q0F16 = `pear://${models[17].hyperdriveKey}/${models[17].modelId}`;
-export const MARIAN_HI_HI_INDIC_1B_Q0F16 = `pear://${models[18].hyperdriveKey}/${models[18].modelId}`;
-export const MARIAN_EN_HI_INDIC_1B_Q0F32 = `pear://${models[19].hyperdriveKey}/${models[19].modelId}`;
-export const MARIAN_HI_EN_INDIC_1B_Q0F32 = `pear://${models[20].hyperdriveKey}/${models[20].modelId}`;
-export const MARIAN_HI_HI_INDIC_1B_Q0F32 = `pear://${models[21].hyperdriveKey}/${models[21].modelId}`;
-export const MARIAN_EN_HI_INDIC_1B_Q4_0 = `pear://${models[22].hyperdriveKey}/${models[22].modelId}`;
-export const MARIAN_HI_EN_INDIC_1B_Q4_0 = `pear://${models[23].hyperdriveKey}/${models[23].modelId}`;
-export const MARIAN_HI_HI_INDIC_1B_Q4_0 = `pear://${models[24].hyperdriveKey}/${models[24].modelId}`;
-export const MARIAN_EN_HI_INDIC_200M_Q0F16 = `pear://${models[25].hyperdriveKey}/${models[25].modelId}`;
-export const MARIAN_HI_EN_INDIC_200M_Q0F16 = `pear://${models[26].hyperdriveKey}/${models[26].modelId}`;
-export const MARIAN_EN_HI_INDIC_200M_Q4_0 = `pear://${models[27].hyperdriveKey}/${models[27].modelId}`;
-export const MARIAN_HI_EN_INDIC_200M_Q4_0 = `pear://${models[28].hyperdriveKey}/${models[28].modelId}`;
-export const MARIAN_HI_HI_INDIC_320M_Q0F16 = `pear://${models[29].hyperdriveKey}/${models[29].modelId}`;
-export const MARIAN_HI_HI_INDIC_320M_Q4_0 = `pear://${models[30].hyperdriveKey}/${models[30].modelId}`;
-export const MARIAN_OPUS_EN_IT = `pear://${models[31].hyperdriveKey}/${models[31].modelId}`;
-export const MARIAN_OPUS_IT_EN = `pear://${models[32].hyperdriveKey}/${models[32].modelId}`;
-export const MARIAN_OPUS_DE_EN_Q0F32 = `pear://${models[33].hyperdriveKey}/${models[33].modelId}`;
-export const MARIAN_OPUS_DE_ES_Q0F32 = `pear://${models[34].hyperdriveKey}/${models[34].modelId}`;
-export const MARIAN_OPUS_DE_IT_Q0F32 = `pear://${models[35].hyperdriveKey}/${models[35].modelId}`;
-export const MARIAN_OPUS_EN_DE_Q0F32 = `pear://${models[36].hyperdriveKey}/${models[36].modelId}`;
-export const MARIAN_OPUS_EN_ES_Q0F32 = `pear://${models[37].hyperdriveKey}/${models[37].modelId}`;
-export const MARIAN_OPUS_EN_IT_Q0F32 = `pear://${models[38].hyperdriveKey}/${models[38].modelId}`;
-export const MARIAN_OPUS_ES_DE_Q0F32 = `pear://${models[39].hyperdriveKey}/${models[39].modelId}`;
-export const MARIAN_OPUS_ES_EN_Q0F32 = `pear://${models[40].hyperdriveKey}/${models[40].modelId}`;
-export const MARIAN_OPUS_ES_IT_Q0F32 = `pear://${models[41].hyperdriveKey}/${models[41].modelId}`;
-export const MARIAN_OPUS_IT_DE_Q0F32 = `pear://${models[42].hyperdriveKey}/${models[42].modelId}`;
-export const MARIAN_OPUS_IT_EN_Q0F32 = `pear://${models[43].hyperdriveKey}/${models[43].modelId}`;
-export const MARIAN_OPUS_IT_ES_Q0F32 = `pear://${models[44].hyperdriveKey}/${models[44].modelId}`;
-export const MARIAN_OPUS_EN_RU_Q0F16 = `pear://${models[45].hyperdriveKey}/${models[45].modelId}`;
-export const MARIAN_OPUS_RU_EN_Q0F16 = `pear://${models[46].hyperdriveKey}/${models[46].modelId}`;
-export const MARIAN_OPUS_EN_RU_Q0F32 = `pear://${models[47].hyperdriveKey}/${models[47].modelId}`;
-export const MARIAN_OPUS_RU_EN_Q0F32 = `pear://${models[48].hyperdriveKey}/${models[48].modelId}`;
-export const MARIAN_OPUS_EN_RU_Q4_0 = `pear://${models[49].hyperdriveKey}/${models[49].modelId}`;
-export const MARIAN_OPUS_RU_EN_Q4_0 = `pear://${models[50].hyperdriveKey}/${models[50].modelId}`;
-export const MARIAN_OPUS_EN_JA_Q4F16_1 = `pear://${models[51].hyperdriveKey}/${models[51].modelId}`;
-export const MARIAN_OPUS_JA_EN_Q4F16_1 = `pear://${models[52].hyperdriveKey}/${models[52].modelId}`;
-export const VAD_SILERO_5_1_2 = `pear://${models[53].hyperdriveKey}/${models[53].modelId}`;
-export const WHISPER_LARGE_3 = `pear://${models[54].hyperdriveKey}/${models[54].modelId}`;
-export const WHISPER_SMALL = `pear://${models[55].hyperdriveKey}/${models[55].modelId}`;
-export const WHISPER_TINY = `pear://${models[56].hyperdriveKey}/${models[56].modelId}`;
-export const WHISPER_TINY_SILERO = `pear://${models[57].hyperdriveKey}/${models[57].modelId}`;
+export const QWEN2_5_OMNI_3B_MULTIMODAL = `pear://${models[8].hyperdriveKey}/${models[8].modelId}`;
+export const MMPROJ_QWEN2_5_OMNI_3B_MULTIMODAL = `pear://${models[9].hyperdriveKey}/${models[9].modelId}`;
+export const QWEN3_0_6B_INST = `pear://${models[10].hyperdriveKey}/${models[10].modelId}`;
+export const QWEN3_4B_INST_Q4 = `pear://${models[11].hyperdriveKey}/${models[11].modelId}`;
+export const QWEN3_600M_INST_Q4 = `pear://${models[12].hyperdriveKey}/${models[12].modelId}`;
+export const SALAMANDRATA_2B_INST_Q4 = `pear://${models[13].hyperdriveKey}/${models[13].modelId}`;
+export const SALAMANDRATA_2B_INST_Q8 = `pear://${models[14].hyperdriveKey}/${models[14].modelId}`;
+export const SMOLVLM2_2_500M_MULTIMODAL_F16 = `pear://${models[15].hyperdriveKey}/${models[15].modelId}`;
+export const MMPROJ_SMOLVLM2_2_500M_MULTIMODAL_F16 = `pear://${models[16].hyperdriveKey}/${models[16].modelId}`;
+export const SMOLVLM2_2_500M_MULTIMODAL_Q8_0 = `pear://${models[17].hyperdriveKey}/${models[17].modelId}`;
+export const MMPROJ_SMOLVLM2_2_500M_MULTIMODAL_Q8_0 = `pear://${models[18].hyperdriveKey}/${models[18].modelId}`;
+export const MARIAN_EN_HI_INDIC_200M_Q0F32 = `pear://${models[19].hyperdriveKey}/${models[19].modelId}`;
+export const MARIAN_HI_EN_INDIC_200M_Q0F32 = `pear://${models[20].hyperdriveKey}/${models[20].modelId}`;
+export const MARIAN_HI_HI_INDIC_320M_Q0F32 = `pear://${models[21].hyperdriveKey}/${models[21].modelId}`;
+export const MARIAN_EN_HI_INDIC_1B_Q0F16 = `pear://${models[22].hyperdriveKey}/${models[22].modelId}`;
+export const MARIAN_HI_EN_INDIC_1B_Q0F16 = `pear://${models[23].hyperdriveKey}/${models[23].modelId}`;
+export const MARIAN_HI_HI_INDIC_1B_Q0F16 = `pear://${models[24].hyperdriveKey}/${models[24].modelId}`;
+export const MARIAN_EN_HI_INDIC_1B_Q0F32 = `pear://${models[25].hyperdriveKey}/${models[25].modelId}`;
+export const MARIAN_HI_EN_INDIC_1B_Q0F32 = `pear://${models[26].hyperdriveKey}/${models[26].modelId}`;
+export const MARIAN_HI_HI_INDIC_1B_Q0F32 = `pear://${models[27].hyperdriveKey}/${models[27].modelId}`;
+export const MARIAN_EN_HI_INDIC_1B_Q4_0 = `pear://${models[28].hyperdriveKey}/${models[28].modelId}`;
+export const MARIAN_HI_EN_INDIC_1B_Q4_0 = `pear://${models[29].hyperdriveKey}/${models[29].modelId}`;
+export const MARIAN_HI_HI_INDIC_1B_Q4_0 = `pear://${models[30].hyperdriveKey}/${models[30].modelId}`;
+export const MARIAN_EN_HI_INDIC_200M_Q0F16 = `pear://${models[31].hyperdriveKey}/${models[31].modelId}`;
+export const MARIAN_HI_EN_INDIC_200M_Q0F16 = `pear://${models[32].hyperdriveKey}/${models[32].modelId}`;
+export const MARIAN_EN_HI_INDIC_200M_Q4_0 = `pear://${models[33].hyperdriveKey}/${models[33].modelId}`;
+export const MARIAN_HI_EN_INDIC_200M_Q4_0 = `pear://${models[34].hyperdriveKey}/${models[34].modelId}`;
+export const MARIAN_HI_HI_INDIC_320M_Q0F16 = `pear://${models[35].hyperdriveKey}/${models[35].modelId}`;
+export const MARIAN_HI_HI_INDIC_320M_Q4_0 = `pear://${models[36].hyperdriveKey}/${models[36].modelId}`;
+export const MARIAN_OPUS_EN_IT = `pear://${models[37].hyperdriveKey}/${models[37].modelId}`;
+export const MARIAN_OPUS_IT_EN = `pear://${models[38].hyperdriveKey}/${models[38].modelId}`;
+export const MARIAN_OPUS_DE_EN_Q0F32 = `pear://${models[39].hyperdriveKey}/${models[39].modelId}`;
+export const MARIAN_OPUS_DE_ES_Q0F32 = `pear://${models[40].hyperdriveKey}/${models[40].modelId}`;
+export const MARIAN_OPUS_DE_IT_Q0F32 = `pear://${models[41].hyperdriveKey}/${models[41].modelId}`;
+export const MARIAN_OPUS_EN_DE_Q0F32 = `pear://${models[42].hyperdriveKey}/${models[42].modelId}`;
+export const MARIAN_OPUS_EN_ES_Q0F32 = `pear://${models[43].hyperdriveKey}/${models[43].modelId}`;
+export const MARIAN_OPUS_EN_IT_Q0F32 = `pear://${models[44].hyperdriveKey}/${models[44].modelId}`;
+export const MARIAN_OPUS_ES_DE_Q0F32 = `pear://${models[45].hyperdriveKey}/${models[45].modelId}`;
+export const MARIAN_OPUS_ES_EN_Q0F32 = `pear://${models[46].hyperdriveKey}/${models[46].modelId}`;
+export const MARIAN_OPUS_ES_IT_Q0F32 = `pear://${models[47].hyperdriveKey}/${models[47].modelId}`;
+export const MARIAN_OPUS_IT_DE_Q0F32 = `pear://${models[48].hyperdriveKey}/${models[48].modelId}`;
+export const MARIAN_OPUS_IT_EN_Q0F32 = `pear://${models[49].hyperdriveKey}/${models[49].modelId}`;
+export const MARIAN_OPUS_IT_ES_Q0F32 = `pear://${models[50].hyperdriveKey}/${models[50].modelId}`;
+export const MARIAN_OPUS_EN_RU_Q0F16 = `pear://${models[51].hyperdriveKey}/${models[51].modelId}`;
+export const MARIAN_OPUS_RU_EN_Q0F16 = `pear://${models[52].hyperdriveKey}/${models[52].modelId}`;
+export const MARIAN_OPUS_EN_RU_Q0F32 = `pear://${models[53].hyperdriveKey}/${models[53].modelId}`;
+export const MARIAN_OPUS_RU_EN_Q0F32 = `pear://${models[54].hyperdriveKey}/${models[54].modelId}`;
+export const MARIAN_OPUS_EN_RU_Q4_0 = `pear://${models[55].hyperdriveKey}/${models[55].modelId}`;
+export const MARIAN_OPUS_RU_EN_Q4_0 = `pear://${models[56].hyperdriveKey}/${models[56].modelId}`;
+export const MARIAN_OPUS_EN_JA_Q4F16_1 = `pear://${models[57].hyperdriveKey}/${models[57].modelId}`;
+export const MARIAN_OPUS_JA_EN_Q4F16_1 = `pear://${models[58].hyperdriveKey}/${models[58].modelId}`;
+export const VAD_SILERO_5_1_2 = `pear://${models[59].hyperdriveKey}/${models[59].modelId}`;
+export const WHISPER_LARGE_3 = `pear://${models[60].hyperdriveKey}/${models[60].modelId}`;
+export const WHISPER_SMALL = `pear://${models[61].hyperdriveKey}/${models[61].modelId}`;
+export const WHISPER_TINY = `pear://${models[62].hyperdriveKey}/${models[62].modelId}`;
+export const WHISPER_TINY_SILERO = `pear://${models[63].hyperdriveKey}/${models[63].modelId}`;
 // Helper function for our curated model list
 export function getModelBySrc(modelId, hyperdriveKey) {
-    if (hyperdriveKey) {
-        return models.find((model) => model.modelId === modelId && model.hyperdriveKey === hyperdriveKey);
-    }
-    return models.find((model) => model.modelId === modelId);
+    return models.find((model) => model.modelId === modelId && model.hyperdriveKey === hyperdriveKey);
 }

package/dist/schemas/common.d.ts

@@ -32,6 +32,7 @@ export declare const requestSchema: z.ZodUnion<readonly [z.ZodObject<{
         frequency_penalty: z.ZodOptional<z.ZodNumber>;
         repeat_penalty: z.ZodOptional<z.ZodNumber>;
         reverse_prompt: z.ZodOptional<z.ZodString>;
+        n_discarded: z.ZodOptional<z.ZodNumber>;
     }, z.core.$strip>;
 }, z.core.$strip>, z.ZodObject<{
     type: z.ZodLiteral<"loadModel">;
@@ -199,6 +200,7 @@ export declare const requestSchema: z.ZodUnion<readonly [z.ZodObject<{
         frequency_penalty?: number | undefined;
         repeat_penalty?: number | undefined;
         reverse_prompt?: string | undefined;
+        n_discarded?: number | undefined;
     };
     projectionModelSrc?: string | undefined;
     vadModelSrc?: string | undefined;
@@ -296,6 +298,7 @@ export declare const requestSchema: z.ZodUnion<readonly [z.ZodObject<{
         frequency_penalty?: number | undefined;
         repeat_penalty?: number | undefined;
         reverse_prompt?: string | undefined;
+        n_discarded?: number | undefined;
     };
     projectionModelSrc?: string | undefined;
     vadModelSrc?: string | undefined;
@@ -483,6 +486,7 @@ export declare const requestSchema: z.ZodUnion<readonly [z.ZodObject<{
             paragraph: "paragraph";
         }>>;
         splitStrategy: z.ZodOptional<z.ZodEnum<{
+            token: "token";
             character: "character";
             word: "word";
             sentence: "sentence";

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

@@ -21,6 +21,7 @@ export declare const llmConfigSchema: z.ZodObject<{
     frequency_penalty: z.ZodOptional<z.ZodNumber>;
     repeat_penalty: z.ZodOptional<z.ZodNumber>;
     reverse_prompt: z.ZodOptional<z.ZodString>;
+    n_discarded: z.ZodOptional<z.ZodNumber>;
 }, z.core.$strip>;
 export declare const embedConfigSchema: z.ZodObject<{
     config: z.ZodDefault<z.ZodString>;

package/dist/schemas/llamacpp-config.js

@@ -32,8 +32,11 @@ export const llmConfigSchema = z.object({
     frequency_penalty: z.number().optional(),
     repeat_penalty: z.number().optional(),
     reverse_prompt: z.string().optional(),
+    n_discarded: z.number().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\t25"),
+    config: z
+        .string()
+        .default("-ngl\t99\n-dev\tgpu\n--embd-separator\t⟪§§§EMBED_SEP§§§⟫\n-c\t512"),
 });

package/dist/schemas/load-model.d.ts

@@ -19,6 +19,7 @@ export declare const loadModelOptionsSchema: z.ZodPipe<z.ZodUnion<readonly [z.Zo
         frequency_penalty: z.ZodOptional<z.ZodOptional<z.ZodNumber>>;
         repeat_penalty: z.ZodOptional<z.ZodOptional<z.ZodNumber>>;
         reverse_prompt: z.ZodOptional<z.ZodOptional<z.ZodString>>;
+        n_discarded: z.ZodOptional<z.ZodOptional<z.ZodNumber>>;
     }, z.core.$strict>>;
     seed: z.ZodOptional<z.ZodBoolean>;
     projectionModelSrc: z.ZodOptional<z.ZodString>;
@@ -187,6 +188,7 @@ export declare const loadModelOptionsSchema: z.ZodPipe<z.ZodUnion<readonly [z.Zo
         frequency_penalty?: number | undefined;
         repeat_penalty?: number | undefined;
         reverse_prompt?: string | undefined;
+        n_discarded?: number | undefined;
     } | undefined;
     projectionModelSrc?: string | undefined;
     delegate?: {
@@ -271,6 +273,7 @@ export declare const loadModelOptionsSchema: z.ZodPipe<z.ZodUnion<readonly [z.Zo
         frequency_penalty?: number | undefined;
         repeat_penalty?: number | undefined;
         reverse_prompt?: string | undefined;
+        n_discarded?: number | undefined;
     } | undefined;
     seed?: boolean | undefined;
     projectionModelSrc?: string | undefined;
@@ -370,6 +373,7 @@ export declare const loadLlmModelRequestSchema: z.ZodObject<{
         frequency_penalty: z.ZodOptional<z.ZodNumber>;
         repeat_penalty: z.ZodOptional<z.ZodNumber>;
         reverse_prompt: z.ZodOptional<z.ZodString>;
+        n_discarded: z.ZodOptional<z.ZodNumber>;
     }, z.core.$strip>;
 }, z.core.$strip>;
 export declare const loadWhisperModelRequestSchema: z.ZodObject<{
@@ -550,6 +554,7 @@ export declare const loadModelRequestSchema: z.ZodPipe<z.ZodDiscriminatedUnion<[
         frequency_penalty: z.ZodOptional<z.ZodNumber>;
         repeat_penalty: z.ZodOptional<z.ZodNumber>;
         reverse_prompt: z.ZodOptional<z.ZodString>;
+        n_discarded: z.ZodOptional<z.ZodNumber>;
     }, z.core.$strip>;
 }, z.core.$strip>, z.ZodObject<{
     type: z.ZodLiteral<"loadModel">;
@@ -717,6 +722,7 @@ export declare const loadModelRequestSchema: z.ZodPipe<z.ZodDiscriminatedUnion<[
         frequency_penalty?: number | undefined;
         repeat_penalty?: number | undefined;
         reverse_prompt?: string | undefined;
+        n_discarded?: number | undefined;
     };
     projectionModelSrc?: string | undefined;
     vadModelSrc?: string | undefined;
@@ -814,6 +820,7 @@ export declare const loadModelRequestSchema: z.ZodPipe<z.ZodDiscriminatedUnion<[
         frequency_penalty?: number | undefined;
         repeat_penalty?: number | undefined;
         reverse_prompt?: string | undefined;
+        n_discarded?: number | undefined;
     };
     projectionModelSrc?: string | undefined;
     vadModelSrc?: string | undefined;
@@ -932,6 +939,7 @@ export declare const loadModelServerParamsSchema: z.ZodObject<{
             frequency_penalty: z.ZodOptional<z.ZodOptional<z.ZodNumber>>;
             repeat_penalty: z.ZodOptional<z.ZodOptional<z.ZodNumber>>;
             reverse_prompt: z.ZodOptional<z.ZodOptional<z.ZodString>>;
+            n_discarded: z.ZodOptional<z.ZodOptional<z.ZodNumber>>;
         }, z.core.$strict>>;
         seed: z.ZodOptional<z.ZodBoolean>;
         projectionModelSrc: z.ZodOptional<z.ZodString>;
@@ -1100,6 +1108,7 @@ export declare const loadModelServerParamsSchema: z.ZodObject<{
             frequency_penalty?: number | undefined;
             repeat_penalty?: number | undefined;
             reverse_prompt?: string | undefined;
+            n_discarded?: number | undefined;
         } | undefined;
         projectionModelSrc?: string | undefined;
         delegate?: {
@@ -1184,6 +1193,7 @@ export declare const loadModelServerParamsSchema: z.ZodObject<{
             frequency_penalty?: number | undefined;
             repeat_penalty?: number | undefined;
             reverse_prompt?: string | undefined;
+            n_discarded?: number | undefined;
         } | undefined;
         seed?: boolean | undefined;
         projectionModelSrc?: string | undefined;

package/dist/schemas/rag.d.ts

@@ -25,6 +25,7 @@ declare const ragSaveEmbeddingsParamsSchema: z.ZodObject<{
             paragraph: "paragraph";
         }>>;
         splitStrategy: z.ZodOptional<z.ZodEnum<{
+            token: "token";
             character: "character";
             word: "word";
             sentence: "sentence";
@@ -45,6 +46,7 @@ export declare const ragSaveEmbeddingsOperationSchema: z.ZodObject<{
             paragraph: "paragraph";
         }>>;
         splitStrategy: z.ZodOptional<z.ZodEnum<{
+            token: "token";
             character: "character";
             word: "word";
             sentence: "sentence";
@@ -95,6 +97,7 @@ export declare const ragRequestSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
             paragraph: "paragraph";
         }>>;
         splitStrategy: z.ZodOptional<z.ZodEnum<{
+            token: "token";
             character: "character";
             word: "word";
             sentence: "sentence";

package/dist/schemas/rag.js

@@ -4,7 +4,9 @@ const llmChunkOptsSchema = z.object({
     chunkSize: z.number().optional(),
     chunkOverlap: z.number().optional(),
     chunkStrategy: z.enum(["character", "paragraph"]).optional(),
-    splitStrategy: z.enum(["character", "word", "sentence", "line"]).optional(),
+    splitStrategy: z
+        .enum(["character", "word", "token", "sentence", "line"])
+        .optional(),
 });
 const ragSaveEmbeddingsResultSchema = z.object({
     status: z.enum(["fulfilled", "rejected"]),

package/dist/server/bare/addons/llamacpp/handlers/completion-stream.js

@@ -54,8 +54,34 @@ function prepareMessagesForCache(cachePathToUse, cacheExists, history) {
 }
 async function* processModelResponse(model, messagesToSend, shouldSaveCache) {
     const response = await model.run(messagesToSend);
-    for await (const token of response.iterate()) {
-        yield token;
+    // Create a promise that rejects when onError is triggered
+    let rejectOnError = null;
+    const errorPromise = new Promise((_, reject) => {
+        rejectOnError = reject;
+    });
+    response.onError((error) => {
+        if (rejectOnError) {
+            rejectOnError(error);
+        }
+    });
+    const iterator = response.iterate();
+    try {
+        while (true) {
+            // Race between getting the next token and an error occurring
+            const result = await Promise.race([iterator.next(), errorPromise]);
+            if (result.done) {
+                break;
+            }
+            yield result.value;
+        }
+    }
+    catch (error) {
+        if (error instanceof Error) {
+            throw error;
+        }
+        else {
+            throw new Error(String(error));
+        }
     }
     if (shouldSaveCache) {
         await model.run([{ role: "session", content: "save" }]);

package/dist/server/bare/registry/model-registry.d.ts

@@ -8,6 +8,7 @@ export interface AnyModel {
         onUpdate(callback: (data: string) => void): {
             await: () => Promise<void>;
         };
+        onError(callback: (error: Error) => void): void;
         onUpdate(callback: (data: Array<{
             text: string;
         }>) => void): {

package/dist/server/rpc/handlers/completion-stream.js

@@ -1,52 +1,38 @@
 import { completion } from "../../bare/addons/llamacpp/index.js";
 export async function* handleCompletionStream(request) {
-    try {
-        // Log tools if provided
-        if (request.tools) {
-            console.log("Tools passed to completionStream:", request.tools);
+    const filteredHistory = request.history.map(({ role, content, attachments }) => ({
+        role,
+        content,
+        attachments: attachments ?? [],
+    }));
+    const stream = completion({
+        history: filteredHistory,
+        modelId: request.modelId,
+        kvCache: request.kvCache,
+    });
+    let stats;
+    let done = false;
+    let buffer = "";
+    while (!done) {
+        const result = await stream.next();
+        if (result.done) {
+            stats = result.value;
+            done = true;
         }
-        const filteredHistory = request.history.map(({ role, content, attachments }) => ({
-            role,
-            content,
-            attachments: attachments ?? [],
-        }));
-        const stream = completion({
-            history: filteredHistory,
-            modelId: request.modelId,
-            kvCache: request.kvCache,
-        });
-        let stats;
-        let done = false;
-        let buffer = "";
-        while (!done) {
-            const result = await stream.next();
-            if (result.done) {
-                stats = result.value;
-                done = true;
-            }
-            else {
-                buffer += result.value;
-                if (request.stream) {
-                    yield {
-                        type: "completionStream",
-                        token: result.value,
-                    };
-                }
+        else {
+            buffer += result.value;
+            if (request.stream) {
+                yield {
+                    type: "completionStream",
+                    token: result.value,
+                };
             }
         }
-        yield {
-            type: "completionStream",
-            token: buffer,
-            done: true,
-            stats,
-        };
-    }
-    catch (error) {
-        console.error("Error during completion stream:", error);
-        yield {
-            type: "completionStream",
-            token: "",
-            error: error instanceof Error ? error.message : "Unknown error",
-        };
     }
+    yield {
+        type: "completionStream",
+        token: buffer,
+        done: true,
+        stats,
+    };
 }

package/dist/server/rpc/handlers/index.d.ts

@@ -9,9 +9,11 @@ import { handleTranslate } from "./translate";
 import { handleLoggingStream } from "./logging-stream";
 import { handleRag } from "./rag";
 import { cancelHandler } from "./cancelHandler";
+import { handleDeleteCache } from "./delete-cache";
 export declare const handlers: {
     completionStream: typeof handleCompletionStream;
     downloadAsset: typeof handleDownloadAsset;
+    deleteCache: typeof handleDeleteCache;
     loadModel: typeof handleLoadModel;
     unloadModel: typeof handleUnloadModel;
     embed: typeof handleEmbed;

package/dist/server/rpc/handlers/index.js

@@ -9,9 +9,11 @@ import { handleTranslate } from "./translate.js";
 import { handleLoggingStream } from "./logging-stream.js";
 import { handleRag } from "./rag.js";
 import { cancelHandler } from "./cancelHandler.js";
+import { handleDeleteCache } from "./delete-cache.js";
 export const handlers = {
     completionStream: handleCompletionStream,
     downloadAsset: handleDownloadAsset,
+    deleteCache: handleDeleteCache,
     loadModel: handleLoadModel,
     unloadModel: handleUnloadModel,
     embed: handleEmbed,

package/dist/server/rpc/handlers/load-model/download-manager.d.ts

@@ -4,28 +4,9 @@ export declare const registerDownload: (key: string, entry: DownloadEntry) => vo
 export declare const unregisterDownload: (key: string) => void;
 export declare const createHyperdriveDownloadKey: (hyperdriveKey: string, modelFileName: string) => string;
 export declare const createHttpDownloadKey: (url: string) => string;
-export declare const getAllActiveDownloads: () => Map<string, DownloadEntry>;
 export declare const setClearCacheFlag: (downloadKey: string, clearCache: boolean) => void;
 export declare const shouldClearCache: (downloadKey: string) => boolean;
 export declare const clearClearCacheFlag: (downloadKey: string) => void;
 export declare const createCancelFunction: (downloadKey: string, clearCache?: boolean) => () => void;
 export declare const cancelAllDownloads: () => void;
-export declare const getDownloadStats: () => {
-    activeCount: number;
-    downloads: ({
-        hyperdriveKey: string;
-        modelFileName: string;
-        key: string;
-        startTime: number;
-        duration: number;
-        type: "hyperdrive" | "http";
-    } | {
-        url: string;
-        modelPath: string;
-        key: string;
-        startTime: number;
-        duration: number;
-        type: "hyperdrive" | "http";
-    })[];
-};
 //# sourceMappingURL=download-manager.d.ts.map

package/dist/server/rpc/handlers/load-model/download-manager.js

@@ -12,7 +12,6 @@ export const unregisterDownload = (key) => {
 };
 export const createHyperdriveDownloadKey = (hyperdriveKey, modelFileName) => `${hyperdriveKey}:${modelFileName}`;
 export const createHttpDownloadKey = (url) => `http:${url}`;
-export const getAllActiveDownloads = () => new Map(activeDownloads);
 export const setClearCacheFlag = (downloadKey, clearCache) => {
     if (clearCache) {
         clearCacheFlags.set(downloadKey, true);
@@ -39,24 +38,6 @@ export const cancelAllDownloads = () => {
     console.log(`🧹 Cancelling ${activeDownloads.size} active downloads`);
     Array.from(activeDownloads.keys()).forEach((key) => createCancelFunction(key)());
 };
-export const getDownloadStats = () => ({
-    activeCount: activeDownloads.size,
-    downloads: Array.from(activeDownloads.entries()).map(([key, entry]) => ({
-        key,
-        startTime: entry.startTime,
-        duration: Date.now() - entry.startTime,
-        type: entry.type,
-        ...(entry.type === "hyperdrive"
-            ? {
-                hyperdriveKey: entry.hyperdriveKey,
-                modelFileName: entry.modelFileName,
-            }
-            : {
-                url: entry.url,
-                modelPath: entry.modelPath,
-            }),
-    })),
-});
 // Register cleanup handlers for graceful shutdown
 let cleanupRegistered = false;
 let isCleaningUp = false;
@@ -79,10 +60,14 @@ const registerCleanupHandlers = () => {
             console.error("❌ Error during download cleanup:", error);
         }
     };
-    process.on("SIGTERM", () => {
-        void cleanup();
-    });
-    process.on("SIGINT", () => {
-        void cleanup();
-    });
+    let handlingExit = false;
+    const exitHandler = async () => {
+        if (handlingExit)
+            return;
+        handlingExit = true;
+        await cleanup();
+        process.exit(0);
+    };
+    process.once("SIGTERM", () => void exitHandler());
+    process.once("SIGINT", () => void exitHandler());
 };

package/dist/server/rpc/handlers/load-model/hyperdrive.js

@@ -10,24 +10,49 @@ import { getEnv } from "../../../worker.js";
 import { getModelsCacheDir, generateShortHash } from "../../../utils/index.js";
 import { getModelBySrc } from "../../../../models/hyperdrive/index.js";
 import { getActiveDownload, registerDownload, unregisterDownload, createHyperdriveDownloadKey, shouldClearCache, clearClearCacheFlag, } from "./download-manager.js";
+function createProgressHandler(blob, progressCallback, downloadKey, initialBytes = 0) {
+    let downloadedBytes = initialBytes;
+    return (index, bytes) => {
+        if (index >= blob.blockOffset &&
+            index < blob.blockOffset + blob.blockLength) {
+            downloadedBytes += bytes;
+            if (progressCallback && downloadKey) {
+                progressCallback({
+                    type: "modelProgress",
+                    downloaded: downloadedBytes,
+                    total: blob.byteLength,
+                    percentage: blob.byteLength > 0 ? (downloadedBytes / blob.byteLength) * 100 : 0,
+                    downloadKey,
+                });
+            }
+        }
+    };
+}
 async function setupHyperdrive(hyperdriveKey, corestoreDir) {
     const corestore = new Corestore(corestoreDir);
     await corestore.ready();
     const drive = new Hyperdrive(corestore, Buffer.from(hyperdriveKey, "hex"));
     await drive.ready();
     const swarm = new Hyperswarm();
-    swarm.join(drive.discoveryKey, { server: false, client: true });
+    const doneFindingPeers = drive.findingPeers();
     swarm.on("connection", (connection) => {
         console.log(`🤝 Connected to peer for hyperdrive download`);
         drive.replicate(connection);
     });
-    const doneFindingPeers = drive.findingPeers();
-    await swarm.flush();
-    doneFindingPeers();
+    swarm.join(drive.discoveryKey, { server: false, client: true });
+    swarm.flush().then(doneFindingPeers, doneFindingPeers);
     return { corestore, drive, swarm, corestoreDir };
 }
 async function cleanupHyperdrive(setup) {
     console.log("🧹 Cleaning up hyperdrive setup");
+    if (!setup.swarm.suspended) {
+        try {
+            await setup.swarm.suspend();
+        }
+        catch {
+            // If suspend fails, continue with cleanup
+        }
+    }
     await setup.swarm.destroy();
     await setup.drive.close();
     await setup.corestore.close();
@@ -42,43 +67,81 @@ export async function deleteCorestoreDirectory(corestoreDir) {
     }
 }
 async function findFileInHyperdrive(drive, modelFileName) {
+    if (drive.core.length === 0) {
+        // this method will update the length of the core using the first peer it finds
+        // this will ensure that the hyperdrive is usable
+        await drive.core.update();
+    }
     const paths = [modelFileName, `/${modelFileName}`, `./${modelFileName}`];
-    let entry = null;
-    let targetPath = modelFileName;
     for (const pathVariation of paths) {
         try {
-            entry = await drive.entry(pathVariation);
+            const entry = await drive.entry(pathVariation, { wait: true });
             if (entry) {
-                targetPath = pathVariation;
-                break;
+                return { entry, targetPath: pathVariation };
             }
         }
         catch {
             // Continue trying other paths
         }
     }
-    if (!entry) {
-        // Try listing available files to help debug
-        try {
-            const rootFiles = drive.list("/", false);
-            const availableFiles = [];
-            for await (const file of rootFiles) {
+    // If not found, list available files for debugging
+    const availableFiles = [];
+    try {
+        // Try both with and without leading slash
+        for (const prefix of ["/", ""]) {
+            const files = drive.list(prefix, { recursive: true, wait: true });
+            for await (const file of files) {
                 availableFiles.push(file.key);
             }
-            if (availableFiles.length === 0) {
-                // Try without leading slash
-                const noSlashFiles = drive.list("", false);
-                for await (const file of noSlashFiles) {
-                    availableFiles.push(file.key);
-                }
-            }
-            throw new Error(`File ${modelFileName} not found in hyperdrive. Available files: ${availableFiles.join(", ")}`);
+            if (availableFiles.length > 0)
+                break;
         }
-        catch (listError) {
-            throw new Error(`File ${modelFileName} not found in hyperdrive. Could not list available files: ${listError instanceof Error ? listError.message : String(listError)}`);
+    }
+    catch {
+        // Ignore listing errors
+    }
+    const filesInfo = availableFiles.length > 0
+        ? `Available files: ${availableFiles.join(", ")}`
+        : "Could not list available files";
+    throw new Error(`File ${modelFileName} not found in hyperdrive. ${filesInfo}`);
+}
+async function checkFileDownloaded(drive, entry) {
+    const blobs = await drive.getBlobs();
+    const blob = entry?.value?.blob;
+    if (!blob) {
+        return {
+            blobs,
+            blocksDownloaded: 0,
+            totalBlocks: 0,
+            totalBytes: 0,
+        };
+    }
+    // Check how many blocks are already downloaded
+    let blocksDownloaded = 0;
+    for (let i = blob.blockOffset; i < blob.blockOffset + blob.blockLength; i++) {
+        const hasBlock = await blobs.core.has(i);
+        if (hasBlock) {
+            blocksDownloaded++;
         }
     }
-    return { entry, targetPath };
+    return {
+        blobs,
+        blocksDownloaded,
+        totalBlocks: blob.blockLength,
+        totalBytes: blob.byteLength,
+    };
+}
+async function validateFileExistence(drive, file) {
+    const { entry, targetPath } = await findFileInHyperdrive(drive, file);
+    const { blobs, blocksDownloaded, totalBlocks, totalBytes } = await checkFileDownloaded(drive, entry);
+    return {
+        entry,
+        blobs,
+        targetPath,
+        blocksDownloaded,
+        totalBlocks,
+        totalBytes,
+    };
 }
 async function validateFileChecksum(filePath, expectedChecksum) {
     console.log(`🔍 Validating SHA256 checksum...`);
@@ -153,13 +216,15 @@ async function performHyperdriveDownload(modelPath, modelFileName, hyperdriveKey
     }
     console.log(`Downloading model from hyperdrive ${hyperdriveKey} to ${modelPath}`);
     const setup = await setupHyperdrive(hyperdriveKey, corestoreDir);
-    let downloadedBytes = 0;
-    let totalBytes = 0;
     // Set up abort handler
     let readStream;
     let writeStream;
     const cleanup = () => {
         console.log("🛑 Aborting hyperdrive download...");
+        setup.swarm.suspended = true;
+        setup.swarm.suspend().catch(() => {
+            // Ignore errors during emergency suspend
+        });
         // Destroy streams if they exist
         if (readStream) {
             readStream.destroy();
@@ -169,48 +234,19 @@ async function performHyperdriveDownload(modelPath, modelFileName, hyperdriveKey
         }
     };
     signal?.addEventListener("abort", cleanup);
+    let downloadHandler;
     try {
-        const result = await findFileInHyperdrive(setup.drive, modelFileName);
-        const { entry, targetPath } = result;
-        totalBytes = entry?.value?.blob?.byteLength || 0;
-        console.log(`📏 Model size: ${totalBytes} bytes (${(totalBytes / 1024 / 1024).toFixed(2)} MB)`);
-        // Get blobs and file blob range
-        const blobs = await setup.drive.getBlobs();
+        const { entry, blobs, targetPath, blocksDownloaded, totalBlocks, totalBytes, } = await validateFileExistence(setup.drive, modelFileName);
         const blob = entry?.value?.blob;
         if (!blob) {
             throw new Error(`No blob found for ${modelFileName}`);
         }
-        // Check what blocks are already downloaded in corestore for this specific file
-        let alreadyDownloaded = 0;
-        for (let i = blob.blockOffset; i < blob.blockOffset + blob.blockLength; i++) {
-            const hasBlock = await blobs.core.has(i);
-            if (hasBlock) {
-                alreadyDownloaded++;
-            }
-        }
-        downloadedBytes = 0;
-        if (alreadyDownloaded > 0) {
-            console.log(`📥 Resuming download (${alreadyDownloaded}/${blob.blockLength} blocks already in corestore)`);
-        }
-        else {
-            console.log(`📥 Starting fresh download`);
-        }
+        // Calculate already downloaded bytes for resuming
+        const bytesPerBlock = totalBlocks > 0 ? totalBytes / totalBlocks : 0;
+        const initialBytes = Math.floor(blocksDownloaded * bytesPerBlock);
         // Track only downloads for blocks within our file's blob range
-        blobs.core.on("download", (index, bytes) => {
-            if (index >= blob.blockOffset &&
-                index < blob.blockOffset + blob.blockLength) {
-                downloadedBytes += bytes;
-                if (progressCallback) {
-                    progressCallback({
-                        type: "modelProgress",
-                        downloaded: downloadedBytes,
-                        total: totalBytes,
-                        percentage: totalBytes > 0 ? (downloadedBytes / totalBytes) * 100 : 0,
-                        downloadKey,
-                    });
-                }
-            }
-        });
+        downloadHandler = createProgressHandler(blob, progressCallback, downloadKey, initialBytes);
+        blobs.core.on("download", downloadHandler);
         // Check if aborted before starting streams
         if (signal?.aborted) {
             throw new Error("Download cancelled");
@@ -223,15 +259,13 @@ async function performHyperdriveDownload(modelPath, modelFileName, hyperdriveKey
         readStream.pipe(writeStream);
         // Wait for download to complete
         await new Promise((resolve, reject) => {
-            writeStream.on("finish", () => resolve(undefined));
+            writeStream.on("finish", resolve);
             writeStream.on("error", reject);
             readStream.on("error", reject);
             // Also reject on abort
-            if (signal) {
-                signal.addEventListener("abort", () => {
-                    reject(new Error("Download cancelled"));
-                }, { once: true });
-            }
+            signal?.addEventListener("abort", () => {
+                reject(new Error("Download cancelled"));
+            }, { once: true });
         });
         console.log(`✅ Model downloaded successfully to ${modelPath}`);
         // Delete corestore directory after successful download (only when not seeding, to save space)
@@ -249,22 +283,29 @@ async function performHyperdriveDownload(modelPath, modelFileName, hyperdriveKey
     }
     finally {
         signal?.removeEventListener("abort", cleanup);
+        if (downloadHandler) {
+            try {
+                const { blobs } = await validateFileExistence(setup.drive, modelFileName);
+                blobs.core.off("download", downloadHandler);
+            }
+            catch {
+                // If we can't get blobs, setup is already closed
+            }
+        }
         if (!seed) {
             await cleanupHyperdrive(setup);
         }
         // Only delete corestore and partial file if user explicitly requested clearCache
-        if (signal?.aborted) {
-            if (shouldClearCache(downloadKey)) {
-                try {
-                    await fsPromises.unlink(modelPath);
-                }
-                catch {
-                    // no-op
-                }
-                await deleteCorestoreDirectory(corestoreDir);
+        if (signal?.aborted && shouldClearCache(downloadKey)) {
+            try {
+                await fsPromises.unlink(modelPath);
+            }
+            catch {
+                // no-op
             }
-            clearClearCacheFlag(downloadKey);
+            await deleteCorestoreDirectory(corestoreDir);
         }
+        clearClearCacheFlag(downloadKey);
     }
 }
 export async function downloadModelFromHyperdrive(hyperdriveKey, modelFileName, seed, progressCallback, expectedChecksum) {

package/dist/server/rpc/handlers/load-model/resolve.js

@@ -1,46 +1,10 @@
-import { getModelBySrc, models } from "../../../../models/hyperdrive/models.js";
+import { models } from "../../../../models/hyperdrive/models.js";
 import { hyperdriveUrlSchema } from "../../../../schemas/index.js";
 import { getModelsCacheDir } from "../../../utils/index.js";
 import { promises as fsPromises } from "bare-fs";
 import path from "bare-path";
 import { downloadModelFromHttp } from "./http.js";
 import { downloadModelFromHyperdrive } from "./hyperdrive.js";
-/**
- * Attempts to resolve a model from cache (without validation)
- * Validation will be handled by the hyperdrive download logic
- */
-async function tryResolveCachedModel(modelId) {
-    const cacheDir = getModelsCacheDir();
-    const cachedPath = path.join(cacheDir, modelId);
-    try {
-        await fsPromises.access(cachedPath);
-        // File exists - return null to let hyperdrive validation handle size/checksum
-        // This ensures partial downloads are properly resumed instead of being
-        // incorrectly validated and deleted
-        return null;
-    }
-    catch {
-        // File doesn't exist
-        return null;
-    }
-}
-/**
- * Resolves a known model by ID from hyperdrive, checking cache first
- */
-async function resolveKnownModel(model, modelIdOrPath, progressCallback, seed) {
-    if (!model) {
-        throw new Error("Model is null");
-    }
-    console.log(`Loading model ID '${modelIdOrPath}' from hyperdrive`);
-    // Try to resolve from cache first
-    const cachedPath = await tryResolveCachedModel(model.modelId);
-    if (cachedPath) {
-        return cachedPath;
-    }
-    // Not cached or invalid, download from hyperdrive
-    const downloadedPath = await downloadModelFromHyperdrive(model.hyperdriveKey, model.modelId, seed, progressCallback, model.sha256Checksum);
-    return downloadedPath;
-}
 /**
  * Resolves a local file path or cached file
  */
@@ -84,10 +48,7 @@ export async function resolveModelPath(modelSrc, progressCallback, seed) {
     let actualPath;
     // Validate seeding is only used with hyperdrive models
     if (seed && !hyperdriveKey) {
-        const model = getModelBySrc(actualModelSrc);
-        if (!model) {
-            throw new Error("Seeding is only supported for hyperdrive models");
-        }
+        throw new Error("Seeding is only supported for hyperdrive models");
     }
     // Check if it's an HTTP/HTTPS URL
     if (actualModelSrc.startsWith("http://") ||
@@ -101,21 +62,13 @@ export async function resolveModelPath(modelSrc, progressCallback, seed) {
         console.log(`Loading from hyperdrive: ${hyperdriveKey}`);
         actualPath = await downloadModelFromHyperdrive(hyperdriveKey, actualModelSrc, seed, progressCallback);
     }
+    else if (actualModelSrc.includes("/") || actualModelSrc.includes("\\")) {
+        // Handle file paths (absolute or relative with slashes)
+        actualPath = actualModelSrc;
+    }
     else {
-        // Try to resolve as a known model ID
-        const model = getModelBySrc(actualModelSrc);
-        if (model) {
-            actualPath = await resolveKnownModel(model, actualModelSrc, progressCallback, seed);
-        }
-        else if (actualModelSrc.includes("/") || actualModelSrc.includes("\\")) {
-            // Handle file paths (absolute or relative with slashes)
-            console.log(`Loading from file: ${actualModelSrc}`);
-            actualPath = actualModelSrc;
-        }
-        else {
-            // Handle local files and cached files
-            actualPath = await resolveLocalOrCachedFile(actualModelSrc);
-        }
+        // Handle local files and cached files
+        actualPath = await resolveLocalOrCachedFile(actualModelSrc);
     }
     return actualPath;
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@qvac/sdk",
-  "version": "0.2.1-dev.1760851161.68dc306",
+  "version": "0.2.1",
   "type": "module",
   "main": "./dist/index.js",
   "types": "./dist/index.d.ts",
@@ -45,7 +45,7 @@
     "bare:ensureprebuilds:whispercpp": "cp -r node_modules/@qvac/transcription-whispercpp/prebuilds/android-arm64 node_modules/@qvac/transcription-whispercpp/prebuilds/android-ia32 && cp -r node_modules/@qvac/transcription-whispercpp/prebuilds/android-arm64 node_modules/@qvac/transcription-whispercpp/prebuilds/android-arm && cp -r node_modules/@qvac/transcription-whispercpp/prebuilds/android-arm64 node_modules/@qvac/transcription-whispercpp/prebuilds/android-x64",
     "bare:ensureprebuilds:embeddings": "cp -r node_modules/@qvac/embed-llamacpp/prebuilds/android-arm64 node_modules/@qvac/embed-llamacpp/prebuilds/android-ia32 && cp -r node_modules/@qvac/embed-llamacpp/prebuilds/android-arm64 node_modules/@qvac/embed-llamacpp/prebuilds/android-arm && cp -r node_modules/@qvac/embed-llamacpp/prebuilds/android-arm64 node_modules/@qvac/embed-llamacpp/prebuilds/android-x64",
     "bare:ensureprebuilds": "bun run bare:ensureprebuilds:llamacpp && bun run bare:ensureprebuilds:whispercpp && bun run bare:ensureprebuilds:embeddings",
-    "prepare": "bun run husky",
+    "prepare:husky": "bun run husky",
     "bare:example": "./scripts/run-with-bare.sh",
     "docs_api:website": "typedoc --out ./docs/api/website --entryPoints index.ts --entryPointStrategy expand --readme none",
     "docs_api:markdown": "typedoc --plugin typedoc-plugin-markdown --out ./docs/api/markdown --entryPoints index.ts --entryPointStrategy expand --readme none && sed -i '1,4d' ./docs/api/markdown/README.md",
@@ -53,13 +53,13 @@
     "update-models": "node ./scripts/update-models.js"
   },
   "dependencies": {
-    "@qvac/embed-llamacpp": "^0.1.0",
-    "@qvac/llm-llamacpp": "^0.2.2",
+    "@qvac/embed-llamacpp": "^0.3.1",
+    "@qvac/llm-llamacpp": "^0.3.2",
     "@qvac/decoder-audio": "^0.1.0",
     "@qvac/error": "^0.1.0",
     "@qvac/langdetect-text": "^0.1.0",
     "@qvac/logging": "^0.1.0",
-    "@qvac/rag": "^0.1.0",
+    "@qvac/rag": "^0.2.1",
     "@qvac/transcription-whispercpp": "^0.1.4",
     "@qvac/translation-nmtcpp": "^0.1.0",
     "fast-safe-stringify": "2.1.1",
@@ -126,4 +126,4 @@
   "patchedDependencies": {
     "bare-rpc@0.2.11": "patches/bare-rpc@0.2.11.patch"
   }
-}
+}