@qvac/sdk 0.1.1 → 0.2.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (49) hide show
  1. package/dist/client/api/completion-stream.d.ts +19 -4
  2. package/dist/client/api/completion-stream.js +19 -4
  3. package/dist/client/api/delete-cache.d.ts +29 -0
  4. package/dist/client/api/delete-cache.js +47 -0
  5. package/dist/client/api/index.d.ts +1 -0
  6. package/dist/client/api/index.js +1 -0
  7. package/dist/client/rpc/bare-client.d.ts +5 -0
  8. package/dist/client/rpc/bare-client.js +37 -1
  9. package/dist/examples/download-with-cancel.js +3 -3
  10. package/dist/examples/kv-cache-cleanup.d.ts +2 -0
  11. package/dist/examples/kv-cache-cleanup.js +88 -0
  12. package/dist/examples/kv-cache-custom-key.d.ts +2 -0
  13. package/dist/examples/kv-cache-custom-key.js +87 -0
  14. package/dist/examples/llamacpp-multimodal.js +3 -3
  15. package/dist/examples/rag/rag-hyperdb-workspaces.js +2 -1
  16. package/dist/index.d.ts +1 -1
  17. package/dist/index.js +1 -1
  18. package/dist/models/hyperdrive/models.d.ts +49 -1
  19. package/dist/models/hyperdrive/models.js +105 -54
  20. package/dist/schemas/common.d.ts +22 -6
  21. package/dist/schemas/common.js +3 -0
  22. package/dist/schemas/completion-stream.d.ts +5 -3
  23. package/dist/schemas/completion-stream.js +3 -2
  24. package/dist/schemas/delete-cache.d.ts +17 -0
  25. package/dist/schemas/delete-cache.js +17 -0
  26. package/dist/schemas/index.d.ts +1 -0
  27. package/dist/schemas/index.js +1 -0
  28. package/dist/schemas/llamacpp-config.d.ts +2 -1
  29. package/dist/schemas/llamacpp-config.js +5 -2
  30. package/dist/schemas/load-model.d.ts +16 -6
  31. package/dist/schemas/rag.d.ts +3 -0
  32. package/dist/schemas/rag.js +3 -1
  33. package/dist/server/bare/addons/llamacpp/handlers/completion-stream.js +135 -53
  34. package/dist/server/bare/addons/llamacpp/kv-cache-utils.d.ts +13 -4
  35. package/dist/server/bare/addons/llamacpp/kv-cache-utils.js +70 -7
  36. package/dist/server/bare/addons/load-model.js +1 -1
  37. package/dist/server/bare/registry/model-registry.d.ts +8 -1
  38. package/dist/server/bare/registry/model-registry.js +9 -0
  39. package/dist/server/rpc/handle-request.js +11 -0
  40. package/dist/server/rpc/handlers/completion-stream.js +31 -45
  41. package/dist/server/rpc/handlers/delete-cache.d.ts +3 -0
  42. package/dist/server/rpc/handlers/delete-cache.js +29 -0
  43. package/dist/server/rpc/handlers/index.d.ts +4 -0
  44. package/dist/server/rpc/handlers/index.js +4 -0
  45. package/dist/server/rpc/handlers/load-model/download-manager.d.ts +0 -19
  46. package/dist/server/rpc/handlers/load-model/download-manager.js +10 -25
  47. package/dist/server/rpc/handlers/load-model/hyperdrive.js +121 -80
  48. package/dist/server/rpc/handlers/load-model/resolve.js +8 -55
  49. package/package.json +6 -6
@@ -65,6 +65,22 @@ export const models = [
65
65
  expectedSize: 1056782912,
66
66
  sha256Checksum: "c876f159707a4e4f70e045106c69db15bfc935a4981706fd4f65c6e7ea1e81c5",
67
67
  },
68
+ {
69
+ hyperdriveKey: "583f04c31d151b29e02aaf98f80d16aff585714ef9e2806a5067651c0f64ae31",
70
+ hyperbeeKey: "generation:qwen2.5-omni:multimodal::3B::1.0.0",
71
+ modelId: "Qwen2.5-Omni-3B-Q4_K_M.gguf",
72
+ addon: "llm",
73
+ expectedSize: 2104931648,
74
+ sha256Checksum: "4b0bd358c1e9ec55dd3055ef6d71c958c821533d85916a10cfa89c4552a86e29",
75
+ },
76
+ {
77
+ hyperdriveKey: "583f04c31d151b29e02aaf98f80d16aff585714ef9e2806a5067651c0f64ae31",
78
+ hyperbeeKey: "generation:qwen2.5-omni:multimodal::3B::1.0.0",
79
+ modelId: "mmproj-Qwen2.5-Omni-3B-Q8_0.gguf",
80
+ addon: "llm",
81
+ expectedSize: 1538031328,
82
+ sha256Checksum: "4e6c816cd33f7298d07cb780c136a396631e50e62f6501660271f8c6e302e565",
83
+ },
68
84
  {
69
85
  hyperdriveKey: "d331ed49c444f90b3b2a70aa5b820752685142bf2d769a0592ed666140477578",
70
86
  hyperbeeKey: "generation:qwen3:instruct::0.6B::1.0.0",
@@ -105,6 +121,38 @@ export const models = [
105
121
  expectedSize: 2401075552,
106
122
  sha256Checksum: "9cfebb63cd7e6b6d1755ea8ef001fcdadfa17831da1db96eafe3fc554913d5e1",
107
123
  },
124
+ {
125
+ hyperdriveKey: "b99c6f5dcadaad827cdcb1a129a3b34aff6ad28056738e0bd559c7ffb807e30f",
126
+ hyperbeeKey: "generation:smolVLM2:multimodal:2:500M:f16:1.0.0:video-instruct",
127
+ modelId: "SmolVLM2-500M-Video-Instruct-f16.gguf",
128
+ addon: "llm",
129
+ expectedSize: 820424704,
130
+ sha256Checksum: "80f7e3f04bc2d3324ac1a9f52f5776fe13a69912adf74f8e7edacf773d140d77",
131
+ },
132
+ {
133
+ hyperdriveKey: "b99c6f5dcadaad827cdcb1a129a3b34aff6ad28056738e0bd559c7ffb807e30f",
134
+ hyperbeeKey: "generation:smolVLM2:multimodal:2:500M:f16:1.0.0:video-instruct",
135
+ modelId: "mmproj-SmolVLM2-500M-Video-Instruct-f16.gguf",
136
+ addon: "llm",
137
+ expectedSize: 199470624,
138
+ sha256Checksum: "b5dc8ebe7cbeab66a5369693960a52515d7824f13d4063ceca78431f2a6b59b0",
139
+ },
140
+ {
141
+ hyperdriveKey: "73b1bc01d01e25fa27be7d7f434337d14f054b0315e8463766ca31e778ac6576",
142
+ hyperbeeKey: "generation:smolVLM2:multimodal:2:500M:q8_0:1.0.0:video-instruct",
143
+ modelId: "SmolVLM2-500M-Video-Instruct-Q8_0.gguf",
144
+ addon: "llm",
145
+ expectedSize: 436808704,
146
+ sha256Checksum: "6f67b8036b2469fcd71728702720c6b51aebd759b78137a8120733b4d66438bc",
147
+ },
148
+ {
149
+ hyperdriveKey: "73b1bc01d01e25fa27be7d7f434337d14f054b0315e8463766ca31e778ac6576",
150
+ hyperbeeKey: "generation:smolVLM2:multimodal:2:500M:q8_0:1.0.0:video-instruct",
151
+ modelId: "mmproj-SmolVLM2-500M-Video-Instruct-Q8_0.gguf",
152
+ addon: "llm",
153
+ expectedSize: 108785184,
154
+ sha256Checksum: "921dc7e259f308e5b027111fa185efcbf33db13f6e35749ddf7f5cdb60ef520b",
155
+ },
108
156
  {
109
157
  hyperdriveKey: "268c2e9b2a3420632e4b6649e32822f42d5dfbda4c7e96daec5b629ed20f99f7",
110
158
  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
475
523
  export const LASER_DOLPHIN_2_6_2X7B_INST_Q2_K = `pear://${models[5].hyperdriveKey}/${models[5].modelId}`;
476
524
  export const LLAMA_3_2_1B_INST_Q4_0 = `pear://${models[6].hyperdriveKey}/${models[6].modelId}`;
477
525
  export const QWEN_3_1_7B_INST_Q4 = `pear://${models[7].hyperdriveKey}/${models[7].modelId}`;
478
- export const QWEN3_0_6B_INST = `pear://${models[8].hyperdriveKey}/${models[8].modelId}`;
479
- export const QWEN3_4B_INST_Q4 = `pear://${models[9].hyperdriveKey}/${models[9].modelId}`;
480
- export const QWEN3_600M_INST_Q4 = `pear://${models[10].hyperdriveKey}/${models[10].modelId}`;
481
- export const SALAMANDRATA_2B_INST_Q4 = `pear://${models[11].hyperdriveKey}/${models[11].modelId}`;
482
- export const SALAMANDRATA_2B_INST_Q8 = `pear://${models[12].hyperdriveKey}/${models[12].modelId}`;
483
- export const MARIAN_EN_HI_INDIC_200M_Q0F32 = `pear://${models[13].hyperdriveKey}/${models[13].modelId}`;
484
- export const MARIAN_HI_EN_INDIC_200M_Q0F32 = `pear://${models[14].hyperdriveKey}/${models[14].modelId}`;
485
- export const MARIAN_HI_HI_INDIC_320M_Q0F32 = `pear://${models[15].hyperdriveKey}/${models[15].modelId}`;
486
- export const MARIAN_EN_HI_INDIC_1B_Q0F16 = `pear://${models[16].hyperdriveKey}/${models[16].modelId}`;
487
- export const MARIAN_HI_EN_INDIC_1B_Q0F16 = `pear://${models[17].hyperdriveKey}/${models[17].modelId}`;
488
- export const MARIAN_HI_HI_INDIC_1B_Q0F16 = `pear://${models[18].hyperdriveKey}/${models[18].modelId}`;
489
- export const MARIAN_EN_HI_INDIC_1B_Q0F32 = `pear://${models[19].hyperdriveKey}/${models[19].modelId}`;
490
- export const MARIAN_HI_EN_INDIC_1B_Q0F32 = `pear://${models[20].hyperdriveKey}/${models[20].modelId}`;
491
- export const MARIAN_HI_HI_INDIC_1B_Q0F32 = `pear://${models[21].hyperdriveKey}/${models[21].modelId}`;
492
- export const MARIAN_EN_HI_INDIC_1B_Q4_0 = `pear://${models[22].hyperdriveKey}/${models[22].modelId}`;
493
- export const MARIAN_HI_EN_INDIC_1B_Q4_0 = `pear://${models[23].hyperdriveKey}/${models[23].modelId}`;
494
- export const MARIAN_HI_HI_INDIC_1B_Q4_0 = `pear://${models[24].hyperdriveKey}/${models[24].modelId}`;
495
- export const MARIAN_EN_HI_INDIC_200M_Q0F16 = `pear://${models[25].hyperdriveKey}/${models[25].modelId}`;
496
- export const MARIAN_HI_EN_INDIC_200M_Q0F16 = `pear://${models[26].hyperdriveKey}/${models[26].modelId}`;
497
- export const MARIAN_EN_HI_INDIC_200M_Q4_0 = `pear://${models[27].hyperdriveKey}/${models[27].modelId}`;
498
- export const MARIAN_HI_EN_INDIC_200M_Q4_0 = `pear://${models[28].hyperdriveKey}/${models[28].modelId}`;
499
- export const MARIAN_HI_HI_INDIC_320M_Q0F16 = `pear://${models[29].hyperdriveKey}/${models[29].modelId}`;
500
- export const MARIAN_HI_HI_INDIC_320M_Q4_0 = `pear://${models[30].hyperdriveKey}/${models[30].modelId}`;
501
- export const MARIAN_OPUS_EN_IT = `pear://${models[31].hyperdriveKey}/${models[31].modelId}`;
502
- export const MARIAN_OPUS_IT_EN = `pear://${models[32].hyperdriveKey}/${models[32].modelId}`;
503
- export const MARIAN_OPUS_DE_EN_Q0F32 = `pear://${models[33].hyperdriveKey}/${models[33].modelId}`;
504
- export const MARIAN_OPUS_DE_ES_Q0F32 = `pear://${models[34].hyperdriveKey}/${models[34].modelId}`;
505
- export const MARIAN_OPUS_DE_IT_Q0F32 = `pear://${models[35].hyperdriveKey}/${models[35].modelId}`;
506
- export const MARIAN_OPUS_EN_DE_Q0F32 = `pear://${models[36].hyperdriveKey}/${models[36].modelId}`;
507
- export const MARIAN_OPUS_EN_ES_Q0F32 = `pear://${models[37].hyperdriveKey}/${models[37].modelId}`;
508
- export const MARIAN_OPUS_EN_IT_Q0F32 = `pear://${models[38].hyperdriveKey}/${models[38].modelId}`;
509
- export const MARIAN_OPUS_ES_DE_Q0F32 = `pear://${models[39].hyperdriveKey}/${models[39].modelId}`;
510
- export const MARIAN_OPUS_ES_EN_Q0F32 = `pear://${models[40].hyperdriveKey}/${models[40].modelId}`;
511
- export const MARIAN_OPUS_ES_IT_Q0F32 = `pear://${models[41].hyperdriveKey}/${models[41].modelId}`;
512
- export const MARIAN_OPUS_IT_DE_Q0F32 = `pear://${models[42].hyperdriveKey}/${models[42].modelId}`;
513
- export const MARIAN_OPUS_IT_EN_Q0F32 = `pear://${models[43].hyperdriveKey}/${models[43].modelId}`;
514
- export const MARIAN_OPUS_IT_ES_Q0F32 = `pear://${models[44].hyperdriveKey}/${models[44].modelId}`;
515
- export const MARIAN_OPUS_EN_RU_Q0F16 = `pear://${models[45].hyperdriveKey}/${models[45].modelId}`;
516
- export const MARIAN_OPUS_RU_EN_Q0F16 = `pear://${models[46].hyperdriveKey}/${models[46].modelId}`;
517
- export const MARIAN_OPUS_EN_RU_Q0F32 = `pear://${models[47].hyperdriveKey}/${models[47].modelId}`;
518
- export const MARIAN_OPUS_RU_EN_Q0F32 = `pear://${models[48].hyperdriveKey}/${models[48].modelId}`;
519
- export const MARIAN_OPUS_EN_RU_Q4_0 = `pear://${models[49].hyperdriveKey}/${models[49].modelId}`;
520
- export const MARIAN_OPUS_RU_EN_Q4_0 = `pear://${models[50].hyperdriveKey}/${models[50].modelId}`;
521
- export const MARIAN_OPUS_EN_JA_Q4F16_1 = `pear://${models[51].hyperdriveKey}/${models[51].modelId}`;
522
- export const MARIAN_OPUS_JA_EN_Q4F16_1 = `pear://${models[52].hyperdriveKey}/${models[52].modelId}`;
523
- export const VAD_SILERO_5_1_2 = `pear://${models[53].hyperdriveKey}/${models[53].modelId}`;
524
- export const WHISPER_LARGE_3 = `pear://${models[54].hyperdriveKey}/${models[54].modelId}`;
525
- export const WHISPER_SMALL = `pear://${models[55].hyperdriveKey}/${models[55].modelId}`;
526
- export const WHISPER_TINY = `pear://${models[56].hyperdriveKey}/${models[56].modelId}`;
527
- export const WHISPER_TINY_SILERO = `pear://${models[57].hyperdriveKey}/${models[57].modelId}`;
526
+ export const QWEN2_5_OMNI_3B_MULTIMODAL = `pear://${models[8].hyperdriveKey}/${models[8].modelId}`;
527
+ export const MMPROJ_QWEN2_5_OMNI_3B_MULTIMODAL = `pear://${models[9].hyperdriveKey}/${models[9].modelId}`;
528
+ export const QWEN3_0_6B_INST = `pear://${models[10].hyperdriveKey}/${models[10].modelId}`;
529
+ export const QWEN3_4B_INST_Q4 = `pear://${models[11].hyperdriveKey}/${models[11].modelId}`;
530
+ export const QWEN3_600M_INST_Q4 = `pear://${models[12].hyperdriveKey}/${models[12].modelId}`;
531
+ export const SALAMANDRATA_2B_INST_Q4 = `pear://${models[13].hyperdriveKey}/${models[13].modelId}`;
532
+ export const SALAMANDRATA_2B_INST_Q8 = `pear://${models[14].hyperdriveKey}/${models[14].modelId}`;
533
+ export const SMOLVLM2_2_500M_MULTIMODAL_F16 = `pear://${models[15].hyperdriveKey}/${models[15].modelId}`;
534
+ export const MMPROJ_SMOLVLM2_2_500M_MULTIMODAL_F16 = `pear://${models[16].hyperdriveKey}/${models[16].modelId}`;
535
+ export const SMOLVLM2_2_500M_MULTIMODAL_Q8_0 = `pear://${models[17].hyperdriveKey}/${models[17].modelId}`;
536
+ export const MMPROJ_SMOLVLM2_2_500M_MULTIMODAL_Q8_0 = `pear://${models[18].hyperdriveKey}/${models[18].modelId}`;
537
+ export const MARIAN_EN_HI_INDIC_200M_Q0F32 = `pear://${models[19].hyperdriveKey}/${models[19].modelId}`;
538
+ export const MARIAN_HI_EN_INDIC_200M_Q0F32 = `pear://${models[20].hyperdriveKey}/${models[20].modelId}`;
539
+ export const MARIAN_HI_HI_INDIC_320M_Q0F32 = `pear://${models[21].hyperdriveKey}/${models[21].modelId}`;
540
+ export const MARIAN_EN_HI_INDIC_1B_Q0F16 = `pear://${models[22].hyperdriveKey}/${models[22].modelId}`;
541
+ export const MARIAN_HI_EN_INDIC_1B_Q0F16 = `pear://${models[23].hyperdriveKey}/${models[23].modelId}`;
542
+ export const MARIAN_HI_HI_INDIC_1B_Q0F16 = `pear://${models[24].hyperdriveKey}/${models[24].modelId}`;
543
+ export const MARIAN_EN_HI_INDIC_1B_Q0F32 = `pear://${models[25].hyperdriveKey}/${models[25].modelId}`;
544
+ export const MARIAN_HI_EN_INDIC_1B_Q0F32 = `pear://${models[26].hyperdriveKey}/${models[26].modelId}`;
545
+ export const MARIAN_HI_HI_INDIC_1B_Q0F32 = `pear://${models[27].hyperdriveKey}/${models[27].modelId}`;
546
+ export const MARIAN_EN_HI_INDIC_1B_Q4_0 = `pear://${models[28].hyperdriveKey}/${models[28].modelId}`;
547
+ export const MARIAN_HI_EN_INDIC_1B_Q4_0 = `pear://${models[29].hyperdriveKey}/${models[29].modelId}`;
548
+ export const MARIAN_HI_HI_INDIC_1B_Q4_0 = `pear://${models[30].hyperdriveKey}/${models[30].modelId}`;
549
+ export const MARIAN_EN_HI_INDIC_200M_Q0F16 = `pear://${models[31].hyperdriveKey}/${models[31].modelId}`;
550
+ export const MARIAN_HI_EN_INDIC_200M_Q0F16 = `pear://${models[32].hyperdriveKey}/${models[32].modelId}`;
551
+ export const MARIAN_EN_HI_INDIC_200M_Q4_0 = `pear://${models[33].hyperdriveKey}/${models[33].modelId}`;
552
+ export const MARIAN_HI_EN_INDIC_200M_Q4_0 = `pear://${models[34].hyperdriveKey}/${models[34].modelId}`;
553
+ export const MARIAN_HI_HI_INDIC_320M_Q0F16 = `pear://${models[35].hyperdriveKey}/${models[35].modelId}`;
554
+ export const MARIAN_HI_HI_INDIC_320M_Q4_0 = `pear://${models[36].hyperdriveKey}/${models[36].modelId}`;
555
+ export const MARIAN_OPUS_EN_IT = `pear://${models[37].hyperdriveKey}/${models[37].modelId}`;
556
+ export const MARIAN_OPUS_IT_EN = `pear://${models[38].hyperdriveKey}/${models[38].modelId}`;
557
+ export const MARIAN_OPUS_DE_EN_Q0F32 = `pear://${models[39].hyperdriveKey}/${models[39].modelId}`;
558
+ export const MARIAN_OPUS_DE_ES_Q0F32 = `pear://${models[40].hyperdriveKey}/${models[40].modelId}`;
559
+ export const MARIAN_OPUS_DE_IT_Q0F32 = `pear://${models[41].hyperdriveKey}/${models[41].modelId}`;
560
+ export const MARIAN_OPUS_EN_DE_Q0F32 = `pear://${models[42].hyperdriveKey}/${models[42].modelId}`;
561
+ export const MARIAN_OPUS_EN_ES_Q0F32 = `pear://${models[43].hyperdriveKey}/${models[43].modelId}`;
562
+ export const MARIAN_OPUS_EN_IT_Q0F32 = `pear://${models[44].hyperdriveKey}/${models[44].modelId}`;
563
+ export const MARIAN_OPUS_ES_DE_Q0F32 = `pear://${models[45].hyperdriveKey}/${models[45].modelId}`;
564
+ export const MARIAN_OPUS_ES_EN_Q0F32 = `pear://${models[46].hyperdriveKey}/${models[46].modelId}`;
565
+ export const MARIAN_OPUS_ES_IT_Q0F32 = `pear://${models[47].hyperdriveKey}/${models[47].modelId}`;
566
+ export const MARIAN_OPUS_IT_DE_Q0F32 = `pear://${models[48].hyperdriveKey}/${models[48].modelId}`;
567
+ export const MARIAN_OPUS_IT_EN_Q0F32 = `pear://${models[49].hyperdriveKey}/${models[49].modelId}`;
568
+ export const MARIAN_OPUS_IT_ES_Q0F32 = `pear://${models[50].hyperdriveKey}/${models[50].modelId}`;
569
+ export const MARIAN_OPUS_EN_RU_Q0F16 = `pear://${models[51].hyperdriveKey}/${models[51].modelId}`;
570
+ export const MARIAN_OPUS_RU_EN_Q0F16 = `pear://${models[52].hyperdriveKey}/${models[52].modelId}`;
571
+ export const MARIAN_OPUS_EN_RU_Q0F32 = `pear://${models[53].hyperdriveKey}/${models[53].modelId}`;
572
+ export const MARIAN_OPUS_RU_EN_Q0F32 = `pear://${models[54].hyperdriveKey}/${models[54].modelId}`;
573
+ export const MARIAN_OPUS_EN_RU_Q4_0 = `pear://${models[55].hyperdriveKey}/${models[55].modelId}`;
574
+ export const MARIAN_OPUS_RU_EN_Q4_0 = `pear://${models[56].hyperdriveKey}/${models[56].modelId}`;
575
+ export const MARIAN_OPUS_EN_JA_Q4F16_1 = `pear://${models[57].hyperdriveKey}/${models[57].modelId}`;
576
+ export const MARIAN_OPUS_JA_EN_Q4F16_1 = `pear://${models[58].hyperdriveKey}/${models[58].modelId}`;
577
+ export const VAD_SILERO_5_1_2 = `pear://${models[59].hyperdriveKey}/${models[59].modelId}`;
578
+ export const WHISPER_LARGE_3 = `pear://${models[60].hyperdriveKey}/${models[60].modelId}`;
579
+ export const WHISPER_SMALL = `pear://${models[61].hyperdriveKey}/${models[61].modelId}`;
580
+ export const WHISPER_TINY = `pear://${models[62].hyperdriveKey}/${models[62].modelId}`;
581
+ export const WHISPER_TINY_SILERO = `pear://${models[63].hyperdriveKey}/${models[63].modelId}`;
528
582
  // Helper function for our curated model list
529
583
  export function getModelBySrc(modelId, hyperdriveKey) {
530
- if (hyperdriveKey) {
531
- return models.find((model) => model.modelId === modelId && model.hyperdriveKey === hyperdriveKey);
532
- }
533
- return models.find((model) => model.modelId === modelId);
584
+ return models.find((model) => model.modelId === modelId && model.hyperdriveKey === hyperdriveKey);
534
585
  }
@@ -25,13 +25,14 @@ export declare const requestSchema: z.ZodUnion<readonly [z.ZodObject<{
25
25
  lora: z.ZodOptional<z.ZodString>;
26
26
  device: z.ZodDefault<z.ZodString>;
27
27
  predict: z.ZodOptional<z.ZodUnion<readonly [z.ZodLiteral<-1>, z.ZodLiteral<-2>, z.ZodNumber]>>;
28
- system_prompt: z.ZodOptional<z.ZodString>;
28
+ system_prompt: z.ZodDefault<z.ZodString>;
29
29
  no_mmap: z.ZodOptional<z.ZodBoolean>;
30
30
  verbosity: z.ZodOptional<z.ZodUnion<readonly [z.ZodLiteral<-1>, z.ZodLiteral<0>, z.ZodLiteral<1>]>>;
31
31
  presence_penalty: z.ZodOptional<z.ZodNumber>;
32
32
  frequency_penalty: z.ZodOptional<z.ZodNumber>;
33
33
  repeat_penalty: z.ZodOptional<z.ZodNumber>;
34
34
  reverse_prompt: z.ZodOptional<z.ZodString>;
35
+ n_discarded: z.ZodOptional<z.ZodNumber>;
35
36
  }, z.core.$strip>;
36
37
  }, z.core.$strip>, z.ZodObject<{
37
38
  type: z.ZodLiteral<"loadModel">;
@@ -186,19 +187,20 @@ export declare const requestSchema: z.ZodUnion<readonly [z.ZodObject<{
186
187
  ctx_size: number;
187
188
  gpu_layers: number;
188
189
  device: string;
190
+ system_prompt: string;
189
191
  temp?: number | undefined;
190
192
  top_p?: number | undefined;
191
193
  top_k?: number | undefined;
192
194
  seed?: number | undefined;
193
195
  lora?: string | undefined;
194
196
  predict?: number | undefined;
195
- system_prompt?: string | undefined;
196
197
  no_mmap?: boolean | undefined;
197
198
  verbosity?: 0 | 1 | -1 | undefined;
198
199
  presence_penalty?: number | undefined;
199
200
  frequency_penalty?: number | undefined;
200
201
  repeat_penalty?: number | undefined;
201
202
  reverse_prompt?: string | undefined;
203
+ n_discarded?: number | undefined;
202
204
  };
203
205
  projectionModelSrc?: string | undefined;
204
206
  vadModelSrc?: string | undefined;
@@ -283,19 +285,20 @@ export declare const requestSchema: z.ZodUnion<readonly [z.ZodObject<{
283
285
  ctx_size: number;
284
286
  gpu_layers: number;
285
287
  device: string;
288
+ system_prompt: string;
286
289
  temp?: number | undefined;
287
290
  top_p?: number | undefined;
288
291
  top_k?: number | undefined;
289
292
  seed?: number | undefined;
290
293
  lora?: string | undefined;
291
294
  predict?: number | undefined;
292
- system_prompt?: string | undefined;
293
295
  no_mmap?: boolean | undefined;
294
296
  verbosity?: 0 | 1 | -1 | undefined;
295
297
  presence_penalty?: number | undefined;
296
298
  frequency_penalty?: number | undefined;
297
299
  repeat_penalty?: number | undefined;
298
300
  reverse_prompt?: string | undefined;
301
+ n_discarded?: number | undefined;
299
302
  };
300
303
  projectionModelSrc?: string | undefined;
301
304
  vadModelSrc?: string | undefined;
@@ -419,7 +422,7 @@ export declare const requestSchema: z.ZodUnion<readonly [z.ZodObject<{
419
422
  }, z.core.$strip>;
420
423
  }, z.core.$strip>>>;
421
424
  stream: z.ZodBoolean;
422
- kvCache: z.ZodOptional<z.ZodBoolean>;
425
+ kvCache: z.ZodOptional<z.ZodUnion<readonly [z.ZodBoolean, z.ZodString]>>;
423
426
  type: z.ZodLiteral<"completionStream">;
424
427
  }, z.core.$strip>, z.ZodObject<{
425
428
  modelId: z.ZodString;
@@ -483,6 +486,7 @@ export declare const requestSchema: z.ZodUnion<readonly [z.ZodObject<{
483
486
  paragraph: "paragraph";
484
487
  }>>;
485
488
  splitStrategy: z.ZodOptional<z.ZodEnum<{
489
+ token: "token";
486
490
  character: "character";
487
491
  word: "word";
488
492
  sentence: "sentence";
@@ -505,7 +509,14 @@ export declare const requestSchema: z.ZodUnion<readonly [z.ZodObject<{
505
509
  ids: z.ZodArray<z.ZodString>;
506
510
  type: z.ZodLiteral<"rag">;
507
511
  operation: z.ZodLiteral<"deleteEmbeddings">;
508
- }, z.core.$strip>], "operation">]>;
512
+ }, z.core.$strip>], "operation">, z.ZodUnion<readonly [z.ZodObject<{
513
+ type: z.ZodLiteral<"deleteCache">;
514
+ all: z.ZodLiteral<true>;
515
+ }, z.core.$strip>, z.ZodObject<{
516
+ type: z.ZodLiteral<"deleteCache">;
517
+ kvCacheKey: z.ZodString;
518
+ modelId: z.ZodOptional<z.ZodString>;
519
+ }, z.core.$strip>]>]>;
509
520
  export declare const responseSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
510
521
  type: z.ZodLiteral<"pong">;
511
522
  number: z.ZodNumber;
@@ -526,6 +537,7 @@ export declare const responseSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
526
537
  stats: z.ZodOptional<z.ZodObject<{
527
538
  timeToFirstToken: z.ZodNumber;
528
539
  tokensPerSecond: z.ZodNumber;
540
+ cacheTokens: z.ZodNumber;
529
541
  }, z.core.$strip>>;
530
542
  error: z.ZodOptional<z.ZodString>;
531
543
  }, z.core.$strip>, z.ZodObject<{
@@ -616,7 +628,11 @@ export declare const responseSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
616
628
  success: z.ZodBoolean;
617
629
  error: z.ZodOptional<z.ZodString>;
618
630
  operation: z.ZodLiteral<"deleteEmbeddings">;
619
- }, z.core.$strip>], "operation">], "type">;
631
+ }, z.core.$strip>], "operation">, z.ZodObject<{
632
+ type: z.ZodLiteral<"deleteCache">;
633
+ success: z.ZodBoolean;
634
+ error: z.ZodOptional<z.ZodString>;
635
+ }, z.core.$strip>], "type">;
620
636
  export declare const rpcOptionsSchema: z.ZodObject<{
621
637
  timeout: z.ZodOptional<z.ZodNumber>;
622
638
  }, z.core.$strip>;
@@ -12,6 +12,7 @@ import { translateRequestSchema, translateResponseSchema } from "./translate.js"
12
12
  import { loggingStreamRequestSchema, loggingStreamResponseSchema, } from "./logging-stream.js";
13
13
  import { errorResponseSchema } from "./error.js";
14
14
  import { ragRequestSchema, ragResponseSchema } from "./rag.js";
15
+ import { deleteCacheRequestSchema, deleteCacheResponseSchema, } from "./delete-cache.js";
15
16
  export const requestSchema = z.union([
16
17
  pingRequestSchema,
17
18
  loadModelRequestSchema,
@@ -25,6 +26,7 @@ export const requestSchema = z.union([
25
26
  cancelRequestSchema,
26
27
  provideRequestSchema,
27
28
  ragRequestSchema,
29
+ deleteCacheRequestSchema,
28
30
  ]);
29
31
  export const responseSchema = z.discriminatedUnion("type", [
30
32
  pingResponseSchema,
@@ -41,6 +43,7 @@ export const responseSchema = z.discriminatedUnion("type", [
41
43
  provideResponseSchema,
42
44
  errorResponseSchema,
43
45
  ragResponseSchema,
46
+ deleteCacheResponseSchema,
44
47
  ]);
45
48
  export const rpcOptionsSchema = z.object({
46
49
  timeout: z.number().min(100).optional(),
@@ -32,7 +32,7 @@ export declare const completionParamsSchema: z.ZodObject<{
32
32
  }, z.core.$strip>>>;
33
33
  }, z.core.$strip>>;
34
34
  modelId: z.ZodString;
35
- kvCache: z.ZodOptional<z.ZodBoolean>;
35
+ kvCache: z.ZodOptional<z.ZodUnion<readonly [z.ZodBoolean, z.ZodString]>>;
36
36
  }, z.core.$strip>;
37
37
  export declare const completionClientParamsSchema: z.ZodObject<{
38
38
  history: z.ZodArray<z.ZodObject<{
@@ -65,7 +65,7 @@ export declare const completionClientParamsSchema: z.ZodObject<{
65
65
  }, z.core.$strip>;
66
66
  }, z.core.$strip>>>;
67
67
  stream: z.ZodBoolean;
68
- kvCache: z.ZodOptional<z.ZodBoolean>;
68
+ kvCache: z.ZodOptional<z.ZodUnion<readonly [z.ZodBoolean, z.ZodString]>>;
69
69
  }, z.core.$strip>;
70
70
  export declare const completionStreamRequestSchema: z.ZodObject<{
71
71
  history: z.ZodArray<z.ZodObject<{
@@ -98,12 +98,13 @@ export declare const completionStreamRequestSchema: z.ZodObject<{
98
98
  }, z.core.$strip>;
99
99
  }, z.core.$strip>>>;
100
100
  stream: z.ZodBoolean;
101
- kvCache: z.ZodOptional<z.ZodBoolean>;
101
+ kvCache: z.ZodOptional<z.ZodUnion<readonly [z.ZodBoolean, z.ZodString]>>;
102
102
  type: z.ZodLiteral<"completionStream">;
103
103
  }, z.core.$strip>;
104
104
  export declare const completionStatsSchema: z.ZodObject<{
105
105
  timeToFirstToken: z.ZodNumber;
106
106
  tokensPerSecond: z.ZodNumber;
107
+ cacheTokens: z.ZodNumber;
107
108
  }, z.core.$strip>;
108
109
  export declare const completionStreamResponseSchema: z.ZodObject<{
109
110
  type: z.ZodLiteral<"completionStream">;
@@ -112,6 +113,7 @@ export declare const completionStreamResponseSchema: z.ZodObject<{
112
113
  stats: z.ZodOptional<z.ZodObject<{
113
114
  timeToFirstToken: z.ZodNumber;
114
115
  tokensPerSecond: z.ZodNumber;
116
+ cacheTokens: z.ZodNumber;
115
117
  }, z.core.$strip>>;
116
118
  error: z.ZodOptional<z.ZodString>;
117
119
  }, z.core.$strip>;
@@ -30,12 +30,12 @@ export const completionParamsSchema = z.object({
30
30
  attachments: z.array(attachmentSchema).optional(),
31
31
  })),
32
32
  modelId: z.string(),
33
- kvCache: z.boolean().optional(),
33
+ kvCache: z.union([z.boolean(), z.string()]).optional(),
34
34
  });
35
35
  export const completionClientParamsSchema = completionParamsSchema.extend({
36
36
  tools: z.array(toolSchema).optional(),
37
37
  stream: z.boolean(),
38
- kvCache: z.boolean().optional(),
38
+ kvCache: z.union([z.boolean(), z.string()]).optional(),
39
39
  });
40
40
  export const completionStreamRequestSchema = completionClientParamsSchema.extend({
41
41
  type: z.literal("completionStream"),
@@ -43,6 +43,7 @@ export const completionStreamRequestSchema = completionClientParamsSchema.extend
43
43
  export const completionStatsSchema = z.object({
44
44
  timeToFirstToken: z.number(),
45
45
  tokensPerSecond: z.number(),
46
+ cacheTokens: z.number(),
46
47
  });
47
48
  export const completionStreamResponseSchema = z.object({
48
49
  type: z.literal("completionStream"),
@@ -0,0 +1,17 @@
1
+ import { z } from "zod";
2
+ export declare const deleteCacheRequestSchema: z.ZodUnion<readonly [z.ZodObject<{
3
+ type: z.ZodLiteral<"deleteCache">;
4
+ all: z.ZodLiteral<true>;
5
+ }, z.core.$strip>, z.ZodObject<{
6
+ type: z.ZodLiteral<"deleteCache">;
7
+ kvCacheKey: z.ZodString;
8
+ modelId: z.ZodOptional<z.ZodString>;
9
+ }, z.core.$strip>]>;
10
+ export declare const deleteCacheResponseSchema: z.ZodObject<{
11
+ type: z.ZodLiteral<"deleteCache">;
12
+ success: z.ZodBoolean;
13
+ error: z.ZodOptional<z.ZodString>;
14
+ }, z.core.$strip>;
15
+ export type DeleteCacheRequest = z.infer<typeof deleteCacheRequestSchema>;
16
+ export type DeleteCacheResponse = z.infer<typeof deleteCacheResponseSchema>;
17
+ //# sourceMappingURL=delete-cache.d.ts.map
@@ -0,0 +1,17 @@
1
+ import { z } from "zod";
2
+ export const deleteCacheRequestSchema = z.union([
3
+ z.object({
4
+ type: z.literal("deleteCache"),
5
+ all: z.literal(true),
6
+ }),
7
+ z.object({
8
+ type: z.literal("deleteCache"),
9
+ kvCacheKey: z.string(),
10
+ modelId: z.string().optional(),
11
+ }),
12
+ ]);
13
+ export const deleteCacheResponseSchema = z.object({
14
+ type: z.literal("deleteCache"),
15
+ success: z.boolean(),
16
+ error: z.string().optional(),
17
+ });
@@ -1,6 +1,7 @@
1
1
  export * from "./cancel";
2
2
  export * from "./completion-stream";
3
3
  export * from "./delegate";
4
+ export * from "./delete-cache";
4
5
  export * from "./download-asset";
5
6
  export * from "./embed";
6
7
  export * from "./load-model";
@@ -2,6 +2,7 @@
2
2
  export * from "./cancel.js";
3
3
  export * from "./completion-stream.js";
4
4
  export * from "./delegate.js";
5
+ export * from "./delete-cache.js";
5
6
  export * from "./download-asset.js";
6
7
  export * from "./embed.js";
7
8
  export * from "./load-model.js";
@@ -14,13 +14,14 @@ export declare const llmConfigSchema: z.ZodObject<{
14
14
  lora: z.ZodOptional<z.ZodString>;
15
15
  device: z.ZodDefault<z.ZodString>;
16
16
  predict: z.ZodOptional<z.ZodUnion<readonly [z.ZodLiteral<-1>, z.ZodLiteral<-2>, z.ZodNumber]>>;
17
- system_prompt: z.ZodOptional<z.ZodString>;
17
+ system_prompt: z.ZodDefault<z.ZodString>;
18
18
  no_mmap: z.ZodOptional<z.ZodBoolean>;
19
19
  verbosity: z.ZodOptional<z.ZodUnion<readonly [z.ZodLiteral<-1>, z.ZodLiteral<0>, z.ZodLiteral<1>]>>;
20
20
  presence_penalty: z.ZodOptional<z.ZodNumber>;
21
21
  frequency_penalty: z.ZodOptional<z.ZodNumber>;
22
22
  repeat_penalty: z.ZodOptional<z.ZodNumber>;
23
23
  reverse_prompt: z.ZodOptional<z.ZodString>;
24
+ n_discarded: z.ZodOptional<z.ZodNumber>;
24
25
  }, z.core.$strip>;
25
26
  export declare const embedConfigSchema: z.ZodObject<{
26
27
  config: z.ZodDefault<z.ZodString>;
@@ -25,15 +25,18 @@ export const llmConfigSchema = z.object({
25
25
  z.number().int().min(1), // positive integer: fixed token count
26
26
  ])
27
27
  .optional(),
28
- system_prompt: z.string().optional(),
28
+ system_prompt: z.string().default("You are a helpful assistant."),
29
29
  no_mmap: z.boolean().optional(),
30
30
  verbosity: verbositySchema.optional(),
31
31
  presence_penalty: z.number().optional(),
32
32
  frequency_penalty: z.number().optional(),
33
33
  repeat_penalty: z.number().optional(),
34
34
  reverse_prompt: z.string().optional(),
35
+ n_discarded: z.number().optional(),
35
36
  });
36
37
  export const embedConfigSchema = z.object({
37
38
  // https://github.com/tetherto/qvac-lib-infer-llamacpp-embed/tree/main?tab=readme-ov-file#4-create-config
38
- config: z.string().default("-ngl\t25"),
39
+ config: z
40
+ .string()
41
+ .default("-ngl\t99\n-dev\tgpu\n--embd-separator\t⟪§§§EMBED_SEP§§§⟫\n-c\t512"),
39
42
  });
@@ -12,13 +12,14 @@ export declare const loadModelOptionsSchema: z.ZodPipe<z.ZodUnion<readonly [z.Zo
12
12
  lora: z.ZodOptional<z.ZodOptional<z.ZodString>>;
13
13
  device: z.ZodOptional<z.ZodDefault<z.ZodString>>;
14
14
  predict: z.ZodOptional<z.ZodOptional<z.ZodUnion<readonly [z.ZodLiteral<-1>, z.ZodLiteral<-2>, z.ZodNumber]>>>;
15
- system_prompt: z.ZodOptional<z.ZodOptional<z.ZodString>>;
15
+ system_prompt: z.ZodOptional<z.ZodDefault<z.ZodString>>;
16
16
  no_mmap: z.ZodOptional<z.ZodOptional<z.ZodBoolean>>;
17
17
  verbosity: z.ZodOptional<z.ZodOptional<z.ZodUnion<readonly [z.ZodLiteral<-1>, z.ZodLiteral<0>, z.ZodLiteral<1>]>>>;
18
18
  presence_penalty: z.ZodOptional<z.ZodOptional<z.ZodNumber>>;
19
19
  frequency_penalty: z.ZodOptional<z.ZodOptional<z.ZodNumber>>;
20
20
  repeat_penalty: z.ZodOptional<z.ZodOptional<z.ZodNumber>>;
21
21
  reverse_prompt: z.ZodOptional<z.ZodOptional<z.ZodString>>;
22
+ n_discarded: z.ZodOptional<z.ZodOptional<z.ZodNumber>>;
22
23
  }, z.core.$strict>>;
23
24
  seed: z.ZodOptional<z.ZodBoolean>;
24
25
  projectionModelSrc: z.ZodOptional<z.ZodString>;
@@ -187,6 +188,7 @@ export declare const loadModelOptionsSchema: z.ZodPipe<z.ZodUnion<readonly [z.Zo
187
188
  frequency_penalty?: number | undefined;
188
189
  repeat_penalty?: number | undefined;
189
190
  reverse_prompt?: string | undefined;
191
+ n_discarded?: number | undefined;
190
192
  } | undefined;
191
193
  projectionModelSrc?: string | undefined;
192
194
  delegate?: {
@@ -271,6 +273,7 @@ export declare const loadModelOptionsSchema: z.ZodPipe<z.ZodUnion<readonly [z.Zo
271
273
  frequency_penalty?: number | undefined;
272
274
  repeat_penalty?: number | undefined;
273
275
  reverse_prompt?: string | undefined;
276
+ n_discarded?: number | undefined;
274
277
  } | undefined;
275
278
  seed?: boolean | undefined;
276
279
  projectionModelSrc?: string | undefined;
@@ -363,13 +366,14 @@ export declare const loadLlmModelRequestSchema: z.ZodObject<{
363
366
  lora: z.ZodOptional<z.ZodString>;
364
367
  device: z.ZodDefault<z.ZodString>;
365
368
  predict: z.ZodOptional<z.ZodUnion<readonly [z.ZodLiteral<-1>, z.ZodLiteral<-2>, z.ZodNumber]>>;
366
- system_prompt: z.ZodOptional<z.ZodString>;
369
+ system_prompt: z.ZodDefault<z.ZodString>;
367
370
  no_mmap: z.ZodOptional<z.ZodBoolean>;
368
371
  verbosity: z.ZodOptional<z.ZodUnion<readonly [z.ZodLiteral<-1>, z.ZodLiteral<0>, z.ZodLiteral<1>]>>;
369
372
  presence_penalty: z.ZodOptional<z.ZodNumber>;
370
373
  frequency_penalty: z.ZodOptional<z.ZodNumber>;
371
374
  repeat_penalty: z.ZodOptional<z.ZodNumber>;
372
375
  reverse_prompt: z.ZodOptional<z.ZodString>;
376
+ n_discarded: z.ZodOptional<z.ZodNumber>;
373
377
  }, z.core.$strip>;
374
378
  }, z.core.$strip>;
375
379
  export declare const loadWhisperModelRequestSchema: z.ZodObject<{
@@ -543,13 +547,14 @@ export declare const loadModelRequestSchema: z.ZodPipe<z.ZodDiscriminatedUnion<[
543
547
  lora: z.ZodOptional<z.ZodString>;
544
548
  device: z.ZodDefault<z.ZodString>;
545
549
  predict: z.ZodOptional<z.ZodUnion<readonly [z.ZodLiteral<-1>, z.ZodLiteral<-2>, z.ZodNumber]>>;
546
- system_prompt: z.ZodOptional<z.ZodString>;
550
+ system_prompt: z.ZodDefault<z.ZodString>;
547
551
  no_mmap: z.ZodOptional<z.ZodBoolean>;
548
552
  verbosity: z.ZodOptional<z.ZodUnion<readonly [z.ZodLiteral<-1>, z.ZodLiteral<0>, z.ZodLiteral<1>]>>;
549
553
  presence_penalty: z.ZodOptional<z.ZodNumber>;
550
554
  frequency_penalty: z.ZodOptional<z.ZodNumber>;
551
555
  repeat_penalty: z.ZodOptional<z.ZodNumber>;
552
556
  reverse_prompt: z.ZodOptional<z.ZodString>;
557
+ n_discarded: z.ZodOptional<z.ZodNumber>;
553
558
  }, z.core.$strip>;
554
559
  }, z.core.$strip>, z.ZodObject<{
555
560
  type: z.ZodLiteral<"loadModel">;
@@ -704,19 +709,20 @@ export declare const loadModelRequestSchema: z.ZodPipe<z.ZodDiscriminatedUnion<[
704
709
  ctx_size: number;
705
710
  gpu_layers: number;
706
711
  device: string;
712
+ system_prompt: string;
707
713
  temp?: number | undefined;
708
714
  top_p?: number | undefined;
709
715
  top_k?: number | undefined;
710
716
  seed?: number | undefined;
711
717
  lora?: string | undefined;
712
718
  predict?: number | undefined;
713
- system_prompt?: string | undefined;
714
719
  no_mmap?: boolean | undefined;
715
720
  verbosity?: 0 | 1 | -1 | undefined;
716
721
  presence_penalty?: number | undefined;
717
722
  frequency_penalty?: number | undefined;
718
723
  repeat_penalty?: number | undefined;
719
724
  reverse_prompt?: string | undefined;
725
+ n_discarded?: number | undefined;
720
726
  };
721
727
  projectionModelSrc?: string | undefined;
722
728
  vadModelSrc?: string | undefined;
@@ -801,19 +807,20 @@ export declare const loadModelRequestSchema: z.ZodPipe<z.ZodDiscriminatedUnion<[
801
807
  ctx_size: number;
802
808
  gpu_layers: number;
803
809
  device: string;
810
+ system_prompt: string;
804
811
  temp?: number | undefined;
805
812
  top_p?: number | undefined;
806
813
  top_k?: number | undefined;
807
814
  seed?: number | undefined;
808
815
  lora?: string | undefined;
809
816
  predict?: number | undefined;
810
- system_prompt?: string | undefined;
811
817
  no_mmap?: boolean | undefined;
812
818
  verbosity?: 0 | 1 | -1 | undefined;
813
819
  presence_penalty?: number | undefined;
814
820
  frequency_penalty?: number | undefined;
815
821
  repeat_penalty?: number | undefined;
816
822
  reverse_prompt?: string | undefined;
823
+ n_discarded?: number | undefined;
817
824
  };
818
825
  projectionModelSrc?: string | undefined;
819
826
  vadModelSrc?: string | undefined;
@@ -925,13 +932,14 @@ export declare const loadModelServerParamsSchema: z.ZodObject<{
925
932
  lora: z.ZodOptional<z.ZodOptional<z.ZodString>>;
926
933
  device: z.ZodOptional<z.ZodDefault<z.ZodString>>;
927
934
  predict: z.ZodOptional<z.ZodOptional<z.ZodUnion<readonly [z.ZodLiteral<-1>, z.ZodLiteral<-2>, z.ZodNumber]>>>;
928
- system_prompt: z.ZodOptional<z.ZodOptional<z.ZodString>>;
935
+ system_prompt: z.ZodOptional<z.ZodDefault<z.ZodString>>;
929
936
  no_mmap: z.ZodOptional<z.ZodOptional<z.ZodBoolean>>;
930
937
  verbosity: z.ZodOptional<z.ZodOptional<z.ZodUnion<readonly [z.ZodLiteral<-1>, z.ZodLiteral<0>, z.ZodLiteral<1>]>>>;
931
938
  presence_penalty: z.ZodOptional<z.ZodOptional<z.ZodNumber>>;
932
939
  frequency_penalty: z.ZodOptional<z.ZodOptional<z.ZodNumber>>;
933
940
  repeat_penalty: z.ZodOptional<z.ZodOptional<z.ZodNumber>>;
934
941
  reverse_prompt: z.ZodOptional<z.ZodOptional<z.ZodString>>;
942
+ n_discarded: z.ZodOptional<z.ZodOptional<z.ZodNumber>>;
935
943
  }, z.core.$strict>>;
936
944
  seed: z.ZodOptional<z.ZodBoolean>;
937
945
  projectionModelSrc: z.ZodOptional<z.ZodString>;
@@ -1100,6 +1108,7 @@ export declare const loadModelServerParamsSchema: z.ZodObject<{
1100
1108
  frequency_penalty?: number | undefined;
1101
1109
  repeat_penalty?: number | undefined;
1102
1110
  reverse_prompt?: string | undefined;
1111
+ n_discarded?: number | undefined;
1103
1112
  } | undefined;
1104
1113
  projectionModelSrc?: string | undefined;
1105
1114
  delegate?: {
@@ -1184,6 +1193,7 @@ export declare const loadModelServerParamsSchema: z.ZodObject<{
1184
1193
  frequency_penalty?: number | undefined;
1185
1194
  repeat_penalty?: number | undefined;
1186
1195
  reverse_prompt?: string | undefined;
1196
+ n_discarded?: number | undefined;
1187
1197
  } | undefined;
1188
1198
  seed?: boolean | undefined;
1189
1199
  projectionModelSrc?: string | undefined;
@@ -25,6 +25,7 @@ declare const ragSaveEmbeddingsParamsSchema: z.ZodObject<{
25
25
  paragraph: "paragraph";
26
26
  }>>;
27
27
  splitStrategy: z.ZodOptional<z.ZodEnum<{
28
+ token: "token";
28
29
  character: "character";
29
30
  word: "word";
30
31
  sentence: "sentence";
@@ -45,6 +46,7 @@ export declare const ragSaveEmbeddingsOperationSchema: z.ZodObject<{
45
46
  paragraph: "paragraph";
46
47
  }>>;
47
48
  splitStrategy: z.ZodOptional<z.ZodEnum<{
49
+ token: "token";
48
50
  character: "character";
49
51
  word: "word";
50
52
  sentence: "sentence";
@@ -95,6 +97,7 @@ export declare const ragRequestSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
95
97
  paragraph: "paragraph";
96
98
  }>>;
97
99
  splitStrategy: z.ZodOptional<z.ZodEnum<{
100
+ token: "token";
98
101
  character: "character";
99
102
  word: "word";
100
103
  sentence: "sentence";