Hugging Face's logo

Spaces:
retopara
/
ragflow
Build error

App Files Community
黄腾 aopstudio commited on
Commit
436c3b5
·
1 Parent(s): 6b089a1

'load llm infomation from a json file and add support for OpenRouter' (#1533)

Browse files

### What problem does this PR solve?

#1467

### Type of change

- [x] New Feature (non-breaking change which adds functionality)

---------

Co-authored-by: Zhedong Cen <[email protected]>

Files changed (8) hide show
  1. api/apps/llm_app.py +2 -2
  2. api/db/init_data.py +19 -894
  3. conf/llm_factories.json +1884 -0
  4. rag/llm/__init__.py +4 -2
  5. rag/llm/chat_model.py +12 -5
  6. rag/llm/cv_model.py +60 -1
  7. web/src/assets/svg/llm/open-router.svg +18 -0
  8. web/src/pages/user-setting/setting-model/index.tsx +1 -0
api/apps/llm_app.py CHANGED
@@ -57,8 +57,8 @@ def set_api_key():
57
  mdl = ChatModel[factory](
58
  req["api_key"], llm.llm_name, base_url=req.get("base_url"))
59
  try:
60
- m, tc = mdl.chat(None, [{"role": "user", "content": "Hello! How are you doing!"}], {
61
- "temperature": 0.9})
62
  if not tc:
63
  raise Exception(m)
64
  except Exception as e:
 
57
  mdl = ChatModel[factory](
58
  req["api_key"], llm.llm_name, base_url=req.get("base_url"))
59
  try:
60
+ m, tc = mdl.chat(None, [{"role": "user", "content": "Hello! How are you doing!"}],
61
+ {"temperature": 0.9,'max_tokens':50})
62
  if not tc:
63
  raise Exception(m)
64
  except Exception as e:
api/db/init_data.py CHANGED
@@ -89,904 +89,29 @@ def init_superuser():
89
  tenant["embd_id"]))
90
 
91
 
92
- factory_infos = [{
93
- "name": "OpenAI",
94
- "logo": "",
95
- "tags": "LLM,TEXT EMBEDDING,SPEECH2TEXT,MODERATION",
96
- "status": "1",
97
- }, {
98
- "name": "Tongyi-Qianwen",
99
- "logo": "",
100
- "tags": "LLM,TEXT EMBEDDING,SPEECH2TEXT,MODERATION",
101
- "status": "1",
102
- }, {
103
- "name": "ZHIPU-AI",
104
- "logo": "",
105
- "tags": "LLM,TEXT EMBEDDING,SPEECH2TEXT,MODERATION",
106
- "status": "1",
107
- },
108
- {
109
- "name": "Ollama",
110
- "logo": "",
111
- "tags": "LLM,TEXT EMBEDDING,SPEECH2TEXT,MODERATION",
112
- "status": "1",
113
- }, {
114
- "name": "Moonshot",
115
- "logo": "",
116
- "tags": "LLM,TEXT EMBEDDING",
117
- "status": "1",
118
- }, {
119
- "name": "FastEmbed",
120
- "logo": "",
121
- "tags": "TEXT EMBEDDING",
122
- "status": "1",
123
- }, {
124
- "name": "Xinference",
125
- "logo": "",
126
- "tags": "LLM,TEXT EMBEDDING,SPEECH2TEXT,MODERATION,TEXT RE-RANK",
127
- "status": "1",
128
- },{
129
- "name": "Youdao",
130
- "logo": "",
131
- "tags": "LLM,TEXT EMBEDDING,SPEECH2TEXT,MODERATION",
132
- "status": "1",
133
- },{
134
- "name": "DeepSeek",
135
- "logo": "",
136
- "tags": "LLM",
137
- "status": "1",
138
- },{
139
- "name": "VolcEngine",
140
- "logo": "",
141
- "tags": "LLM, TEXT EMBEDDING",
142
- "status": "1",
143
- },{
144
- "name": "BaiChuan",
145
- "logo": "",
146
- "tags": "LLM,TEXT EMBEDDING",
147
- "status": "1",
148
- },{
149
- "name": "Jina",
150
- "logo": "",
151
- "tags": "TEXT EMBEDDING, TEXT RE-RANK",
152
- "status": "1",
153
- },{
154
- "name": "BAAI",
155
- "logo": "",
156
- "tags": "TEXT EMBEDDING, TEXT RE-RANK",
157
- "status": "1",
158
- },{
159
- "name": "MiniMax",
160
- "logo": "",
161
- "tags": "LLM,TEXT EMBEDDING",
162
- "status": "1",
163
- },{
164
- "name": "Mistral",
165
- "logo": "",
166
- "tags": "LLM,TEXT EMBEDDING",
167
- "status": "1",
168
- },{
169
- "name": "Azure-OpenAI",
170
- "logo": "",
171
- "tags": "LLM,TEXT EMBEDDING,SPEECH2TEXT,MODERATION",
172
- "status": "1",
173
- },{
174
- "name": "Bedrock",
175
- "logo": "",
176
- "tags": "LLM,TEXT EMBEDDING",
177
- "status": "1",
178
- },{
179
- "name": "Gemini",
180
- "logo": "",
181
- "tags": "LLM,TEXT EMBEDDING,IMAGE2TEXT",
182
- "status": "1",
183
- },
184
- {
185
- "name": "Groq",
186
- "logo": "",
187
- "tags": "LLM",
188
- "status": "1",
189
- }
190
- # {
191
- # "name": "文心一言",
192
- # "logo": "",
193
- # "tags": "LLM,TEXT EMBEDDING,SPEECH2TEXT,MODERATION",
194
- # "status": "1",
195
- # },
196
- ]
197
-
198
-
199
  def init_llm_factory():
200
- llm_infos = [
201
- # ---------------------- OpenAI ------------------------
202
- {
203
- "fid": factory_infos[0]["name"],
204
- "llm_name": "gpt-4o",
205
- "tags": "LLM,CHAT,128K",
206
- "max_tokens": 128000,
207
- "model_type": LLMType.CHAT.value + "," + LLMType.IMAGE2TEXT.value
208
- }, {
209
- "fid": factory_infos[0]["name"],
210
- "llm_name": "gpt-3.5-turbo",
211
- "tags": "LLM,CHAT,4K",
212
- "max_tokens": 4096,
213
- "model_type": LLMType.CHAT.value
214
- }, {
215
- "fid": factory_infos[0]["name"],
216
- "llm_name": "gpt-3.5-turbo-16k-0613",
217
- "tags": "LLM,CHAT,16k",
218
- "max_tokens": 16385,
219
- "model_type": LLMType.CHAT.value
220
- }, {
221
- "fid": factory_infos[0]["name"],
222
- "llm_name": "text-embedding-ada-002",
223
- "tags": "TEXT EMBEDDING,8K",
224
- "max_tokens": 8191,
225
- "model_type": LLMType.EMBEDDING.value
226
- }, {
227
- "fid": factory_infos[0]["name"],
228
- "llm_name": "text-embedding-3-small",
229
- "tags": "TEXT EMBEDDING,8K",
230
- "max_tokens": 8191,
231
- "model_type": LLMType.EMBEDDING.value
232
- }, {
233
- "fid": factory_infos[0]["name"],
234
- "llm_name": "text-embedding-3-large",
235
- "tags": "TEXT EMBEDDING,8K",
236
- "max_tokens": 8191,
237
- "model_type": LLMType.EMBEDDING.value
238
- }, {
239
- "fid": factory_infos[0]["name"],
240
- "llm_name": "whisper-1",
241
- "tags": "SPEECH2TEXT",
242
- "max_tokens": 25 * 1024 * 1024,
243
- "model_type": LLMType.SPEECH2TEXT.value
244
- }, {
245
- "fid": factory_infos[0]["name"],
246
- "llm_name": "gpt-4",
247
- "tags": "LLM,CHAT,8K",
248
- "max_tokens": 8191,
249
- "model_type": LLMType.CHAT.value
250
- }, {
251
- "fid": factory_infos[0]["name"],
252
- "llm_name": "gpt-4-turbo",
253
- "tags": "LLM,CHAT,8K",
254
- "max_tokens": 8191,
255
- "model_type": LLMType.CHAT.value
256
- },{
257
- "fid": factory_infos[0]["name"],
258
- "llm_name": "gpt-4-32k",
259
- "tags": "LLM,CHAT,32K",
260
- "max_tokens": 32768,
261
- "model_type": LLMType.CHAT.value
262
- }, {
263
- "fid": factory_infos[0]["name"],
264
- "llm_name": "gpt-4-vision-preview",
265
- "tags": "LLM,CHAT,IMAGE2TEXT",
266
- "max_tokens": 765,
267
- "model_type": LLMType.IMAGE2TEXT.value
268
- },
269
- # ----------------------- Qwen -----------------------
270
- {
271
- "fid": factory_infos[1]["name"],
272
- "llm_name": "qwen-turbo",
273
- "tags": "LLM,CHAT,8K",
274
- "max_tokens": 8191,
275
- "model_type": LLMType.CHAT.value
276
- }, {
277
- "fid": factory_infos[1]["name"],
278
- "llm_name": "qwen-plus",
279
- "tags": "LLM,CHAT,32K",
280
- "max_tokens": 32768,
281
- "model_type": LLMType.CHAT.value
282
- }, {
283
- "fid": factory_infos[1]["name"],
284
- "llm_name": "qwen-max-1201",
285
- "tags": "LLM,CHAT,6K",
286
- "max_tokens": 5899,
287
- "model_type": LLMType.CHAT.value
288
- }, {
289
- "fid": factory_infos[1]["name"],
290
- "llm_name": "text-embedding-v2",
291
- "tags": "TEXT EMBEDDING,2K",
292
- "max_tokens": 2048,
293
- "model_type": LLMType.EMBEDDING.value
294
- }, {
295
- "fid": factory_infos[1]["name"],
296
- "llm_name": "paraformer-realtime-8k-v1",
297
- "tags": "SPEECH2TEXT",
298
- "max_tokens": 25 * 1024 * 1024,
299
- "model_type": LLMType.SPEECH2TEXT.value
300
- }, {
301
- "fid": factory_infos[1]["name"],
302
- "llm_name": "qwen-vl-max",
303
- "tags": "LLM,CHAT,IMAGE2TEXT",
304
- "max_tokens": 765,
305
- "model_type": LLMType.IMAGE2TEXT.value
306
- },
307
- # ---------------------- ZhipuAI ----------------------
308
- {
309
- "fid": factory_infos[2]["name"],
310
- "llm_name": "glm-3-turbo",
311
- "tags": "LLM,CHAT,",
312
- "max_tokens": 128 * 1000,
313
- "model_type": LLMType.CHAT.value
314
- }, {
315
- "fid": factory_infos[2]["name"],
316
- "llm_name": "glm-4",
317
- "tags": "LLM,CHAT,",
318
- "max_tokens": 128 * 1000,
319
- "model_type": LLMType.CHAT.value
320
- }, {
321
- "fid": factory_infos[2]["name"],
322
- "llm_name": "glm-4v",
323
- "tags": "LLM,CHAT,IMAGE2TEXT",
324
- "max_tokens": 2000,
325
- "model_type": LLMType.IMAGE2TEXT.value
326
- },
327
- {
328
- "fid": factory_infos[2]["name"],
329
- "llm_name": "embedding-2",
330
- "tags": "TEXT EMBEDDING",
331
- "max_tokens": 512,
332
- "model_type": LLMType.EMBEDDING.value
333
- },
334
- # ------------------------ Moonshot -----------------------
335
- {
336
- "fid": factory_infos[4]["name"],
337
- "llm_name": "moonshot-v1-8k",
338
- "tags": "LLM,CHAT,",
339
- "max_tokens": 7900,
340
- "model_type": LLMType.CHAT.value
341
- }, {
342
- "fid": factory_infos[4]["name"],
343
- "llm_name": "moonshot-v1-32k",
344
- "tags": "LLM,CHAT,",
345
- "max_tokens": 32768,
346
- "model_type": LLMType.CHAT.value
347
- }, {
348
- "fid": factory_infos[4]["name"],
349
- "llm_name": "moonshot-v1-128k",
350
- "tags": "LLM,CHAT",
351
- "max_tokens": 128 * 1000,
352
- "model_type": LLMType.CHAT.value
353
- },
354
- # ------------------------ FastEmbed -----------------------
355
- {
356
- "fid": factory_infos[5]["name"],
357
- "llm_name": "BAAI/bge-small-en-v1.5",
358
- "tags": "TEXT EMBEDDING,",
359
- "max_tokens": 512,
360
- "model_type": LLMType.EMBEDDING.value
361
- }, {
362
- "fid": factory_infos[5]["name"],
363
- "llm_name": "BAAI/bge-small-zh-v1.5",
364
- "tags": "TEXT EMBEDDING,",
365
- "max_tokens": 512,
366
- "model_type": LLMType.EMBEDDING.value
367
- }, {
368
- }, {
369
- "fid": factory_infos[5]["name"],
370
- "llm_name": "BAAI/bge-base-en-v1.5",
371
- "tags": "TEXT EMBEDDING,",
372
- "max_tokens": 512,
373
- "model_type": LLMType.EMBEDDING.value
374
- }, {
375
- }, {
376
- "fid": factory_infos[5]["name"],
377
- "llm_name": "BAAI/bge-large-en-v1.5",
378
- "tags": "TEXT EMBEDDING,",
379
- "max_tokens": 512,
380
- "model_type": LLMType.EMBEDDING.value
381
- }, {
382
- "fid": factory_infos[5]["name"],
383
- "llm_name": "sentence-transformers/all-MiniLM-L6-v2",
384
- "tags": "TEXT EMBEDDING,",
385
- "max_tokens": 512,
386
- "model_type": LLMType.EMBEDDING.value
387
- }, {
388
- "fid": factory_infos[5]["name"],
389
- "llm_name": "nomic-ai/nomic-embed-text-v1.5",
390
- "tags": "TEXT EMBEDDING,",
391
- "max_tokens": 8192,
392
- "model_type": LLMType.EMBEDDING.value
393
- }, {
394
- "fid": factory_infos[5]["name"],
395
- "llm_name": "jinaai/jina-embeddings-v2-small-en",
396
- "tags": "TEXT EMBEDDING,",
397
- "max_tokens": 2147483648,
398
- "model_type": LLMType.EMBEDDING.value
399
- }, {
400
- "fid": factory_infos[5]["name"],
401
- "llm_name": "jinaai/jina-embeddings-v2-base-en",
402
- "tags": "TEXT EMBEDDING,",
403
- "max_tokens": 2147483648,
404
- "model_type": LLMType.EMBEDDING.value
405
- },
406
- # ------------------------ Youdao -----------------------
407
- {
408
- "fid": factory_infos[7]["name"],
409
- "llm_name": "maidalun1020/bce-embedding-base_v1",
410
- "tags": "TEXT EMBEDDING,",
411
- "max_tokens": 512,
412
- "model_type": LLMType.EMBEDDING.value
413
- },
414
- {
415
- "fid": factory_infos[7]["name"],
416
- "llm_name": "maidalun1020/bce-reranker-base_v1",
417
- "tags": "RE-RANK, 512",
418
- "max_tokens": 512,
419
- "model_type": LLMType.RERANK.value
420
- },
421
- # ------------------------ DeepSeek -----------------------
422
- {
423
- "fid": factory_infos[8]["name"],
424
- "llm_name": "deepseek-chat",
425
- "tags": "LLM,CHAT,",
426
- "max_tokens": 32768,
427
- "model_type": LLMType.CHAT.value
428
- },
429
- {
430
- "fid": factory_infos[8]["name"],
431
- "llm_name": "deepseek-coder",
432
- "tags": "LLM,CHAT,",
433
- "max_tokens": 16385,
434
- "model_type": LLMType.CHAT.value
435
- },
436
- # ------------------------ VolcEngine -----------------------
437
- {
438
- "fid": factory_infos[9]["name"],
439
- "llm_name": "Skylark2-pro-32k",
440
- "tags": "LLM,CHAT,32k",
441
- "max_tokens": 32768,
442
- "model_type": LLMType.CHAT.value
443
- },
444
- {
445
- "fid": factory_infos[9]["name"],
446
- "llm_name": "Skylark2-pro-4k",
447
- "tags": "LLM,CHAT,4k",
448
- "max_tokens": 4096,
449
- "model_type": LLMType.CHAT.value
450
- },
451
- # ------------------------ BaiChuan -----------------------
452
- {
453
- "fid": factory_infos[10]["name"],
454
- "llm_name": "Baichuan2-Turbo",
455
- "tags": "LLM,CHAT,32K",
456
- "max_tokens": 32768,
457
- "model_type": LLMType.CHAT.value
458
- },
459
- {
460
- "fid": factory_infos[10]["name"],
461
- "llm_name": "Baichuan2-Turbo-192k",
462
- "tags": "LLM,CHAT,192K",
463
- "max_tokens": 196608,
464
- "model_type": LLMType.CHAT.value
465
- },
466
- {
467
- "fid": factory_infos[10]["name"],
468
- "llm_name": "Baichuan3-Turbo",
469
- "tags": "LLM,CHAT,32K",
470
- "max_tokens": 32768,
471
- "model_type": LLMType.CHAT.value
472
- },
473
- {
474
- "fid": factory_infos[10]["name"],
475
- "llm_name": "Baichuan3-Turbo-128k",
476
- "tags": "LLM,CHAT,128K",
477
- "max_tokens": 131072,
478
- "model_type": LLMType.CHAT.value
479
- },
480
- {
481
- "fid": factory_infos[10]["name"],
482
- "llm_name": "Baichuan4",
483
- "tags": "LLM,CHAT,128K",
484
- "max_tokens": 131072,
485
- "model_type": LLMType.CHAT.value
486
- },
487
- {
488
- "fid": factory_infos[10]["name"],
489
- "llm_name": "Baichuan-Text-Embedding",
490
- "tags": "TEXT EMBEDDING",
491
- "max_tokens": 512,
492
- "model_type": LLMType.EMBEDDING.value
493
- },
494
- # ------------------------ Jina -----------------------
495
- {
496
- "fid": factory_infos[11]["name"],
497
- "llm_name": "jina-reranker-v1-base-en",
498
- "tags": "RE-RANK,8k",
499
- "max_tokens": 8196,
500
- "model_type": LLMType.RERANK.value
501
- },
502
- {
503
- "fid": factory_infos[11]["name"],
504
- "llm_name": "jina-reranker-v1-turbo-en",
505
- "tags": "RE-RANK,8k",
506
- "max_tokens": 8196,
507
- "model_type": LLMType.RERANK.value
508
- },
509
- {
510
- "fid": factory_infos[11]["name"],
511
- "llm_name": "jina-reranker-v1-tiny-en",
512
- "tags": "RE-RANK,8k",
513
- "max_tokens": 8196,
514
- "model_type": LLMType.RERANK.value
515
- },
516
- {
517
- "fid": factory_infos[11]["name"],
518
- "llm_name": "jina-colbert-v1-en",
519
- "tags": "RE-RANK,8k",
520
- "max_tokens": 8196,
521
- "model_type": LLMType.RERANK.value
522
- },
523
- {
524
- "fid": factory_infos[11]["name"],
525
- "llm_name": "jina-embeddings-v2-base-en",
526
- "tags": "TEXT EMBEDDING",
527
- "max_tokens": 8196,
528
- "model_type": LLMType.EMBEDDING.value
529
- },
530
- {
531
- "fid": factory_infos[11]["name"],
532
- "llm_name": "jina-embeddings-v2-base-de",
533
- "tags": "TEXT EMBEDDING",
534
- "max_tokens": 8196,
535
- "model_type": LLMType.EMBEDDING.value
536
- },
537
- {
538
- "fid": factory_infos[11]["name"],
539
- "llm_name": "jina-embeddings-v2-base-es",
540
- "tags": "TEXT EMBEDDING",
541
- "max_tokens": 8196,
542
- "model_type": LLMType.EMBEDDING.value
543
- },
544
- {
545
- "fid": factory_infos[11]["name"],
546
- "llm_name": "jina-embeddings-v2-base-code",
547
- "tags": "TEXT EMBEDDING",
548
- "max_tokens": 8196,
549
- "model_type": LLMType.EMBEDDING.value
550
- },
551
- {
552
- "fid": factory_infos[11]["name"],
553
- "llm_name": "jina-embeddings-v2-base-zh",
554
- "tags": "TEXT EMBEDDING",
555
- "max_tokens": 8196,
556
- "model_type": LLMType.EMBEDDING.value
557
- },
558
- # ------------------------ BAAI -----------------------
559
- {
560
- "fid": factory_infos[12]["name"],
561
- "llm_name": "BAAI/bge-large-zh-v1.5",
562
- "tags": "TEXT EMBEDDING,",
563
- "max_tokens": 1024,
564
- "model_type": LLMType.EMBEDDING.value
565
- },
566
- {
567
- "fid": factory_infos[12]["name"],
568
- "llm_name": "BAAI/bge-reranker-v2-m3",
569
- "tags": "RE-RANK,2k",
570
- "max_tokens": 2048,
571
- "model_type": LLMType.RERANK.value
572
- },
573
- # ------------------------ Minimax -----------------------
574
- {
575
- "fid": factory_infos[13]["name"],
576
- "llm_name": "abab6.5",
577
- "tags": "LLM,CHAT,8k",
578
- "max_tokens": 8192,
579
- "model_type": LLMType.CHAT.value
580
- },
581
- {
582
- "fid": factory_infos[13]["name"],
583
- "llm_name": "abab6.5s",
584
- "tags": "LLM,CHAT,245k",
585
- "max_tokens": 245760,
586
- "model_type": LLMType.CHAT.value
587
- },
588
- {
589
- "fid": factory_infos[13]["name"],
590
- "llm_name": "abab6.5t",
591
- "tags": "LLM,CHAT,8k",
592
- "max_tokens": 8192,
593
- "model_type": LLMType.CHAT.value
594
- },
595
- {
596
- "fid": factory_infos[13]["name"],
597
- "llm_name": "abab6.5g",
598
- "tags": "LLM,CHAT,8k",
599
- "max_tokens": 8192,
600
- "model_type": LLMType.CHAT.value
601
- },
602
- {
603
- "fid": factory_infos[13]["name"],
604
- "llm_name": "abab5.5s",
605
- "tags": "LLM,CHAT,8k",
606
- "max_tokens": 8192,
607
- "model_type": LLMType.CHAT.value
608
- },
609
- # ------------------------ Mistral -----------------------
610
- {
611
- "fid": factory_infos[14]["name"],
612
- "llm_name": "open-mixtral-8x22b",
613
- "tags": "LLM,CHAT,64k",
614
- "max_tokens": 64000,
615
- "model_type": LLMType.CHAT.value
616
- },
617
- {
618
- "fid": factory_infos[14]["name"],
619
- "llm_name": "open-mixtral-8x7b",
620
- "tags": "LLM,CHAT,32k",
621
- "max_tokens": 32000,
622
- "model_type": LLMType.CHAT.value
623
- },
624
- {
625
- "fid": factory_infos[14]["name"],
626
- "llm_name": "open-mistral-7b",
627
- "tags": "LLM,CHAT,32k",
628
- "max_tokens": 32000,
629
- "model_type": LLMType.CHAT.value
630
- },
631
- {
632
- "fid": factory_infos[14]["name"],
633
- "llm_name": "mistral-large-latest",
634
- "tags": "LLM,CHAT,32k",
635
- "max_tokens": 32000,
636
- "model_type": LLMType.CHAT.value
637
- },
638
- {
639
- "fid": factory_infos[14]["name"],
640
- "llm_name": "mistral-small-latest",
641
- "tags": "LLM,CHAT,32k",
642
- "max_tokens": 32000,
643
- "model_type": LLMType.CHAT.value
644
- },
645
- {
646
- "fid": factory_infos[14]["name"],
647
- "llm_name": "mistral-medium-latest",
648
- "tags": "LLM,CHAT,32k",
649
- "max_tokens": 32000,
650
- "model_type": LLMType.CHAT.value
651
- },
652
- {
653
- "fid": factory_infos[14]["name"],
654
- "llm_name": "codestral-latest",
655
- "tags": "LLM,CHAT,32k",
656
- "max_tokens": 32000,
657
- "model_type": LLMType.CHAT.value
658
- },
659
- {
660
- "fid": factory_infos[14]["name"],
661
- "llm_name": "mistral-embed",
662
- "tags": "LLM,CHAT,8k",
663
- "max_tokens": 8192,
664
- "model_type": LLMType.EMBEDDING
665
- },
666
- # ------------------------ Azure OpenAI -----------------------
667
- # Please ensure the llm_name is the same as the name in Azure
668
- # OpenAI deployment name (e.g., azure-gpt-4o). And the llm_name
669
- # must different from the OpenAI llm_name
670
- #
671
- # Each model must be deployed in the Azure OpenAI service, otherwise,
672
- # you will receive an error message 'The API deployment for
673
- # this resource does not exist'
674
- {
675
- "fid": factory_infos[15]["name"],
676
- "llm_name": "azure-gpt-4o",
677
- "tags": "LLM,CHAT,128K",
678
- "max_tokens": 128000,
679
- "model_type": LLMType.CHAT.value + "," + LLMType.IMAGE2TEXT.value
680
- }, {
681
- "fid": factory_infos[15]["name"],
682
- "llm_name": "azure-gpt-35-turbo",
683
- "tags": "LLM,CHAT,4K",
684
- "max_tokens": 4096,
685
- "model_type": LLMType.CHAT.value
686
- }, {
687
- "fid": factory_infos[15]["name"],
688
- "llm_name": "azure-gpt-35-turbo-16k",
689
- "tags": "LLM,CHAT,16k",
690
- "max_tokens": 16385,
691
- "model_type": LLMType.CHAT.value
692
- }, {
693
- "fid": factory_infos[15]["name"],
694
- "llm_name": "azure-text-embedding-ada-002",
695
- "tags": "TEXT EMBEDDING,8K",
696
- "max_tokens": 8191,
697
- "model_type": LLMType.EMBEDDING.value
698
- }, {
699
- "fid": factory_infos[15]["name"],
700
- "llm_name": "azure-text-embedding-3-small",
701
- "tags": "TEXT EMBEDDING,8K",
702
- "max_tokens": 8191,
703
- "model_type": LLMType.EMBEDDING.value
704
- }, {
705
- "fid": factory_infos[15]["name"],
706
- "llm_name": "azure-text-embedding-3-large",
707
- "tags": "TEXT EMBEDDING,8K",
708
- "max_tokens": 8191,
709
- "model_type": LLMType.EMBEDDING.value
710
- },{
711
- "fid": factory_infos[15]["name"],
712
- "llm_name": "azure-whisper-1",
713
- "tags": "SPEECH2TEXT",
714
- "max_tokens": 25 * 1024 * 1024,
715
- "model_type": LLMType.SPEECH2TEXT.value
716
- },
717
- {
718
- "fid": factory_infos[15]["name"],
719
- "llm_name": "azure-gpt-4",
720
- "tags": "LLM,CHAT,8K",
721
- "max_tokens": 8191,
722
- "model_type": LLMType.CHAT.value
723
- }, {
724
- "fid": factory_infos[15]["name"],
725
- "llm_name": "azure-gpt-4-turbo",
726
- "tags": "LLM,CHAT,8K",
727
- "max_tokens": 8191,
728
- "model_type": LLMType.CHAT.value
729
- }, {
730
- "fid": factory_infos[15]["name"],
731
- "llm_name": "azure-gpt-4-32k",
732
- "tags": "LLM,CHAT,32K",
733
- "max_tokens": 32768,
734
- "model_type": LLMType.CHAT.value
735
- }, {
736
- "fid": factory_infos[15]["name"],
737
- "llm_name": "azure-gpt-4-vision-preview",
738
- "tags": "LLM,CHAT,IMAGE2TEXT",
739
- "max_tokens": 765,
740
- "model_type": LLMType.IMAGE2TEXT.value
741
- },
742
- # ------------------------ Bedrock -----------------------
743
- {
744
- "fid": factory_infos[16]["name"],
745
- "llm_name": "ai21.j2-ultra-v1",
746
- "tags": "LLM,CHAT,8k",
747
- "max_tokens": 8191,
748
- "model_type": LLMType.CHAT.value
749
- }, {
750
- "fid": factory_infos[16]["name"],
751
- "llm_name": "ai21.j2-mid-v1",
752
- "tags": "LLM,CHAT,8k",
753
- "max_tokens": 8191,
754
- "model_type": LLMType.CHAT.value
755
- }, {
756
- "fid": factory_infos[16]["name"],
757
- "llm_name": "cohere.command-text-v14",
758
- "tags": "LLM,CHAT,4k",
759
- "max_tokens": 4096,
760
- "model_type": LLMType.CHAT.value
761
- }, {
762
- "fid": factory_infos[16]["name"],
763
- "llm_name": "cohere.command-light-text-v14",
764
- "tags": "LLM,CHAT,4k",
765
- "max_tokens": 4096,
766
- "model_type": LLMType.CHAT.value
767
- }, {
768
- "fid": factory_infos[16]["name"],
769
- "llm_name": "cohere.command-r-v1:0",
770
- "tags": "LLM,CHAT,128k",
771
- "max_tokens": 128 * 1024,
772
- "model_type": LLMType.CHAT.value
773
- }, {
774
- "fid": factory_infos[16]["name"],
775
- "llm_name": "cohere.command-r-plus-v1:0",
776
- "tags": "LLM,CHAT,128k",
777
- "max_tokens": 128000,
778
- "model_type": LLMType.CHAT.value
779
- }, {
780
- "fid": factory_infos[16]["name"],
781
- "llm_name": "anthropic.claude-v2",
782
- "tags": "LLM,CHAT,100k",
783
- "max_tokens": 100 * 1024,
784
- "model_type": LLMType.CHAT.value
785
- }, {
786
- "fid": factory_infos[16]["name"],
787
- "llm_name": "anthropic.claude-v2:1",
788
- "tags": "LLM,CHAT,200k",
789
- "max_tokens": 200 * 1024,
790
- "model_type": LLMType.CHAT.value
791
- }, {
792
- "fid": factory_infos[16]["name"],
793
- "llm_name": "anthropic.claude-3-sonnet-20240229-v1:0",
794
- "tags": "LLM,CHAT,200k",
795
- "max_tokens": 200 * 1024,
796
- "model_type": LLMType.CHAT.value
797
- }, {
798
- "fid": factory_infos[16]["name"],
799
- "llm_name": "anthropic.claude-3-5-sonnet-20240620-v1:0",
800
- "tags": "LLM,CHAT,200k",
801
- "max_tokens": 200 * 1024,
802
- "model_type": LLMType.CHAT.value
803
- }, {
804
- "fid": factory_infos[16]["name"],
805
- "llm_name": "anthropic.claude-3-haiku-20240307-v1:0",
806
- "tags": "LLM,CHAT,200k",
807
- "max_tokens": 200 * 1024,
808
- "model_type": LLMType.CHAT.value
809
- }, {
810
- "fid": factory_infos[16]["name"],
811
- "llm_name": "anthropic.claude-3-opus-20240229-v1:0",
812
- "tags": "LLM,CHAT,200k",
813
- "max_tokens": 200 * 1024,
814
- "model_type": LLMType.CHAT.value
815
- }, {
816
- "fid": factory_infos[16]["name"],
817
- "llm_name": "anthropic.claude-instant-v1",
818
- "tags": "LLM,CHAT,100k",
819
- "max_tokens": 100 * 1024,
820
- "model_type": LLMType.CHAT.value
821
- }, {
822
- "fid": factory_infos[16]["name"],
823
- "llm_name": "amazon.titan-text-express-v1",
824
- "tags": "LLM,CHAT,8k",
825
- "max_tokens": 8192,
826
- "model_type": LLMType.CHAT.value
827
- }, {
828
- "fid": factory_infos[16]["name"],
829
- "llm_name": "amazon.titan-text-premier-v1:0",
830
- "tags": "LLM,CHAT,32k",
831
- "max_tokens": 32 * 1024,
832
- "model_type": LLMType.CHAT.value
833
- }, {
834
- "fid": factory_infos[16]["name"],
835
- "llm_name": "amazon.titan-text-lite-v1",
836
- "tags": "LLM,CHAT,4k",
837
- "max_tokens": 4096,
838
- "model_type": LLMType.CHAT.value
839
- }, {
840
- "fid": factory_infos[16]["name"],
841
- "llm_name": "meta.llama2-13b-chat-v1",
842
- "tags": "LLM,CHAT,4k",
843
- "max_tokens": 4096,
844
- "model_type": LLMType.CHAT.value
845
- }, {
846
- "fid": factory_infos[16]["name"],
847
- "llm_name": "meta.llama2-70b-chat-v1",
848
- "tags": "LLM,CHAT,4k",
849
- "max_tokens": 4096,
850
- "model_type": LLMType.CHAT.value
851
- }, {
852
- "fid": factory_infos[16]["name"],
853
- "llm_name": "meta.llama3-8b-instruct-v1:0",
854
- "tags": "LLM,CHAT,8k",
855
- "max_tokens": 8192,
856
- "model_type": LLMType.CHAT.value
857
- }, {
858
- "fid": factory_infos[16]["name"],
859
- "llm_name": "meta.llama3-70b-instruct-v1:0",
860
- "tags": "LLM,CHAT,8k",
861
- "max_tokens": 8192,
862
- "model_type": LLMType.CHAT.value
863
- }, {
864
- "fid": factory_infos[16]["name"],
865
- "llm_name": "mistral.mistral-7b-instruct-v0:2",
866
- "tags": "LLM,CHAT,8k",
867
- "max_tokens": 8192,
868
- "model_type": LLMType.CHAT.value
869
- }, {
870
- "fid": factory_infos[16]["name"],
871
- "llm_name": "mistral.mixtral-8x7b-instruct-v0:1",
872
- "tags": "LLM,CHAT,4k",
873
- "max_tokens": 4096,
874
- "model_type": LLMType.CHAT.value
875
- }, {
876
- "fid": factory_infos[16]["name"],
877
- "llm_name": "mistral.mistral-large-2402-v1:0",
878
- "tags": "LLM,CHAT,8k",
879
- "max_tokens": 8192,
880
- "model_type": LLMType.CHAT.value
881
- }, {
882
- "fid": factory_infos[16]["name"],
883
- "llm_name": "mistral.mistral-small-2402-v1:0",
884
- "tags": "LLM,CHAT,8k",
885
- "max_tokens": 8192,
886
- "model_type": LLMType.CHAT.value
887
- }, {
888
- "fid": factory_infos[16]["name"],
889
- "llm_name": "amazon.titan-embed-text-v2:0",
890
- "tags": "TEXT EMBEDDING",
891
- "max_tokens": 8192,
892
- "model_type": LLMType.EMBEDDING.value
893
- }, {
894
- "fid": factory_infos[16]["name"],
895
- "llm_name": "cohere.embed-english-v3",
896
- "tags": "TEXT EMBEDDING",
897
- "max_tokens": 2048,
898
- "model_type": LLMType.EMBEDDING.value
899
- }, {
900
- "fid": factory_infos[16]["name"],
901
- "llm_name": "cohere.embed-multilingual-v3",
902
- "tags": "TEXT EMBEDDING",
903
- "max_tokens": 2048,
904
- "model_type": LLMType.EMBEDDING.value
905
- }, {
906
- "fid": factory_infos[17]["name"],
907
- "llm_name": "gemini-1.5-pro-latest",
908
- "tags": "LLM,CHAT,1024K",
909
- "max_tokens": 1024*1024,
910
- "model_type": LLMType.CHAT.value
911
- }, {
912
- "fid": factory_infos[17]["name"],
913
- "llm_name": "gemini-1.5-flash-latest",
914
- "tags": "LLM,CHAT,1024K",
915
- "max_tokens": 1024*1024,
916
- "model_type": LLMType.CHAT.value
917
- }, {
918
- "fid": factory_infos[17]["name"],
919
- "llm_name": "gemini-1.0-pro",
920
- "tags": "LLM,CHAT,30K",
921
- "max_tokens": 30*1024,
922
- "model_type": LLMType.CHAT.value
923
- }, {
924
- "fid": factory_infos[17]["name"],
925
- "llm_name": "gemini-1.0-pro-vision-latest",
926
- "tags": "LLM,IMAGE2TEXT,12K",
927
- "max_tokens": 12*1024,
928
- "model_type": LLMType.IMAGE2TEXT.value
929
- }, {
930
- "fid": factory_infos[17]["name"],
931
- "llm_name": "text-embedding-004",
932
- "tags": "TEXT EMBEDDING",
933
- "max_tokens": 2048,
934
- "model_type": LLMType.EMBEDDING.value
935
- },
936
- # ------------------------ Groq -----------------------
937
- {
938
- "fid": factory_infos[18]["name"],
939
- "llm_name": "gemma-7b-it",
940
- "tags": "LLM,CHAT,15k",
941
- "max_tokens": 8192,
942
-
943
- "model_type": LLMType.CHAT.value
944
- },
945
- {
946
- "fid": factory_infos[18]["name"],
947
- "llm_name": "gemma2-9b-it",
948
- "tags": "LLM,CHAT,15k",
949
- "max_tokens": 8192,
950
-
951
- "model_type": LLMType.CHAT.value
952
- },
953
- {
954
- "fid": factory_infos[18]["name"],
955
- "llm_name": "llama3-70b-8192",
956
- "tags": "LLM,CHAT,6k",
957
- "max_tokens": 8192,
958
-
959
- "model_type": LLMType.CHAT.value
960
- },
961
- {
962
- "fid": factory_infos[18]["name"],
963
- "llm_name": "llama3-8b-8192",
964
- "tags": "LLM,CHAT,30k",
965
- "max_tokens": 8192,
966
-
967
- "model_type": LLMType.CHAT.value
968
- },
969
- {
970
- "fid": factory_infos[18]["name"],
971
- "llm_name": "mixtral-8x7b-32768",
972
- "tags": "LLM,CHAT,5k",
973
- "max_tokens": 32768,
974
-
975
- "model_type": LLMType.CHAT.value
976
- }
977
- ]
978
- for info in factory_infos:
979
  try:
980
- LLMFactoriesService.save(**info)
981
- except Exception as e:
982
- pass
983
-
984
- LLMService.filter_delete([(LLM.fid == "MiniMax" or LLM.fid == "Minimax")])
985
- for info in llm_infos:
986
- try:
987
- LLMService.save(**info)
988
  except Exception as e:
989
  pass
 
 
 
 
 
 
 
 
 
 
990
 
991
  LLMFactoriesService.filter_delete([LLMFactories.name == "Local"])
992
  LLMService.filter_delete([LLM.fid == "Local"])
 
89
  tenant["embd_id"]))
90
 
91
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
92
  def init_llm_factory():
93
+ factory_llm_infos = json.load(
94
+ open(
95
+ os.path.join(get_project_base_directory(), "conf", "llm_factories.json"),
96
+ "r",
97
+ )
98
+ )
99
+ for factory_llm_info in factory_llm_infos["factory_llm_infos"]:
100
+ llm_infos = factory_llm_info.pop("llm")
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
101
  try:
102
+ LLMFactoriesService.save(**factory_llm_info)
 
 
 
 
 
 
 
103
  except Exception as e:
104
  pass
105
+ for llm_info in llm_infos:
106
+ llm_info["fid"] = factory_llm_info["name"]
107
+ try:
108
+ LLMService.save(**llm_info)
109
+ except Exception as e:
110
+ pass
111
+ try:
112
+ LLMService.filter_delete([(LLM.fid == "MiniMax" or LLM.fid == "Minimax")])
113
+ except Exception as e:
114
+ pass
115
 
116
  LLMFactoriesService.filter_delete([LLMFactories.name == "Local"])
117
  LLMService.filter_delete([LLM.fid == "Local"])
conf/llm_factories.json ADDED
@@ -0,0 +1,1884 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "factory_llm_infos": [
3
+ {
4
+ "name": "OpenAI",
5
+ "logo": "",
6
+ "tags": "LLM,TEXT EMBEDDING,SPEECH2TEXT,MODERATION",
7
+ "status": "1",
8
+ "llm": [
9
+ {
10
+ "llm_name": "gpt-4o",
11
+ "tags": "LLM,CHAT,128K",
12
+ "max_tokens": 128000,
13
+ "model_type": "chat,image2text"
14
+ },
15
+ {
16
+ "llm_name": "gpt-3.5-turbo",
17
+ "tags": "LLM,CHAT,4K",
18
+ "max_tokens": 4096,
19
+ "model_type": "chat"
20
+ },
21
+ {
22
+ "llm_name": "gpt-3.5-turbo-16k-0613",
23
+ "tags": "LLM,CHAT,16k",
24
+ "max_tokens": 16385,
25
+ "model_type": "chat"
26
+ },
27
+ {
28
+ "llm_name": "text-embedding-ada-002",
29
+ "tags": "TEXT EMBEDDING,8K",
30
+ "max_tokens": 8191,
31
+ "model_type": "embedding"
32
+ },
33
+ {
34
+ "llm_name": "text-embedding-3-small",
35
+ "tags": "TEXT EMBEDDING,8K",
36
+ "max_tokens": 8191,
37
+ "model_type": "embedding"
38
+ },
39
+ {
40
+ "llm_name": "text-embedding-3-large",
41
+ "tags": "TEXT EMBEDDING,8K",
42
+ "max_tokens": 8191,
43
+ "model_type": "embedding"
44
+ },
45
+ {
46
+ "llm_name": "whisper-1",
47
+ "tags": "SPEECH2TEXT",
48
+ "max_tokens": 26214400,
49
+ "model_type": "speech2text"
50
+ },
51
+ {
52
+ "llm_name": "gpt-4",
53
+ "tags": "LLM,CHAT,8K",
54
+ "max_tokens": 8191,
55
+ "model_type": "chat"
56
+ },
57
+ {
58
+ "llm_name": "gpt-4-turbo",
59
+ "tags": "LLM,CHAT,8K",
60
+ "max_tokens": 8191,
61
+ "model_type": "chat"
62
+ },
63
+ {
64
+ "llm_name": "gpt-4-32k",
65
+ "tags": "LLM,CHAT,32K",
66
+ "max_tokens": 32768,
67
+ "model_type": "chat"
68
+ },
69
+ {
70
+ "llm_name": "gpt-4-vision-preview",
71
+ "tags": "LLM,CHAT,IMAGE2TEXT",
72
+ "max_tokens": 765,
73
+ "model_type": "image2text"
74
+ }
75
+ ]
76
+ },
77
+ {
78
+ "name": "Tongyi-Qianwen",
79
+ "logo": "",
80
+ "tags": "LLM,TEXT EMBEDDING,SPEECH2TEXT,MODERATION",
81
+ "status": "1",
82
+ "llm": [
83
+ {
84
+ "llm_name": "qwen-turbo",
85
+ "tags": "LLM,CHAT,8K",
86
+ "max_tokens": 8191,
87
+ "model_type": "chat"
88
+ },
89
+ {
90
+ "llm_name": "qwen-plus",
91
+ "tags": "LLM,CHAT,32K",
92
+ "max_tokens": 32768,
93
+ "model_type": "chat"
94
+ },
95
+ {
96
+ "llm_name": "qwen-max-1201",
97
+ "tags": "LLM,CHAT,6K",
98
+ "max_tokens": 5899,
99
+ "model_type": "chat"
100
+ },
101
+ {
102
+ "llm_name": "text-embedding-v2",
103
+ "tags": "TEXT EMBEDDING,2K",
104
+ "max_tokens": 2048,
105
+ "model_type": "embedding"
106
+ },
107
+ {
108
+ "llm_name": "paraformer-realtime-8k-v1",
109
+ "tags": "SPEECH2TEXT",
110
+ "max_tokens": 26214400,
111
+ "model_type": "speech2text"
112
+ },
113
+ {
114
+ "llm_name": "qwen-vl-max",
115
+ "tags": "LLM,CHAT,IMAGE2TEXT",
116
+ "max_tokens": 765,
117
+ "model_type": "image2text"
118
+ }
119
+ ]
120
+ },
121
+ {
122
+ "name": "ZHIPU-AI",
123
+ "logo": "",
124
+ "tags": "LLM,TEXT EMBEDDING,SPEECH2TEXT,MODERATION",
125
+ "status": "1",
126
+ "llm": [
127
+ {
128
+ "llm_name": "glm-3-turbo",
129
+ "tags": "LLM,CHAT,",
130
+ "max_tokens": 128000,
131
+ "model_type": "chat"
132
+ },
133
+ {
134
+ "llm_name": "glm-4",
135
+ "tags": "LLM,CHAT,",
136
+ "max_tokens": 128000,
137
+ "model_type": "chat"
138
+ },
139
+ {
140
+ "llm_name": "glm-4v",
141
+ "tags": "LLM,CHAT,IMAGE2TEXT",
142
+ "max_tokens": 2000,
143
+ "model_type": "image2text"
144
+ },
145
+ {
146
+ "llm_name": "embedding-2",
147
+ "tags": "TEXT EMBEDDING",
148
+ "max_tokens": 512,
149
+ "model_type": "embedding"
150
+ }
151
+ ]
152
+ },
153
+ {
154
+ "name": "Ollama",
155
+ "logo": "",
156
+ "tags": "LLM,TEXT EMBEDDING,SPEECH2TEXT,MODERATION",
157
+ "status": "1",
158
+ "llm": []
159
+ },
160
+ {
161
+ "name": "Moonshot",
162
+ "logo": "",
163
+ "tags": "LLM,TEXT EMBEDDING",
164
+ "status": "1",
165
+ "llm": [
166
+ {
167
+ "llm_name": "moonshot-v1-8k",
168
+ "tags": "LLM,CHAT,",
169
+ "max_tokens": 7900,
170
+ "model_type": "chat"
171
+ },
172
+ {
173
+ "llm_name": "moonshot-v1-32k",
174
+ "tags": "LLM,CHAT,",
175
+ "max_tokens": 32768,
176
+ "model_type": "chat"
177
+ },
178
+ {
179
+ "llm_name": "moonshot-v1-128k",
180
+ "tags": "LLM,CHAT",
181
+ "max_tokens": 128000,
182
+ "model_type": "chat"
183
+ }
184
+ ]
185
+ },
186
+ {
187
+ "name": "FastEmbed",
188
+ "logo": "",
189
+ "tags": "TEXT EMBEDDING",
190
+ "status": "1",
191
+ "llm": [
192
+ {
193
+ "llm_name": "BAAI/bge-small-en-v1.5",
194
+ "tags": "TEXT EMBEDDING,",
195
+ "max_tokens": 512,
196
+ "model_type": "embedding"
197
+ },
198
+ {
199
+ "llm_name": "BAAI/bge-small-zh-v1.5",
200
+ "tags": "TEXT EMBEDDING,",
201
+ "max_tokens": 512,
202
+ "model_type": "embedding"
203
+ },
204
+ {
205
+ "llm_name": "BAAI/bge-base-en-v1.5",
206
+ "tags": "TEXT EMBEDDING,",
207
+ "max_tokens": 512,
208
+ "model_type": "embedding"
209
+ },
210
+ {
211
+ "llm_name": "BAAI/bge-large-en-v1.5",
212
+ "tags": "TEXT EMBEDDING,",
213
+ "max_tokens": 512,
214
+ "model_type": "embedding"
215
+ },
216
+ {
217
+ "llm_name": "sentence-transformers/all-MiniLM-L6-v2",
218
+ "tags": "TEXT EMBEDDING,",
219
+ "max_tokens": 512,
220
+ "model_type": "embedding"
221
+ },
222
+ {
223
+ "llm_name": "nomic-ai/nomic-embed-text-v1.5",
224
+ "tags": "TEXT EMBEDDING,",
225
+ "max_tokens": 8192,
226
+ "model_type": "embedding"
227
+ },
228
+ {
229
+ "llm_name": "jinaai/jina-embeddings-v2-small-en",
230
+ "tags": "TEXT EMBEDDING,",
231
+ "max_tokens": 2147483648,
232
+ "model_type": "embedding"
233
+ },
234
+ {
235
+ "llm_name": "jinaai/jina-embeddings-v2-base-en",
236
+ "tags": "TEXT EMBEDDING,",
237
+ "max_tokens": 2147483648,
238
+ "model_type": "embedding"
239
+ }
240
+ ]
241
+ },
242
+ {
243
+ "name": "Xinference",
244
+ "logo": "",
245
+ "tags": "LLM,TEXT EMBEDDING,SPEECH2TEXT,MODERATION,TEXT RE-RANK",
246
+ "status": "1",
247
+ "llm": []
248
+ },
249
+ {
250
+ "name": "Youdao",
251
+ "logo": "",
252
+ "tags": "LLM,TEXT EMBEDDING,SPEECH2TEXT,MODERATION",
253
+ "status": "1",
254
+ "llm": [
255
+ {
256
+ "llm_name": "maidalun1020/bce-embedding-base_v1",
257
+ "tags": "TEXT EMBEDDING,",
258
+ "max_tokens": 512,
259
+ "model_type": "embedding"
260
+ },
261
+ {
262
+ "llm_name": "maidalun1020/bce-reranker-base_v1",
263
+ "tags": "RE-RANK, 512",
264
+ "max_tokens": 512,
265
+ "model_type": "rerank"
266
+ }
267
+ ]
268
+ },
269
+ {
270
+ "name": "DeepSeek",
271
+ "logo": "",
272
+ "tags": "LLM",
273
+ "status": "1",
274
+ "llm": [
275
+ {
276
+ "llm_name": "deepseek-chat",
277
+ "tags": "LLM,CHAT,",
278
+ "max_tokens": 32768,
279
+ "model_type": "chat"
280
+ },
281
+ {
282
+ "llm_name": "deepseek-coder",
283
+ "tags": "LLM,CHAT,",
284
+ "max_tokens": 16385,
285
+ "model_type": "chat"
286
+ }
287
+ ]
288
+ },
289
+ {
290
+ "name": "VolcEngine",
291
+ "logo": "",
292
+ "tags": "LLM, TEXT EMBEDDING",
293
+ "status": "1",
294
+ "llm": [
295
+ {
296
+ "llm_name": "Skylark2-pro-32k",
297
+ "tags": "LLM,CHAT,32k",
298
+ "max_tokens": 32768,
299
+ "model_type": "chat"
300
+ },
301
+ {
302
+ "llm_name": "Skylark2-pro-4k",
303
+ "tags": "LLM,CHAT,4k",
304
+ "max_tokens": 4096,
305
+ "model_type": "chat"
306
+ }
307
+ ]
308
+ },
309
+ {
310
+ "name": "BaiChuan",
311
+ "logo": "",
312
+ "tags": "LLM,TEXT EMBEDDING",
313
+ "status": "1",
314
+ "llm": [
315
+ {
316
+ "llm_name": "Baichuan2-Turbo",
317
+ "tags": "LLM,CHAT,32K",
318
+ "max_tokens": 32768,
319
+ "model_type": "chat"
320
+ },
321
+ {
322
+ "llm_name": "Baichuan2-Turbo-192k",
323
+ "tags": "LLM,CHAT,192K",
324
+ "max_tokens": 196608,
325
+ "model_type": "chat"
326
+ },
327
+ {
328
+ "llm_name": "Baichuan3-Turbo",
329
+ "tags": "LLM,CHAT,32K",
330
+ "max_tokens": 32768,
331
+ "model_type": "chat"
332
+ },
333
+ {
334
+ "llm_name": "Baichuan3-Turbo-128k",
335
+ "tags": "LLM,CHAT,128K",
336
+ "max_tokens": 131072,
337
+ "model_type": "chat"
338
+ },
339
+ {
340
+ "llm_name": "Baichuan4",
341
+ "tags": "LLM,CHAT,128K",
342
+ "max_tokens": 131072,
343
+ "model_type": "chat"
344
+ },
345
+ {
346
+ "llm_name": "Baichuan-Text-Embedding",
347
+ "tags": "TEXT EMBEDDING",
348
+ "max_tokens": 512,
349
+ "model_type": "embedding"
350
+ }
351
+ ]
352
+ },
353
+ {
354
+ "name": "Jina",
355
+ "logo": "",
356
+ "tags": "TEXT EMBEDDING, TEXT RE-RANK",
357
+ "status": "1",
358
+ "llm": [
359
+ {
360
+ "llm_name": "jina-reranker-v1-base-en",
361
+ "tags": "RE-RANK,8k",
362
+ "max_tokens": 8196,
363
+ "model_type": "rerank"
364
+ },
365
+ {
366
+ "llm_name": "jina-reranker-v1-turbo-en",
367
+ "tags": "RE-RANK,8k",
368
+ "max_tokens": 8196,
369
+ "model_type": "rerank"
370
+ },
371
+ {
372
+ "llm_name": "jina-reranker-v1-tiny-en",
373
+ "tags": "RE-RANK,8k",
374
+ "max_tokens": 8196,
375
+ "model_type": "rerank"
376
+ },
377
+ {
378
+ "llm_name": "jina-colbert-v1-en",
379
+ "tags": "RE-RANK,8k",
380
+ "max_tokens": 8196,
381
+ "model_type": "rerank"
382
+ },
383
+ {
384
+ "llm_name": "jina-embeddings-v2-base-en",
385
+ "tags": "TEXT EMBEDDING",
386
+ "max_tokens": 8196,
387
+ "model_type": "embedding"
388
+ },
389
+ {
390
+ "llm_name": "jina-embeddings-v2-base-de",
391
+ "tags": "TEXT EMBEDDING",
392
+ "max_tokens": 8196,
393
+ "model_type": "embedding"
394
+ },
395
+ {
396
+ "llm_name": "jina-embeddings-v2-base-es",
397
+ "tags": "TEXT EMBEDDING",
398
+ "max_tokens": 8196,
399
+ "model_type": "embedding"
400
+ },
401
+ {
402
+ "llm_name": "jina-embeddings-v2-base-code",
403
+ "tags": "TEXT EMBEDDING",
404
+ "max_tokens": 8196,
405
+ "model_type": "embedding"
406
+ },
407
+ {
408
+ "llm_name": "jina-embeddings-v2-base-zh",
409
+ "tags": "TEXT EMBEDDING",
410
+ "max_tokens": 8196,
411
+ "model_type": "embedding"
412
+ }
413
+ ]
414
+ },
415
+ {
416
+ "name": "BAAI",
417
+ "logo": "",
418
+ "tags": "TEXT EMBEDDING, TEXT RE-RANK",
419
+ "status": "1",
420
+ "llm": [
421
+ {
422
+ "llm_name": "BAAI/bge-large-zh-v1.5",
423
+ "tags": "TEXT EMBEDDING,",
424
+ "max_tokens": 1024,
425
+ "model_type": "embedding"
426
+ },
427
+ {
428
+ "llm_name": "BAAI/bge-reranker-v2-m3",
429
+ "tags": "RE-RANK,2k",
430
+ "max_tokens": 2048,
431
+ "model_type": "rerank"
432
+ }
433
+ ]
434
+ },
435
+ {
436
+ "name": "Minimax",
437
+ "logo": "",
438
+ "tags": "LLM,TEXT EMBEDDING",
439
+ "status": "1",
440
+ "llm": [
441
+ {
442
+ "llm_name": "abab6.5",
443
+ "tags": "LLM,CHAT,8k",
444
+ "max_tokens": 8192,
445
+ "model_type": "chat"
446
+ },
447
+ {
448
+ "llm_name": "abab6.5s",
449
+ "tags": "LLM,CHAT,245k",
450
+ "max_tokens": 245760,
451
+ "model_type": "chat"
452
+ },
453
+ {
454
+ "llm_name": "abab6.5t",
455
+ "tags": "LLM,CHAT,8k",
456
+ "max_tokens": 8192,
457
+ "model_type": "chat"
458
+ },
459
+ {
460
+ "llm_name": "abab6.5g",
461
+ "tags": "LLM,CHAT,8k",
462
+ "max_tokens": 8192,
463
+ "model_type": "chat"
464
+ },
465
+ {
466
+ "llm_name": "abab5.5s",
467
+ "tags": "LLM,CHAT,8k",
468
+ "max_tokens": 8192,
469
+ "model_type": "chat"
470
+ }
471
+ ]
472
+ },
473
+ {
474
+ "name": "Mistral",
475
+ "logo": "",
476
+ "tags": "LLM,TEXT EMBEDDING",
477
+ "status": "1",
478
+ "llm": [
479
+ {
480
+ "llm_name": "open-mixtral-8x22b",
481
+ "tags": "LLM,CHAT,64k",
482
+ "max_tokens": 64000,
483
+ "model_type": "chat"
484
+ },
485
+ {
486
+ "llm_name": "open-mixtral-8x7b",
487
+ "tags": "LLM,CHAT,32k",
488
+ "max_tokens": 32000,
489
+ "model_type": "chat"
490
+ },
491
+ {
492
+ "llm_name": "open-mistral-7b",
493
+ "tags": "LLM,CHAT,32k",
494
+ "max_tokens": 32000,
495
+ "model_type": "chat"
496
+ },
497
+ {
498
+ "llm_name": "mistral-large-latest",
499
+ "tags": "LLM,CHAT,32k",
500
+ "max_tokens": 32000,
501
+ "model_type": "chat"
502
+ },
503
+ {
504
+ "llm_name": "mistral-small-latest",
505
+ "tags": "LLM,CHAT,32k",
506
+ "max_tokens": 32000,
507
+ "model_type": "chat"
508
+ },
509
+ {
510
+ "llm_name": "mistral-medium-latest",
511
+ "tags": "LLM,CHAT,32k",
512
+ "max_tokens": 32000,
513
+ "model_type": "chat"
514
+ },
515
+ {
516
+ "llm_name": "codestral-latest",
517
+ "tags": "LLM,CHAT,32k",
518
+ "max_tokens": 32000,
519
+ "model_type": "chat"
520
+ },
521
+ {
522
+ "llm_name": "mistral-embed",
523
+ "tags": "LLM,CHAT,8k",
524
+ "max_tokens": 8192,
525
+ "model_type": "embedding"
526
+ }
527
+ ]
528
+ },
529
+ {
530
+ "name": "Azure-OpenAI",
531
+ "logo": "",
532
+ "tags": "LLM,TEXT EMBEDDING,SPEECH2TEXT,MODERATION",
533
+ "status": "1",
534
+ "llm": [
535
+ {
536
+ "llm_name": "azure-gpt-4o",
537
+ "tags": "LLM,CHAT,128K",
538
+ "max_tokens": 128000,
539
+ "model_type": "chat,image2text"
540
+ },
541
+ {
542
+ "llm_name": "azure-gpt-35-turbo",
543
+ "tags": "LLM,CHAT,4K",
544
+ "max_tokens": 4096,
545
+ "model_type": "chat"
546
+ },
547
+ {
548
+ "llm_name": "azure-gpt-35-turbo-16k",
549
+ "tags": "LLM,CHAT,16k",
550
+ "max_tokens": 16385,
551
+ "model_type": "chat"
552
+ },
553
+ {
554
+ "llm_name": "azure-text-embedding-ada-002",
555
+ "tags": "TEXT EMBEDDING,8K",
556
+ "max_tokens": 8191,
557
+ "model_type": "embedding"
558
+ },
559
+ {
560
+ "llm_name": "azure-text-embedding-3-small",
561
+ "tags": "TEXT EMBEDDING,8K",
562
+ "max_tokens": 8191,
563
+ "model_type": "embedding"
564
+ },
565
+ {
566
+ "llm_name": "azure-text-embedding-3-large",
567
+ "tags": "TEXT EMBEDDING,8K",
568
+ "max_tokens": 8191,
569
+ "model_type": "embedding"
570
+ },
571
+ {
572
+ "llm_name": "azure-whisper-1",
573
+ "tags": "SPEECH2TEXT",
574
+ "max_tokens": 26214400,
575
+ "model_type": "speech2text"
576
+ },
577
+ {
578
+ "llm_name": "azure-gpt-4",
579
+ "tags": "LLM,CHAT,8K",
580
+ "max_tokens": 8191,
581
+ "model_type": "chat"
582
+ },
583
+ {
584
+ "llm_name": "azure-gpt-4-turbo",
585
+ "tags": "LLM,CHAT,8K",
586
+ "max_tokens": 8191,
587
+ "model_type": "chat"
588
+ },
589
+ {
590
+ "llm_name": "azure-gpt-4-32k",
591
+ "tags": "LLM,CHAT,32K",
592
+ "max_tokens": 32768,
593
+ "model_type": "chat"
594
+ },
595
+ {
596
+ "llm_name": "azure-gpt-4-vision-preview",
597
+ "tags": "LLM,CHAT,IMAGE2TEXT",
598
+ "max_tokens": 765,
599
+ "model_type": "image2text"
600
+ }
601
+ ]
602
+ },
603
+ {
604
+ "name": "Bedrock",
605
+ "logo": "",
606
+ "tags": "LLM,TEXT EMBEDDING",
607
+ "status": "1",
608
+ "llm": [
609
+ {
610
+ "llm_name": "ai21.j2-ultra-v1",
611
+ "tags": "LLM,CHAT,8k",
612
+ "max_tokens": 8191,
613
+ "model_type": "chat"
614
+ },
615
+ {
616
+ "llm_name": "ai21.j2-mid-v1",
617
+ "tags": "LLM,CHAT,8k",
618
+ "max_tokens": 8191,
619
+ "model_type": "chat"
620
+ },
621
+ {
622
+ "llm_name": "cohere.command-text-v14",
623
+ "tags": "LLM,CHAT,4k",
624
+ "max_tokens": 4096,
625
+ "model_type": "chat"
626
+ },
627
+ {
628
+ "llm_name": "cohere.command-light-text-v14",
629
+ "tags": "LLM,CHAT,4k",
630
+ "max_tokens": 4096,
631
+ "model_type": "chat"
632
+ },
633
+ {
634
+ "llm_name": "cohere.command-r-v1:0",
635
+ "tags": "LLM,CHAT,128k",
636
+ "max_tokens": 131072,
637
+ "model_type": "chat"
638
+ },
639
+ {
640
+ "llm_name": "cohere.command-r-plus-v1:0",
641
+ "tags": "LLM,CHAT,128k",
642
+ "max_tokens": 128000,
643
+ "model_type": "chat"
644
+ },
645
+ {
646
+ "llm_name": "anthropic.claude-v2",
647
+ "tags": "LLM,CHAT,100k",
648
+ "max_tokens": 102400,
649
+ "model_type": "chat"
650
+ },
651
+ {
652
+ "llm_name": "anthropic.claude-v2:1",
653
+ "tags": "LLM,CHAT,200k",
654
+ "max_tokens": 204800,
655
+ "model_type": "chat"
656
+ },
657
+ {
658
+ "llm_name": "anthropic.claude-3-sonnet-20240229-v1:0",
659
+ "tags": "LLM,CHAT,200k",
660
+ "max_tokens": 204800,
661
+ "model_type": "chat"
662
+ },
663
+ {
664
+ "llm_name": "anthropic.claude-3-5-sonnet-20240620-v1:0",
665
+ "tags": "LLM,CHAT,200k",
666
+ "max_tokens": 204800,
667
+ "model_type": "chat"
668
+ },
669
+ {
670
+ "llm_name": "anthropic.claude-3-haiku-20240307-v1:0",
671
+ "tags": "LLM,CHAT,200k",
672
+ "max_tokens": 204800,
673
+ "model_type": "chat"
674
+ },
675
+ {
676
+ "llm_name": "anthropic.claude-3-opus-20240229-v1:0",
677
+ "tags": "LLM,CHAT,200k",
678
+ "max_tokens": 204800,
679
+ "model_type": "chat"
680
+ },
681
+ {
682
+ "llm_name": "anthropic.claude-instant-v1",
683
+ "tags": "LLM,CHAT,100k",
684
+ "max_tokens": 102400,
685
+ "model_type": "chat"
686
+ },
687
+ {
688
+ "llm_name": "amazon.titan-text-express-v1",
689
+ "tags": "LLM,CHAT,8k",
690
+ "max_tokens": 8192,
691
+ "model_type": "chat"
692
+ },
693
+ {
694
+ "llm_name": "amazon.titan-text-premier-v1:0",
695
+ "tags": "LLM,CHAT,32k",
696
+ "max_tokens": 32768,
697
+ "model_type": "chat"
698
+ },
699
+ {
700
+ "llm_name": "amazon.titan-text-lite-v1",
701
+ "tags": "LLM,CHAT,4k",
702
+ "max_tokens": 4096,
703
+ "model_type": "chat"
704
+ },
705
+ {
706
+ "llm_name": "meta.llama2-13b-chat-v1",
707
+ "tags": "LLM,CHAT,4k",
708
+ "max_tokens": 4096,
709
+ "model_type": "chat"
710
+ },
711
+ {
712
+ "llm_name": "meta.llama2-70b-chat-v1",
713
+ "tags": "LLM,CHAT,4k",
714
+ "max_tokens": 4096,
715
+ "model_type": "chat"
716
+ },
717
+ {
718
+ "llm_name": "meta.llama3-8b-instruct-v1:0",
719
+ "tags": "LLM,CHAT,8k",
720
+ "max_tokens": 8192,
721
+ "model_type": "chat"
722
+ },
723
+ {
724
+ "llm_name": "meta.llama3-70b-instruct-v1:0",
725
+ "tags": "LLM,CHAT,8k",
726
+ "max_tokens": 8192,
727
+ "model_type": "chat"
728
+ },
729
+ {
730
+ "llm_name": "mistral.mistral-7b-instruct-v0:2",
731
+ "tags": "LLM,CHAT,8k",
732
+ "max_tokens": 8192,
733
+ "model_type": "chat"
734
+ },
735
+ {
736
+ "llm_name": "mistral.mixtral-8x7b-instruct-v0:1",
737
+ "tags": "LLM,CHAT,4k",
738
+ "max_tokens": 4096,
739
+ "model_type": "chat"
740
+ },
741
+ {
742
+ "llm_name": "mistral.mistral-large-2402-v1:0",
743
+ "tags": "LLM,CHAT,8k",
744
+ "max_tokens": 8192,
745
+ "model_type": "chat"
746
+ },
747
+ {
748
+ "llm_name": "mistral.mistral-small-2402-v1:0",
749
+ "tags": "LLM,CHAT,8k",
750
+ "max_tokens": 8192,
751
+ "model_type": "chat"
752
+ },
753
+ {
754
+ "llm_name": "amazon.titan-embed-text-v2:0",
755
+ "tags": "TEXT EMBEDDING",
756
+ "max_tokens": 8192,
757
+ "model_type": "embedding"
758
+ },
759
+ {
760
+ "llm_name": "cohere.embed-english-v3",
761
+ "tags": "TEXT EMBEDDING",
762
+ "max_tokens": 2048,
763
+ "model_type": "embedding"
764
+ },
765
+ {
766
+ "llm_name": "cohere.embed-multilingual-v3",
767
+ "tags": "TEXT EMBEDDING",
768
+ "max_tokens": 2048,
769
+ "model_type": "embedding"
770
+ }
771
+ ]
772
+ },
773
+ {
774
+ "name": "Gemini",
775
+ "logo": "",
776
+ "tags": "LLM,TEXT EMBEDDING,IMAGE2TEXT",
777
+ "status": "1",
778
+ "llm": [
779
+ {
780
+ "llm_name": "gemini-1.5-pro-latest",
781
+ "tags": "LLM,CHAT,1024K",
782
+ "max_tokens": 1048576,
783
+ "model_type": "chat"
784
+ },
785
+ {
786
+ "llm_name": "gemini-1.5-flash-latest",
787
+ "tags": "LLM,CHAT,1024K",
788
+ "max_tokens": 1048576,
789
+ "model_type": "chat"
790
+ },
791
+ {
792
+ "llm_name": "gemini-1.0-pro",
793
+ "tags": "LLM,CHAT,30K",
794
+ "max_tokens": 30720,
795
+ "model_type": "chat"
796
+ },
797
+ {
798
+ "llm_name": "gemini-1.0-pro-vision-latest",
799
+ "tags": "LLM,IMAGE2TEXT,12K",
800
+ "max_tokens": 12288,
801
+ "model_type": "image2text"
802
+ },
803
+ {
804
+ "llm_name": "text-embedding-004",
805
+ "tags": "TEXT EMBEDDING",
806
+ "max_tokens": 2048,
807
+ "model_type": "embedding"
808
+ }
809
+ ]
810
+ },
811
+ {
812
+ "name": "Groq",
813
+ "logo": "",
814
+ "tags": "LLM",
815
+ "status": "1",
816
+ "llm": [
817
+ {
818
+ "llm_name": "gemma-7b-it",
819
+ "tags": "LLM,CHAT,15k",
820
+ "max_tokens": 8192,
821
+ "model_type": "chat"
822
+ },
823
+ {
824
+ "llm_name": "gemma2-9b-it",
825
+ "tags": "LLM,CHAT,15k",
826
+ "max_tokens": 8192,
827
+ "model_type": "chat"
828
+ },
829
+ {
830
+ "llm_name": "llama3-70b-8192",
831
+ "tags": "LLM,CHAT,6k",
832
+ "max_tokens": 8192,
833
+ "model_type": "chat"
834
+ },
835
+ {
836
+ "llm_name": "llama3-8b-8192",
837
+ "tags": "LLM,CHAT,30k",
838
+ "max_tokens": 8192,
839
+ "model_type": "chat"
840
+ },
841
+ {
842
+ "llm_name": "mixtral-8x7b-32768",
843
+ "tags": "LLM,CHAT,5k",
844
+ "max_tokens": 32768,
845
+ "model_type": "chat"
846
+ }
847
+ ]
848
+ },
849
+ {
850
+ "name": "OpenRouter",
851
+ "logo": "",
852
+ "tags": "LLM,IMAGE2TEXT",
853
+ "status": "1",
854
+ "llm": [
855
+ {
856
+ "llm_name": "nousresearch/hermes-2-theta-llama-3-8b",
857
+ "tags": "LLM CHAT 16K",
858
+ "max_tokens": 16384,
859
+ "model_type": "chat"
860
+ },
861
+ {
862
+ "llm_name": "alpindale/magnum-72b",
863
+ "tags": "LLM CHAT 16K",
864
+ "max_tokens": 16384,
865
+ "model_type": "chat"
866
+ },
867
+ {
868
+ "llm_name": "google/gemma-2-9b-it",
869
+ "tags": "LLM CHAT 8K",
870
+ "max_tokens": 8192,
871
+ "model_type": "chat"
872
+ },
873
+ {
874
+ "llm_name": "google/gemma-2-9b-it:free",
875
+ "tags": "LLM CHAT 8K",
876
+ "max_tokens": 8192,
877
+ "model_type": "chat"
878
+ },
879
+ {
880
+ "llm_name": "sao10k/l3-stheno-8b",
881
+ "tags": "LLM CHAT 32K",
882
+ "max_tokens": 32000,
883
+ "model_type": "chat"
884
+ },
885
+ {
886
+ "llm_name": "openrouter/flavor-of-the-week",
887
+ "tags": "LLM CHAT 32K",
888
+ "max_tokens": 32000,
889
+ "model_type": "chat"
890
+ },
891
+ {
892
+ "llm_name": "ai21/jamba-instruct",
893
+ "tags": "LLM CHAT 250K",
894
+ "max_tokens": 256000,
895
+ "model_type": "chat"
896
+ },
897
+ {
898
+ "llm_name": "nvidia/nemotron-4-340b-instruct",
899
+ "tags": "LLM CHAT 4K",
900
+ "max_tokens": 4096,
901
+ "model_type": "chat"
902
+ },
903
+ {
904
+ "llm_name": "anthropic/claude-3.5-sonnet",
905
+ "tags": "LLM IMAGE2TEXT 195K",
906
+ "max_tokens": 200000,
907
+ "model_type": "image2text"
908
+ },
909
+ {
910
+ "llm_name": "anthropic/claude-3.5-sonnet:beta",
911
+ "tags": "LLM IMAGE2TEXT 195K",
912
+ "max_tokens": 200000,
913
+ "model_type": "image2text"
914
+ },
915
+ {
916
+ "llm_name": "sao10k/l3-euryale-70b",
917
+ "tags": "LLM CHAT 8K",
918
+ "max_tokens": 8192,
919
+ "model_type": "chat"
920
+ },
921
+ {
922
+ "llm_name": "microsoft/phi-3-medium-4k-instruct",
923
+ "tags": "LLM CHAT 4K",
924
+ "max_tokens": 4000,
925
+ "model_type": "chat"
926
+ },
927
+ {
928
+ "llm_name": "cognitivecomputations/dolphin-mixtral-8x22b",
929
+ "tags": "LLM CHAT 64K",
930
+ "max_tokens": 65536,
931
+ "model_type": "chat"
932
+ },
933
+ {
934
+ "llm_name": "qwen/qwen-2-72b-instruct",
935
+ "tags": "LLM CHAT 32K",
936
+ "max_tokens": 32768,
937
+ "model_type": "chat"
938
+ },
939
+ {
940
+ "llm_name": "openchat/openchat-8b",
941
+ "tags": "LLM CHAT 8K",
942
+ "max_tokens": 8192,
943
+ "model_type": "chat"
944
+ },
945
+ {
946
+ "llm_name": "mistralai/mistral-7b-instruct",
947
+ "tags": "LLM CHAT 32K",
948
+ "max_tokens": 32768,
949
+ "model_type": "chat"
950
+ },
951
+ {
952
+ "llm_name": "mistralai/mistral-7b-instruct-v0.3",
953
+ "tags": "LLM CHAT 32K",
954
+ "max_tokens": 32768,
955
+ "model_type": "chat"
956
+ },
957
+ {
958
+ "llm_name": "nousresearch/hermes-2-pro-llama-3-8b",
959
+ "tags": "LLM CHAT 8K",
960
+ "max_tokens": 8192,
961
+ "model_type": "chat"
962
+ },
963
+ {
964
+ "llm_name": "microsoft/phi-3-mini-128k-instruct",
965
+ "tags": "LLM CHAT 125K",
966
+ "max_tokens": 128000,
967
+ "model_type": "chat"
968
+ },
969
+ {
970
+ "llm_name": "microsoft/phi-3-mini-128k-instruct:free",
971
+ "tags": "LLM CHAT 125K",
972
+ "max_tokens": 128000,
973
+ "model_type": "chat"
974
+ },
975
+ {
976
+ "llm_name": "microsoft/phi-3-medium-128k-instruct",
977
+ "tags": "LLM CHAT 125K",
978
+ "max_tokens": 128000,
979
+ "model_type": "chat"
980
+ },
981
+ {
982
+ "llm_name": "microsoft/phi-3-medium-128k-instruct:free",
983
+ "tags": "LLM CHAT 125K",
984
+ "max_tokens": 128000,
985
+ "model_type": "chat"
986
+ },
987
+ {
988
+ "llm_name": "neversleep/llama-3-lumimaid-70b",
989
+ "tags": "LLM CHAT 8K",
990
+ "max_tokens": 8192,
991
+ "model_type": "chat"
992
+ },
993
+ {
994
+ "llm_name": "google/gemini-flash-1.5",
995
+ "tags": "LLM IMAGE2TEXT 2734K",
996
+ "max_tokens": 2800000,
997
+ "model_type": "image2text"
998
+ },
999
+ {
1000
+ "llm_name": "perplexity/llama-3-sonar-small-32k-chat",
1001
+ "tags": "LLM CHAT 32K",
1002
+ "max_tokens": 32768,
1003
+ "model_type": "chat"
1004
+ },
1005
+ {
1006
+ "llm_name": "perplexity/llama-3-sonar-small-32k-online",
1007
+ "tags": "LLM CHAT 28K",
1008
+ "max_tokens": 28000,
1009
+ "model_type": "chat"
1010
+ },
1011
+ {
1012
+ "llm_name": "perplexity/llama-3-sonar-large-32k-chat",
1013
+ "tags": "LLM CHAT 32K",
1014
+ "max_tokens": 32768,
1015
+ "model_type": "chat"
1016
+ },
1017
+ {
1018
+ "llm_name": "perplexity/llama-3-sonar-large-32k-online",
1019
+ "tags": "LLM CHAT 28K",
1020
+ "max_tokens": 28000,
1021
+ "model_type": "chat"
1022
+ },
1023
+ {
1024
+ "llm_name": "deepseek/deepseek-chat",
1025
+ "tags": "LLM CHAT 125K",
1026
+ "max_tokens": 128000,
1027
+ "model_type": "chat"
1028
+ },
1029
+ {
1030
+ "llm_name": "deepseek/deepseek-coder",
1031
+ "tags": "LLM CHAT 125K",
1032
+ "max_tokens": 128000,
1033
+ "model_type": "chat"
1034
+ },
1035
+ {
1036
+ "llm_name": "openai/gpt-4o",
1037
+ "tags": "LLM IMAGE2TEXT 125K",
1038
+ "max_tokens": 128000,
1039
+ "model_type": "image2text"
1040
+ },
1041
+ {
1042
+ "llm_name": "openai/gpt-4o-2024-05-13",
1043
+ "tags": "LLM IMAGE2TEXT 125K",
1044
+ "max_tokens": 128000,
1045
+ "model_type": "image2text"
1046
+ },
1047
+ {
1048
+ "llm_name": "meta-llama/llama-3-8b",
1049
+ "tags": "LLM CHAT 8K",
1050
+ "max_tokens": 8192,
1051
+ "model_type": "chat"
1052
+ },
1053
+ {
1054
+ "llm_name": "meta-llama/llama-3-70b",
1055
+ "tags": "LLM CHAT 8K",
1056
+ "max_tokens": 8192,
1057
+ "model_type": "chat"
1058
+ },
1059
+ {
1060
+ "llm_name": "meta-llama/llama-guard-2-8b",
1061
+ "tags": "LLM CHAT 8K",
1062
+ "max_tokens": 8192,
1063
+ "model_type": "chat"
1064
+ },
1065
+ {
1066
+ "llm_name": "liuhaotian/llava-yi-34b",
1067
+ "tags": "LLM IMAGE2TEXT 4K",
1068
+ "max_tokens": 4096,
1069
+ "model_type": "image2text"
1070
+ },
1071
+ {
1072
+ "llm_name": "allenai/olmo-7b-instruct",
1073
+ "tags": "LLM CHAT 2K",
1074
+ "max_tokens": 2048,
1075
+ "model_type": "chat"
1076
+ },
1077
+ {
1078
+ "llm_name": "qwen/qwen-110b-chat",
1079
+ "tags": "LLM CHAT 32K",
1080
+ "max_tokens": 32768,
1081
+ "model_type": "chat"
1082
+ },
1083
+ {
1084
+ "llm_name": "qwen/qwen-72b-chat",
1085
+ "tags": "LLM CHAT 32K",
1086
+ "max_tokens": 32768,
1087
+ "model_type": "chat"
1088
+ },
1089
+ {
1090
+ "llm_name": "qwen/qwen-32b-chat",
1091
+ "tags": "LLM CHAT 32K",
1092
+ "max_tokens": 32768,
1093
+ "model_type": "chat"
1094
+ },
1095
+ {
1096
+ "llm_name": "qwen/qwen-14b-chat",
1097
+ "tags": "LLM CHAT 32K",
1098
+ "max_tokens": 32768,
1099
+ "model_type": "chat"
1100
+ },
1101
+ {
1102
+ "llm_name": "qwen/qwen-7b-chat",
1103
+ "tags": "LLM CHAT 32K",
1104
+ "max_tokens": 32768,
1105
+ "model_type": "chat"
1106
+ },
1107
+ {
1108
+ "llm_name": "qwen/qwen-4b-chat",
1109
+ "tags": "LLM CHAT 32K",
1110
+ "max_tokens": 32768,
1111
+ "model_type": "chat"
1112
+ },
1113
+ {
1114
+ "llm_name": "meta-llama/llama-3-8b-instruct:free",
1115
+ "tags": "LLM CHAT 8K",
1116
+ "max_tokens": 8192,
1117
+ "model_type": "chat"
1118
+ },
1119
+ {
1120
+ "llm_name": "neversleep/llama-3-lumimaid-8b",
1121
+ "tags": "LLM CHAT 24K",
1122
+ "max_tokens": 24576,
1123
+ "model_type": "chat"
1124
+ },
1125
+ {
1126
+ "llm_name": "neversleep/llama-3-lumimaid-8b:extended",
1127
+ "tags": "LLM CHAT 24K",
1128
+ "max_tokens": 24576,
1129
+ "model_type": "chat"
1130
+ },
1131
+ {
1132
+ "llm_name": "snowflake/snowflake-arctic-instruct",
1133
+ "tags": "LLM CHAT 4K",
1134
+ "max_tokens": 4096,
1135
+ "model_type": "chat"
1136
+ },
1137
+ {
1138
+ "llm_name": "fireworks/firellava-13b",
1139
+ "tags": "LLM IMAGE2TEXT 4K",
1140
+ "max_tokens": 4096,
1141
+ "model_type": "image2text"
1142
+ },
1143
+ {
1144
+ "llm_name": "lynn/soliloquy-l3",
1145
+ "tags": "LLM CHAT 24K",
1146
+ "max_tokens": 24576,
1147
+ "model_type": "chat"
1148
+ },
1149
+ {
1150
+ "llm_name": "sao10k/fimbulvetr-11b-v2",
1151
+ "tags": "LLM CHAT 8K",
1152
+ "max_tokens": 8192,
1153
+ "model_type": "chat"
1154
+ },
1155
+ {
1156
+ "llm_name": "meta-llama/llama-3-8b-instruct:extended",
1157
+ "tags": "LLM CHAT 16K",
1158
+ "max_tokens": 16384,
1159
+ "model_type": "chat"
1160
+ },
1161
+ {
1162
+ "llm_name": "meta-llama/llama-3-8b-instruct:nitro",
1163
+ "tags": "LLM CHAT 8K",
1164
+ "max_tokens": 8192,
1165
+ "model_type": "chat"
1166
+ },
1167
+ {
1168
+ "llm_name": "meta-llama/llama-3-70b-instruct:nitro",
1169
+ "tags": "LLM CHAT 8K",
1170
+ "max_tokens": 8192,
1171
+ "model_type": "chat"
1172
+ },
1173
+ {
1174
+ "llm_name": "meta-llama/llama-3-8b-instruct",
1175
+ "tags": "LLM CHAT 8K",
1176
+ "max_tokens": 8192,
1177
+ "model_type": "chat"
1178
+ },
1179
+ {
1180
+ "llm_name": "meta-llama/llama-3-70b-instruct",
1181
+ "tags": "LLM CHAT 8K",
1182
+ "max_tokens": 8192,
1183
+ "model_type": "chat"
1184
+ },
1185
+ {
1186
+ "llm_name": "mistralai/mixtral-8x22b-instruct",
1187
+ "tags": "LLM CHAT 64K",
1188
+ "max_tokens": 65536,
1189
+ "model_type": "chat"
1190
+ },
1191
+ {
1192
+ "llm_name": "microsoft/wizardlm-2-8x22b",
1193
+ "tags": "LLM CHAT 64K",
1194
+ "max_tokens": 65536,
1195
+ "model_type": "chat"
1196
+ },
1197
+ {
1198
+ "llm_name": "microsoft/wizardlm-2-7b",
1199
+ "tags": "LLM CHAT 32K",
1200
+ "max_tokens": 32000,
1201
+ "model_type": "chat"
1202
+ },
1203
+ {
1204
+ "llm_name": "undi95/toppy-m-7b:nitro",
1205
+ "tags": "LLM CHAT 4K",
1206
+ "max_tokens": 4096,
1207
+ "model_type": "chat"
1208
+ },
1209
+ {
1210
+ "llm_name": "mistralai/mixtral-8x22b",
1211
+ "tags": "LLM CHAT 64K",
1212
+ "max_tokens": 65536,
1213
+ "model_type": "chat"
1214
+ },
1215
+ {
1216
+ "llm_name": "openai/gpt-4-turbo",
1217
+ "tags": "LLM IMAGE2TEXT 125K",
1218
+ "max_tokens": 128000,
1219
+ "model_type": "image2text"
1220
+ },
1221
+ {
1222
+ "llm_name": "google/gemini-pro-1.5",
1223
+ "tags": "LLM IMAGE2TEXT 2734K",
1224
+ "max_tokens": 2800000,
1225
+ "model_type": "image2text"
1226
+ },
1227
+ {
1228
+ "llm_name": "cohere/command-r-plus",
1229
+ "tags": "LLM CHAT 125K",
1230
+ "max_tokens": 128000,
1231
+ "model_type": "chat"
1232
+ },
1233
+ {
1234
+ "llm_name": "databricks/dbrx-instruct",
1235
+ "tags": "LLM CHAT 32K",
1236
+ "max_tokens": 32768,
1237
+ "model_type": "chat"
1238
+ },
1239
+ {
1240
+ "llm_name": "sophosympatheia/midnight-rose-70b",
1241
+ "tags": "LLM CHAT 4K",
1242
+ "max_tokens": 4096,
1243
+ "model_type": "chat"
1244
+ },
1245
+ {
1246
+ "llm_name": "cohere/command",
1247
+ "tags": "LLM CHAT 4K",
1248
+ "max_tokens": 4096,
1249
+ "model_type": "chat"
1250
+ },
1251
+ {
1252
+ "llm_name": "cohere/command-r",
1253
+ "tags": "LLM CHAT 125K",
1254
+ "max_tokens": 128000,
1255
+ "model_type": "chat"
1256
+ },
1257
+ {
1258
+ "llm_name": "anthropic/claude-3-haiku",
1259
+ "tags": "LLM IMAGE2TEXT 195K",
1260
+ "max_tokens": 200000,
1261
+ "model_type": "image2text"
1262
+ },
1263
+ {
1264
+ "llm_name": "anthropic/claude-3-haiku:beta",
1265
+ "tags": "LLM IMAGE2TEXT 195K",
1266
+ "max_tokens": 200000,
1267
+ "model_type": "image2text"
1268
+ },
1269
+ {
1270
+ "llm_name": "google/gemma-7b-it:nitro",
1271
+ "tags": "LLM CHAT 8K",
1272
+ "max_tokens": 8192,
1273
+ "model_type": "chat"
1274
+ },
1275
+ {
1276
+ "llm_name": "mistralai/mixtral-8x7b-instruct:nitro",
1277
+ "tags": "LLM CHAT 32K",
1278
+ "max_tokens": 32768,
1279
+ "model_type": "chat"
1280
+ },
1281
+ {
1282
+ "llm_name": "mistralai/mistral-7b-instruct:nitro",
1283
+ "tags": "LLM CHAT 32K",
1284
+ "max_tokens": 32768,
1285
+ "model_type": "chat"
1286
+ },
1287
+ {
1288
+ "llm_name": "meta-llama/llama-2-70b-chat:nitro",
1289
+ "tags": "LLM CHAT 4K",
1290
+ "max_tokens": 4096,
1291
+ "model_type": "chat"
1292
+ },
1293
+ {
1294
+ "llm_name": "gryphe/mythomax-l2-13b:nitro",
1295
+ "tags": "LLM CHAT 4K",
1296
+ "max_tokens": 4096,
1297
+ "model_type": "chat"
1298
+ },
1299
+ {
1300
+ "llm_name": "anthropic/claude-3-opus",
1301
+ "tags": "LLM IMAGE2TEXT 195K",
1302
+ "max_tokens": 200000,
1303
+ "model_type": "image2text"
1304
+ },
1305
+ {
1306
+ "llm_name": "anthropic/claude-3-sonnet",
1307
+ "tags": "LLM IMAGE2TEXT 195K",
1308
+ "max_tokens": 200000,
1309
+ "model_type": "image2text"
1310
+ },
1311
+ {
1312
+ "llm_name": "anthropic/claude-3-opus:beta",
1313
+ "tags": "LLM IMAGE2TEXT 195K",
1314
+ "max_tokens": 200000,
1315
+ "model_type": "image2text"
1316
+ },
1317
+ {
1318
+ "llm_name": "anthropic/claude-3-sonnet:beta",
1319
+ "tags": "LLM IMAGE2TEXT 195K",
1320
+ "max_tokens": 200000,
1321
+ "model_type": "image2text"
1322
+ },
1323
+ {
1324
+ "llm_name": "mistralai/mistral-large",
1325
+ "tags": "LLM CHAT 32K",
1326
+ "max_tokens": 32000,
1327
+ "model_type": "chat"
1328
+ },
1329
+ {
1330
+ "llm_name": "google/gemma-7b-it",
1331
+ "tags": "LLM CHAT 8K",
1332
+ "max_tokens": 8192,
1333
+ "model_type": "chat"
1334
+ },
1335
+ {
1336
+ "llm_name": "google/gemma-7b-it:free",
1337
+ "tags": "LLM CHAT 8K",
1338
+ "max_tokens": 8192,
1339
+ "model_type": "chat"
1340
+ },
1341
+ {
1342
+ "llm_name": "nousresearch/nous-hermes-2-mistral-7b-dpo",
1343
+ "tags": "LLM CHAT 8K",
1344
+ "max_tokens": 8192,
1345
+ "model_type": "chat"
1346
+ },
1347
+ {
1348
+ "llm_name": "meta-llama/codellama-70b-instruct",
1349
+ "tags": "LLM CHAT 2K",
1350
+ "max_tokens": 2048,
1351
+ "model_type": "chat"
1352
+ },
1353
+ {
1354
+ "llm_name": "recursal/eagle-7b",
1355
+ "tags": "LLM CHAT 9K",
1356
+ "max_tokens": 10000,
1357
+ "model_type": "chat"
1358
+ },
1359
+ {
1360
+ "llm_name": "openai/gpt-3.5-turbo-0613",
1361
+ "tags": "LLM CHAT 4K",
1362
+ "max_tokens": 4095,
1363
+ "model_type": "chat"
1364
+ },
1365
+ {
1366
+ "llm_name": "openai/gpt-4-turbo-preview",
1367
+ "tags": "LLM CHAT 125K",
1368
+ "max_tokens": 128000,
1369
+ "model_type": "chat"
1370
+ },
1371
+ {
1372
+ "llm_name": "undi95/remm-slerp-l2-13b:extended",
1373
+ "tags": "LLM CHAT 6K",
1374
+ "max_tokens": 6144,
1375
+ "model_type": "chat"
1376
+ },
1377
+ {
1378
+ "llm_name": "nousresearch/nous-hermes-2-mixtral-8x7b-dpo",
1379
+ "tags": "LLM CHAT 32K",
1380
+ "max_tokens": 32768,
1381
+ "model_type": "chat"
1382
+ },
1383
+ {
1384
+ "llm_name": "nousresearch/nous-hermes-2-mixtral-8x7b-sft",
1385
+ "tags": "LLM CHAT 32K",
1386
+ "max_tokens": 32768,
1387
+ "model_type": "chat"
1388
+ },
1389
+ {
1390
+ "llm_name": "mistralai/mistral-tiny",
1391
+ "tags": "LLM CHAT 32K",
1392
+ "max_tokens": 32000,
1393
+ "model_type": "chat"
1394
+ },
1395
+ {
1396
+ "llm_name": "mistralai/mistral-small",
1397
+ "tags": "LLM CHAT 32K",
1398
+ "max_tokens": 32000,
1399
+ "model_type": "chat"
1400
+ },
1401
+ {
1402
+ "llm_name": "mistralai/mistral-medium",
1403
+ "tags": "LLM CHAT 32K",
1404
+ "max_tokens": 32000,
1405
+ "model_type": "chat"
1406
+ },
1407
+ {
1408
+ "llm_name": "austism/chronos-hermes-13b",
1409
+ "tags": "LLM CHAT 4K",
1410
+ "max_tokens": 4096,
1411
+ "model_type": "chat"
1412
+ },
1413
+ {
1414
+ "llm_name": "neversleep/noromaid-mixtral-8x7b-instruct",
1415
+ "tags": "LLM CHAT 8K",
1416
+ "max_tokens": 8000,
1417
+ "model_type": "chat"
1418
+ },
1419
+ {
1420
+ "llm_name": "nousresearch/nous-hermes-yi-34b",
1421
+ "tags": "LLM CHAT 4K",
1422
+ "max_tokens": 4096,
1423
+ "model_type": "chat"
1424
+ },
1425
+ {
1426
+ "llm_name": "mistralai/mistral-7b-instruct-v0.2",
1427
+ "tags": "LLM CHAT 32K",
1428
+ "max_tokens": 32768,
1429
+ "model_type": "chat"
1430
+ },
1431
+ {
1432
+ "llm_name": "cognitivecomputations/dolphin-mixtral-8x7b",
1433
+ "tags": "LLM CHAT 32K",
1434
+ "max_tokens": 32768,
1435
+ "model_type": "chat"
1436
+ },
1437
+ {
1438
+ "llm_name": "google/gemini-pro",
1439
+ "tags": "LLM CHAT 89K",
1440
+ "max_tokens": 91728,
1441
+ "model_type": "chat"
1442
+ },
1443
+ {
1444
+ "llm_name": "google/gemini-pro-vision",
1445
+ "tags": "LLM IMAGE2TEXT 44K",
1446
+ "max_tokens": 45875,
1447
+ "model_type": "image2text"
1448
+ },
1449
+ {
1450
+ "llm_name": "mistralai/mixtral-8x7b",
1451
+ "tags": "LLM CHAT 32K",
1452
+ "max_tokens": 32768,
1453
+ "model_type": "chat"
1454
+ },
1455
+ {
1456
+ "llm_name": "mistralai/mixtral-8x7b-instruct",
1457
+ "tags": "LLM CHAT 32K",
1458
+ "max_tokens": 32768,
1459
+ "model_type": "chat"
1460
+ },
1461
+ {
1462
+ "llm_name": "rwkv/rwkv-5-world-3b",
1463
+ "tags": "LLM CHAT 9K",
1464
+ "max_tokens": 10000,
1465
+ "model_type": "chat"
1466
+ },
1467
+ {
1468
+ "llm_name": "recursal/rwkv-5-3b-ai-town",
1469
+ "tags": "LLM CHAT 9K",
1470
+ "max_tokens": 10000,
1471
+ "model_type": "chat"
1472
+ },
1473
+ {
1474
+ "llm_name": "togethercomputer/stripedhyena-nous-7b",
1475
+ "tags": "LLM CHAT 32K",
1476
+ "max_tokens": 32768,
1477
+ "model_type": "chat"
1478
+ },
1479
+ {
1480
+ "llm_name": "togethercomputer/stripedhyena-hessian-7b",
1481
+ "tags": "LLM CHAT 32K",
1482
+ "max_tokens": 32768,
1483
+ "model_type": "chat"
1484
+ },
1485
+ {
1486
+ "llm_name": "koboldai/psyfighter-13b-2",
1487
+ "tags": "LLM CHAT 4K",
1488
+ "max_tokens": 4096,
1489
+ "model_type": "chat"
1490
+ },
1491
+ {
1492
+ "llm_name": "gryphe/mythomist-7b",
1493
+ "tags": "LLM CHAT 32K",
1494
+ "max_tokens": 32768,
1495
+ "model_type": "chat"
1496
+ },
1497
+ {
1498
+ "llm_name": "openrouter/cinematika-7b",
1499
+ "tags": "LLM CHAT 8K",
1500
+ "max_tokens": 8000,
1501
+ "model_type": "chat"
1502
+ },
1503
+ {
1504
+ "llm_name": "nousresearch/nous-capybara-7b",
1505
+ "tags": "LLM CHAT 8K",
1506
+ "max_tokens": 8192,
1507
+ "model_type": "chat"
1508
+ },
1509
+ {
1510
+ "llm_name": "nousresearch/nous-capybara-7b:free",
1511
+ "tags": "LLM CHAT 8K",
1512
+ "max_tokens": 8192,
1513
+ "model_type": "chat"
1514
+ },
1515
+ {
1516
+ "llm_name": "openchat/openchat-7b",
1517
+ "tags": "LLM CHAT 8K",
1518
+ "max_tokens": 8192,
1519
+ "model_type": "chat"
1520
+ },
1521
+ {
1522
+ "llm_name": "openchat/openchat-7b:free",
1523
+ "tags": "LLM CHAT 8K",
1524
+ "max_tokens": 8192,
1525
+ "model_type": "chat"
1526
+ },
1527
+ {
1528
+ "llm_name": "neversleep/noromaid-20b",
1529
+ "tags": "LLM CHAT 8K",
1530
+ "max_tokens": 8192,
1531
+ "model_type": "chat"
1532
+ },
1533
+ {
1534
+ "llm_name": "gryphe/mythomist-7b:free",
1535
+ "tags": "LLM CHAT 32K",
1536
+ "max_tokens": 32768,
1537
+ "model_type": "chat"
1538
+ },
1539
+ {
1540
+ "llm_name": "intel/neural-chat-7b",
1541
+ "tags": "LLM CHAT 4K",
1542
+ "max_tokens": 4096,
1543
+ "model_type": "chat"
1544
+ },
1545
+ {
1546
+ "llm_name": "anthropic/claude-2",
1547
+ "tags": "LLM CHAT 195K",
1548
+ "max_tokens": 200000,
1549
+ "model_type": "chat"
1550
+ },
1551
+ {
1552
+ "llm_name": "anthropic/claude-2.1",
1553
+ "tags": "LLM CHAT 195K",
1554
+ "max_tokens": 200000,
1555
+ "model_type": "chat"
1556
+ },
1557
+ {
1558
+ "llm_name": "anthropic/claude-instant-1.1",
1559
+ "tags": "LLM CHAT 98K",
1560
+ "max_tokens": 100000,
1561
+ "model_type": "chat"
1562
+ },
1563
+ {
1564
+ "llm_name": "anthropic/claude-2:beta",
1565
+ "tags": "LLM CHAT 195K",
1566
+ "max_tokens": 200000,
1567
+ "model_type": "chat"
1568
+ },
1569
+ {
1570
+ "llm_name": "anthropic/claude-2.1:beta",
1571
+ "tags": "LLM CHAT 195K",
1572
+ "max_tokens": 200000,
1573
+ "model_type": "chat"
1574
+ },
1575
+ {
1576
+ "llm_name": "teknium/openhermes-2.5-mistral-7b",
1577
+ "tags": "LLM CHAT 4K",
1578
+ "max_tokens": 4096,
1579
+ "model_type": "chat"
1580
+ },
1581
+ {
1582
+ "llm_name": "nousresearch/nous-capybara-34b",
1583
+ "tags": "LLM CHAT 195K",
1584
+ "max_tokens": 200000,
1585
+ "model_type": "chat"
1586
+ },
1587
+ {
1588
+ "llm_name": "openai/gpt-4-vision-preview",
1589
+ "tags": "LLM IMAGE2TEXT 125K",
1590
+ "max_tokens": 128000,
1591
+ "model_type": "image2text"
1592
+ },
1593
+ {
1594
+ "llm_name": "lizpreciatior/lzlv-70b-fp16-hf",
1595
+ "tags": "LLM CHAT 4K",
1596
+ "max_tokens": 4096,
1597
+ "model_type": "chat"
1598
+ },
1599
+ {
1600
+ "llm_name": "undi95/toppy-m-7b",
1601
+ "tags": "LLM CHAT 4K",
1602
+ "max_tokens": 4096,
1603
+ "model_type": "chat"
1604
+ },
1605
+ {
1606
+ "llm_name": "alpindale/goliath-120b",
1607
+ "tags": "LLM CHAT 6K",
1608
+ "max_tokens": 6144,
1609
+ "model_type": "chat"
1610
+ },
1611
+ {
1612
+ "llm_name": "undi95/toppy-m-7b:free",
1613
+ "tags": "LLM CHAT 4K",
1614
+ "max_tokens": 4096,
1615
+ "model_type": "chat"
1616
+ },
1617
+ {
1618
+ "llm_name": "openrouter/auto",
1619
+ "tags": "LLM CHAT 195K",
1620
+ "max_tokens": 200000,
1621
+ "model_type": "chat"
1622
+ },
1623
+ {
1624
+ "llm_name": "openai/gpt-3.5-turbo-1106",
1625
+ "tags": "LLM CHAT 16K",
1626
+ "max_tokens": 16385,
1627
+ "model_type": "chat"
1628
+ },
1629
+ {
1630
+ "llm_name": "openai/gpt-4-1106-preview",
1631
+ "tags": "LLM CHAT 125K",
1632
+ "max_tokens": 128000,
1633
+ "model_type": "chat"
1634
+ },
1635
+ {
1636
+ "llm_name": "huggingfaceh4/zephyr-7b-beta:free",
1637
+ "tags": "LLM CHAT 4K",
1638
+ "max_tokens": 4096,
1639
+ "model_type": "chat"
1640
+ },
1641
+ {
1642
+ "llm_name": "google/palm-2-chat-bison-32k",
1643
+ "tags": "LLM CHAT 89K",
1644
+ "max_tokens": 91750,
1645
+ "model_type": "chat"
1646
+ },
1647
+ {
1648
+ "llm_name": "google/palm-2-codechat-bison-32k",
1649
+ "tags": "LLM CHAT 89K",
1650
+ "max_tokens": 91750,
1651
+ "model_type": "chat"
1652
+ },
1653
+ {
1654
+ "llm_name": "teknium/openhermes-2-mistral-7b",
1655
+ "tags": "LLM CHAT 8K",
1656
+ "max_tokens": 8192,
1657
+ "model_type": "chat"
1658
+ },
1659
+ {
1660
+ "llm_name": "open-orca/mistral-7b-openorca",
1661
+ "tags": "LLM CHAT 8K",
1662
+ "max_tokens": 8192,
1663
+ "model_type": "chat"
1664
+ },
1665
+ {
1666
+ "llm_name": "gryphe/mythomax-l2-13b:extended",
1667
+ "tags": "LLM CHAT 8K",
1668
+ "max_tokens": 8192,
1669
+ "model_type": "chat"
1670
+ },
1671
+ {
1672
+ "llm_name": "xwin-lm/xwin-lm-70b",
1673
+ "tags": "LLM CHAT 8K",
1674
+ "max_tokens": 8192,
1675
+ "model_type": "chat"
1676
+ },
1677
+ {
1678
+ "llm_name": "openai/gpt-3.5-turbo-instruct",
1679
+ "tags": "LLM CHAT 4K",
1680
+ "max_tokens": 4095,
1681
+ "model_type": "chat"
1682
+ },
1683
+ {
1684
+ "llm_name": "mistralai/mistral-7b-instruct-v0.1",
1685
+ "tags": "LLM CHAT 4K",
1686
+ "max_tokens": 4096,
1687
+ "model_type": "chat"
1688
+ },
1689
+ {
1690
+ "llm_name": "mistralai/mistral-7b-instruct:free",
1691
+ "tags": "LLM CHAT 32K",
1692
+ "max_tokens": 32768,
1693
+ "model_type": "chat"
1694
+ },
1695
+ {
1696
+ "llm_name": "pygmalionai/mythalion-13b",
1697
+ "tags": "LLM CHAT 8K",
1698
+ "max_tokens": 8192,
1699
+ "model_type": "chat"
1700
+ },
1701
+ {
1702
+ "llm_name": "openai/gpt-3.5-turbo-16k",
1703
+ "tags": "LLM CHAT 16K",
1704
+ "max_tokens": 16385,
1705
+ "model_type": "chat"
1706
+ },
1707
+ {
1708
+ "llm_name": "openai/gpt-4-32k",
1709
+ "tags": "LLM CHAT 32K",
1710
+ "max_tokens": 32767,
1711
+ "model_type": "chat"
1712
+ },
1713
+ {
1714
+ "llm_name": "openai/gpt-4-32k-0314",
1715
+ "tags": "LLM CHAT 32K",
1716
+ "max_tokens": 32767,
1717
+ "model_type": "chat"
1718
+ },
1719
+ {
1720
+ "llm_name": "meta-llama/codellama-34b-instruct",
1721
+ "tags": "LLM CHAT 8K",
1722
+ "max_tokens": 8192,
1723
+ "model_type": "chat"
1724
+ },
1725
+ {
1726
+ "llm_name": "phind/phind-codellama-34b",
1727
+ "tags": "LLM CHAT 4K",
1728
+ "max_tokens": 4096,
1729
+ "model_type": "chat"
1730
+ },
1731
+ {
1732
+ "llm_name": "nousresearch/nous-hermes-llama2-13b",
1733
+ "tags": "LLM CHAT 4K",
1734
+ "max_tokens": 4096,
1735
+ "model_type": "chat"
1736
+ },
1737
+ {
1738
+ "llm_name": "mancer/weaver",
1739
+ "tags": "LLM CHAT 8K",
1740
+ "max_tokens": 8000,
1741
+ "model_type": "chat"
1742
+ },
1743
+ {
1744
+ "llm_name": "anthropic/claude-2.0",
1745
+ "tags": "LLM CHAT 98K",
1746
+ "max_tokens": 100000,
1747
+ "model_type": "chat"
1748
+ },
1749
+ {
1750
+ "llm_name": "anthropic/claude-instant-1",
1751
+ "tags": "LLM CHAT 98K",
1752
+ "max_tokens": 100000,
1753
+ "model_type": "chat"
1754
+ },
1755
+ {
1756
+ "llm_name": "anthropic/claude-1",
1757
+ "tags": "LLM CHAT 98K",
1758
+ "max_tokens": 100000,
1759
+ "model_type": "chat"
1760
+ },
1761
+ {
1762
+ "llm_name": "anthropic/claude-1.2",
1763
+ "tags": "LLM CHAT 98K",
1764
+ "max_tokens": 100000,
1765
+ "model_type": "chat"
1766
+ },
1767
+ {
1768
+ "llm_name": "anthropic/claude-instant-1.0",
1769
+ "tags": "LLM CHAT 98K",
1770
+ "max_tokens": 100000,
1771
+ "model_type": "chat"
1772
+ },
1773
+ {
1774
+ "llm_name": "anthropic/claude-2.0:beta",
1775
+ "tags": "LLM CHAT 98K",
1776
+ "max_tokens": 100000,
1777
+ "model_type": "chat"
1778
+ },
1779
+ {
1780
+ "llm_name": "anthropic/claude-instant-1:beta",
1781
+ "tags": "LLM CHAT 98K",
1782
+ "max_tokens": 100000,
1783
+ "model_type": "chat"
1784
+ },
1785
+ {
1786
+ "llm_name": "undi95/remm-slerp-l2-13b",
1787
+ "tags": "LLM CHAT 4K",
1788
+ "max_tokens": 4096,
1789
+ "model_type": "chat"
1790
+ },
1791
+ {
1792
+ "llm_name": "google/palm-2-chat-bison",
1793
+ "tags": "LLM CHAT 25K",
1794
+ "max_tokens": 25804,
1795
+ "model_type": "chat"
1796
+ },
1797
+ {
1798
+ "llm_name": "google/palm-2-codechat-bison",
1799
+ "tags": "LLM CHAT 19K",
1800
+ "max_tokens": 20070,
1801
+ "model_type": "chat"
1802
+ },
1803
+ {
1804
+ "llm_name": "gryphe/mythomax-l2-13b",
1805
+ "tags": "LLM CHAT 4K",
1806
+ "max_tokens": 4096,
1807
+ "model_type": "chat"
1808
+ },
1809
+ {
1810
+ "llm_name": "meta-llama/llama-2-13b-chat",
1811
+ "tags": "LLM CHAT 4K",
1812
+ "max_tokens": 4096,
1813
+ "model_type": "chat"
1814
+ },
1815
+ {
1816
+ "llm_name": "meta-llama/llama-2-70b-chat",
1817
+ "tags": "LLM CHAT 4K",
1818
+ "max_tokens": 4096,
1819
+ "model_type": "chat"
1820
+ },
1821
+ {
1822
+ "llm_name": "openai/gpt-3.5-turbo",
1823
+ "tags": "LLM CHAT 16K",
1824
+ "max_tokens": 16385,
1825
+ "model_type": "chat"
1826
+ },
1827
+ {
1828
+ "llm_name": "openai/gpt-3.5-turbo-0125",
1829
+ "tags": "LLM CHAT 16K",
1830
+ "max_tokens": 16385,
1831
+ "model_type": "chat"
1832
+ },
1833
+ {
1834
+ "llm_name": "openai/gpt-3.5-turbo-0301",
1835
+ "tags": "LLM CHAT 4K",
1836
+ "max_tokens": 4095,
1837
+ "model_type": "chat"
1838
+ },
1839
+ {
1840
+ "llm_name": "openai/gpt-4",
1841
+ "tags": "LLM CHAT 8K",
1842
+ "max_tokens": 8191,
1843
+ "model_type": "chat"
1844
+ },
1845
+ {
1846
+ "llm_name": "openai/gpt-4-0314",
1847
+ "tags": "LLM CHAT 8K",
1848
+ "max_tokens": 8191,
1849
+ "model_type": "chat"
1850
+ },
1851
+ {
1852
+ "llm_name": "01-ai/yi-large",
1853
+ "tags": "LLM CHAT 32K",
1854
+ "max_tokens": 32768,
1855
+ "model_type": "chat"
1856
+ },
1857
+ {
1858
+ "llm_name": "01-ai/yi-34b-200k",
1859
+ "tags": "LLM CHAT 195K",
1860
+ "max_tokens": 200000,
1861
+ "model_type": "chat"
1862
+ },
1863
+ {
1864
+ "llm_name": "01-ai/yi-34b-chat",
1865
+ "tags": "LLM CHAT 4K",
1866
+ "max_tokens": 4096,
1867
+ "model_type": "chat"
1868
+ },
1869
+ {
1870
+ "llm_name": "01-ai/yi-34b",
1871
+ "tags": "LLM CHAT 4K",
1872
+ "max_tokens": 4096,
1873
+ "model_type": "chat"
1874
+ },
1875
+ {
1876
+ "llm_name": "01-ai/yi-6b",
1877
+ "tags": "LLM CHAT 4K",
1878
+ "max_tokens": 4096,
1879
+ "model_type": "chat"
1880
+ }
1881
+ ]
1882
+ }
1883
+ ]
1884
+ }
rag/llm/__init__.py CHANGED
@@ -45,7 +45,8 @@ CvModel = {
45
  "Tongyi-Qianwen": QWenCV,
46
  "ZHIPU-AI": Zhipu4V,
47
  "Moonshot": LocalCV,
48
- 'Gemini':GeminiCV
 
49
  }
50
 
51
 
@@ -65,7 +66,8 @@ ChatModel = {
65
  "Mistral": MistralChat,
66
  'Gemini' : GeminiChat,
67
  "Bedrock": BedrockChat,
68
- "Groq": GroqChat
 
69
  }
70
 
71
 
 
45
  "Tongyi-Qianwen": QWenCV,
46
  "ZHIPU-AI": Zhipu4V,
47
  "Moonshot": LocalCV,
48
+ 'Gemini':GeminiCV,
49
+ 'OpenRouter':OpenRouterCV
50
  }
51
 
52
 
 
66
  "Mistral": MistralChat,
67
  'Gemini' : GeminiChat,
68
  "Bedrock": BedrockChat,
69
+ "Groq": GroqChat,
70
+ 'OpenRouter':OpenRouterChat
71
  }
72
 
73
 
rag/llm/chat_model.py CHANGED
@@ -685,7 +685,6 @@ class GeminiChat(Base):
685
  yield response._chunks[-1].usage_metadata.total_token_count
686
 
687
 
688
-
689
  class GroqChat:
690
  def __init__(self, key, model_name,base_url=''):
691
  self.client = Groq(api_key=key)
@@ -697,7 +696,6 @@ class GroqChat:
697
  for k in list(gen_conf.keys()):
698
  if k not in ["temperature", "top_p", "max_tokens"]:
699
  del gen_conf[k]
700
-
701
  ans = ""
702
  try:
703
  response = self.client.chat.completions.create(
@@ -707,7 +705,7 @@ class GroqChat:
707
  )
708
  ans = response.choices[0].message.content
709
  if response.choices[0].finish_reason == "length":
710
- ans += "...\nFor the content length reason, it stopped, continue?" if self.is_english(
711
  [ans]) else "······\n由于长度的原因，回答被截断了，要继续吗？"
712
  return ans, response.usage.total_tokens
713
  except Exception as e:
@@ -734,11 +732,20 @@ class GroqChat:
734
  ans += resp.choices[0].delta.content
735
  total_tokens += 1
736
  if resp.choices[0].finish_reason == "length":
737
- ans += "...\nFor the content length reason, it stopped, continue?" if self.is_english(
738
  [ans]) else "······\n由于长度的原因，回答被截断了，要继续吗？"
739
  yield ans
740
 
741
  except Exception as e:
742
  yield ans + "\n**ERROR**: " + str(e)
743
 
744
- yield total_tokens
 
 
 
 
 
 
 
 
 
 
685
  yield response._chunks[-1].usage_metadata.total_token_count
686
 
687
 
 
688
  class GroqChat:
689
  def __init__(self, key, model_name,base_url=''):
690
  self.client = Groq(api_key=key)
 
696
  for k in list(gen_conf.keys()):
697
  if k not in ["temperature", "top_p", "max_tokens"]:
698
  del gen_conf[k]
 
699
  ans = ""
700
  try:
701
  response = self.client.chat.completions.create(
 
705
  )
706
  ans = response.choices[0].message.content
707
  if response.choices[0].finish_reason == "length":
708
+ ans += "...\nFor the content length reason, it stopped, continue?" if is_english(
709
  [ans]) else "······\n由于长度的原因，回答被截断了，要继续吗？"
710
  return ans, response.usage.total_tokens
711
  except Exception as e:
 
732
  ans += resp.choices[0].delta.content
733
  total_tokens += 1
734
  if resp.choices[0].finish_reason == "length":
735
+ ans += "...\nFor the content length reason, it stopped, continue?" if is_english(
736
  [ans]) else "······\n由于长度的原因，回答被截断了，要继续吗？"
737
  yield ans
738
 
739
  except Exception as e:
740
  yield ans + "\n**ERROR**: " + str(e)
741
 
742
+ yield total_tokens
743
+
744
+
745
+ ## openrouter
746
+ class OpenRouterChat(Base):
747
+ def __init__(self, key, model_name, base_url="https://openrouter.ai/api/v1"):
748
+ self.base_url = "https://openrouter.ai/api/v1"
749
+ self.client = OpenAI(base_url=self.base_url, api_key=key)
750
+ self.model_name = model_name
751
+
rag/llm/cv_model.py CHANGED
@@ -23,6 +23,8 @@ from openai import OpenAI
23
  import os
24
  import base64
25
  from io import BytesIO
 
 
26
 
27
  from api.utils import get_uuid
28
  from api.utils.file_utils import get_project_base_directory
@@ -212,7 +214,7 @@ class GeminiCV(Base):
212
  self.model = GenerativeModel(model_name=self.model_name)
213
  self.model._client = _client
214
  self.lang = lang
215
-
216
  def describe(self, image, max_tokens=2048):
217
  from PIL.Image import open
218
  gen_config = {'max_output_tokens':max_tokens}
@@ -227,6 +229,63 @@ class GeminiCV(Base):
227
  )
228
  return res.text,res.usage_metadata.total_token_count
229
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
230
  class LocalCV(Base):
231
  def __init__(self, key, model_name="glm-4v", lang="Chinese", **kwargs):
232
  pass
 
23
  import os
24
  import base64
25
  from io import BytesIO
26
+ import json
27
+ import requests
28
 
29
  from api.utils import get_uuid
30
  from api.utils.file_utils import get_project_base_directory
 
214
  self.model = GenerativeModel(model_name=self.model_name)
215
  self.model._client = _client
216
  self.lang = lang
217
+
218
  def describe(self, image, max_tokens=2048):
219
  from PIL.Image import open
220
  gen_config = {'max_output_tokens':max_tokens}
 
229
  )
230
  return res.text,res.usage_metadata.total_token_count
231
 
232
+
233
+ class OpenRouterCV(Base):
234
+ def __init__(
235
+ self,
236
+ key,
237
+ model_name,
238
+ lang="Chinese",
239
+ base_url="https://openrouter.ai/api/v1/chat/completions",
240
+ ):
241
+ self.model_name = model_name
242
+ self.lang = lang
243
+ self.base_url = "https://openrouter.ai/api/v1/chat/completions"
244
+ self.key = key
245
+
246
+ def describe(self, image, max_tokens=300):
247
+ b64 = self.image2base64(image)
248
+ response = requests.post(
249
+ url=self.base_url,
250
+ headers={
251
+ "Authorization": f"Bearer {self.key}",
252
+ },
253
+ data=json.dumps(
254
+ {
255
+ "model": self.model_name,
256
+ "messages": self.prompt(b64),
257
+ "max_tokens": max_tokens,
258
+ }
259
+ ),
260
+ )
261
+ response = response.json()
262
+ return (
263
+ response["choices"][0]["message"]["content"].strip(),
264
+ response["usage"]["total_tokens"],
265
+ )
266
+
267
+ def prompt(self, b64):
268
+ return [
269
+ {
270
+ "role": "user",
271
+ "content": [
272
+ {
273
+ "type": "image_url",
274
+ "image_url": {"url": f"data:image/jpeg;base64,{b64}"},
275
+ },
276
+ {
277
+ "type": "text",
278
+ "text": (
279
+ "请用中文详细描述一下图中的内容，比如时间，地点，人物，事情，人物心情等，如果有数据请提取出数据。"
280
+ if self.lang.lower() == "chinese"
281
+ else "Please describe the content of this picture, like where, when, who, what happen. If it has number data, please extract them out."
282
+ ),
283
+ },
284
+ ],
285
+ }
286
+ ]
287
+
288
+
289
  class LocalCV(Base):
290
  def __init__(self, key, model_name="glm-4v", lang="Chinese", **kwargs):
291
  pass
web/src/assets/svg/llm/open-router.svg ADDED
web/src/pages/user-setting/setting-model/index.tsx CHANGED
@@ -63,6 +63,7 @@ const IconMap = {
63
  Bedrock: 'bedrock',
64
  Gemini:'gemini',
65
  Groq: 'Groq',
 
66
  };
67
 
68
  const LlmIcon = ({ name }: { name: string }) => {
 
63
  Bedrock: 'bedrock',
64
  Gemini:'gemini',
65
  Groq: 'Groq',
66
+ OpenRouter:'open-router'
67
  };
68
 
69
  const LlmIcon = ({ name }: { name: string }) => {