1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623import logging
import re
import time
from abc import abstractmethod
from collections.abc import Generator, Mapping
from typing import Union
from pydantic import ConfigDict
from dify_plugin.entities.model import (
ModelPropertyKey,
ModelType,
ParameterRule,
ParameterType,
PriceType,
)
from dify_plugin.entities.model.llm import (
LLMMode,
LLMResult,
LLMResultChunk,
LLMResultChunkDelta,
LLMUsage,
)
from dify_plugin.entities.model.message import (
AssistantPromptMessage,
PromptMessage,
PromptMessageContentType,
PromptMessageTool,
SystemPromptMessage,
UserPromptMessage,
)
from dify_plugin.interfaces.model.ai_model import AIModel
logger = logging.getLogger(__name__)
class LargeLanguageModel(AIModel):
"""
Model class for large language model.
"""
model_type: ModelType = ModelType.LLM
# pydantic configs
model_config = ConfigDict(protected_namespaces=())
############################################################
# Methods that can be implemented by plugin #
############################################################
@abstractmethod
def _invoke(
self,
model: str,
credentials: dict,
prompt_messages: list[PromptMessage],
model_parameters: dict,
tools: list[PromptMessageTool] | None = None,
stop: list[str] | None = None,
stream: bool = True,
user: str | None = None,
) -> Union[LLMResult, Generator[LLMResultChunk, None, None]]:
"""
Invoke large language model
:param model: model name
:param credentials: model credentials
:param prompt_messages: prompt messages
:param model_parameters: model parameters
:param tools: tools for tool calling
:param stop: stop words
:param stream: is stream response
:param user: unique user id
:return: full response or stream response chunk generator result
"""
raise NotImplementedError
@abstractmethod
def get_num_tokens(
self,
model: str,
credentials: dict,
prompt_messages: list[PromptMessage],
tools: list[PromptMessageTool] | None = None,
) -> int:
"""
Get number of tokens for given prompt messages
:param model: model name
:param credentials: model credentials
:param prompt_messages: prompt messages
:param tools: tools for tool calling
:return:
"""
raise NotImplementedError
############################################################
# For plugin implementation use only #
############################################################
def enforce_stop_tokens(self, text: str, stop: list[str]) -> str:
"""Cut off the text as soon as any stop words occur."""
return re.split("|".join(stop), text, maxsplit=1)[0]
def get_parameter_rules(self, model: str, credentials: dict) -> list[ParameterRule]:
"""
Get parameter rules
:param model: model name
:param credentials: model credentials
:return: parameter rules
"""
model_schema = self.get_model_schema(model, credentials)
if model_schema:
return model_schema.parameter_rules
return []
def get_model_mode(self, model: str, credentials: Mapping | None = None) -> LLMMode:
"""
Get model mode
:param model: model name
:param credentials: model credentials
:return: model mode
"""
model_schema = self.get_model_schema(model, credentials)
mode = LLMMode.CHAT
if model_schema and model_schema.model_properties.get(ModelPropertyKey.MODE):
mode = LLMMode.value_of(model_schema.model_properties[ModelPropertyKey.MODE])
return mode
def _calc_response_usage(
self, model: str, credentials: dict, prompt_tokens: int, completion_tokens: int
) -> LLMUsage:
"""
Calculate response usage
:param model: model name
:param credentials: model credentials
:param prompt_tokens: prompt tokens
:param completion_tokens: completion tokens
:return: usage
"""
# get prompt price info
prompt_price_info = self.get_price(
model=model,
credentials=credentials,
price_type=PriceType.INPUT,
tokens=prompt_tokens,
)
# get completion price info
completion_price_info = self.get_price(
model=model,
credentials=credentials,
price_type=PriceType.OUTPUT,
tokens=completion_tokens,
)
# Calculate latency from thread-local storage
current_time = time.perf_counter()
latency = current_time - self.started_at
# transform usage
usage = LLMUsage(
prompt_tokens=prompt_tokens,
prompt_unit_price=prompt_price_info.unit_price,
prompt_price_unit=prompt_price_info.unit,
prompt_price=prompt_price_info.total_amount,
completion_tokens=completion_tokens,
completion_unit_price=completion_price_info.unit_price,
completion_price_unit=completion_price_info.unit,
completion_price=completion_price_info.total_amount,
total_tokens=prompt_tokens + completion_tokens,
total_price=prompt_price_info.total_amount + completion_price_info.total_amount,
currency=prompt_price_info.currency,
latency=latency,
)
return usage
def _validate_and_filter_model_parameters(self, model: str, model_parameters: dict, credentials: dict) -> dict:
"""
Validate model parameters
:param model: model name
:param model_parameters: model parameters
:param credentials: model credentials
:return:
"""
parameter_rules = self.get_parameter_rules(model, credentials)
# validate model parameters
filtered_model_parameters = {}
for parameter_rule in parameter_rules:
parameter_name = parameter_rule.name
parameter_value = model_parameters.get(parameter_name)
if parameter_value is None:
if parameter_rule.use_template and parameter_rule.use_template in model_parameters:
# if parameter value is None, use template value variable name instead
parameter_value = model_parameters[parameter_rule.use_template]
else:
if parameter_rule.required:
if parameter_rule.default is not None:
filtered_model_parameters[parameter_name] = parameter_rule.default
continue
else:
raise ValueError(f"Model Parameter {parameter_name} is required.")
else:
continue
# validate parameter value type
if parameter_rule.type == ParameterType.INT:
if not isinstance(parameter_value, int):
raise ValueError(f"Model Parameter {parameter_name} should be int.")
# validate parameter value range
if parameter_rule.min is not None and parameter_value < parameter_rule.min:
raise ValueError(
f"Model Parameter {parameter_name} should be greater than or equal to {parameter_rule.min}."
)
if parameter_rule.max is not None and parameter_value > parameter_rule.max:
raise ValueError(
f"Model Parameter {parameter_name} should be less than or equal to {parameter_rule.max}."
)
elif parameter_rule.type == ParameterType.FLOAT:
if not isinstance(parameter_value, float | int):
raise ValueError(f"Model Parameter {parameter_name} should be float.")
# validate parameter value precision
if parameter_rule.precision is not None:
if parameter_rule.precision == 0:
if parameter_value != int(parameter_value):
raise ValueError(f"Model Parameter {parameter_name} should be int.")
else:
if parameter_value != round(parameter_value, parameter_rule.precision):
raise ValueError(
f"Model Parameter {parameter_name} should be round to "
f"{parameter_rule.precision} decimal places."
)
# validate parameter value range
if parameter_rule.min is not None and parameter_value < parameter_rule.min:
raise ValueError(
f"Model Parameter {parameter_name} should be greater than or equal to {parameter_rule.min}."
)
if parameter_rule.max is not None and parameter_value > parameter_rule.max:
raise ValueError(
f"Model Parameter {parameter_name} should be less than or equal to {parameter_rule.max}."
)
elif parameter_rule.type == ParameterType.BOOLEAN:
if not isinstance(parameter_value, bool):
raise ValueError(f"Model Parameter {parameter_name} should be bool.")
elif parameter_rule.type == ParameterType.STRING:
if not isinstance(parameter_value, str):
raise ValueError(f"Model Parameter {parameter_name} should be string.")
# validate options
if parameter_rule.options and parameter_value not in parameter_rule.options:
raise ValueError(f"Model Parameter {parameter_name} should be one of {parameter_rule.options}.")
elif parameter_rule.type == ParameterType.TEXT:
if not isinstance(parameter_value, str):
raise ValueError(f"Model Parameter {parameter_name} should be string.")
else:
raise ValueError(f"Model Parameter {parameter_name} type {parameter_rule.type} is not supported.")
filtered_model_parameters[parameter_name] = parameter_value
return filtered_model_parameters
def _code_block_mode_wrapper(
self,
model: str,
credentials: dict,
prompt_messages: list[PromptMessage],
model_parameters: dict,
tools: list[PromptMessageTool] | None = None,
stop: list[str] | None = None,
stream: bool = True,
user: str | None = None,
) -> Union[LLMResult, Generator[LLMResultChunk, None, None]]:
"""
Code block mode wrapper, ensure the response is a code block with output markdown quote
:param model: model name
:param credentials: model credentials
:param prompt_messages: prompt messages
:param model_parameters: model parameters
:param tools: tools for tool calling
:param stop: stop words
:param stream: is stream response
:param user: unique user id
:param callbacks: callbacks
:return: full response or stream response chunk generator result
"""
block_prompts = """You should always follow the instructions and output a valid {{block}} object.
The structure of the {{block}} object you can found in the instructions, use {"answer": "$your_answer"} as the default structure
if you are not sure about the structure.
<instructions>
{{instructions}}
</instructions>
""" # noqa: E501
code_block = model_parameters.get("response_format", "")
if not code_block:
return self._invoke(
model=model,
credentials=credentials,
prompt_messages=prompt_messages,
model_parameters=model_parameters,
tools=tools,
stop=stop,
stream=stream,
user=user,
)
model_parameters.pop("response_format")
stop = stop or []
stop.extend(["\n```", "```\n"])
block_prompts = block_prompts.replace("{{block}}", code_block)
# check if there is a system message
if len(prompt_messages) > 0 and isinstance(prompt_messages[0], SystemPromptMessage):
# override the system message
prompt_messages[0] = SystemPromptMessage(
content=block_prompts.replace("{{instructions}}", str(prompt_messages[0].content))
)
else:
# insert the system message
prompt_messages.insert(
0,
SystemPromptMessage(
content=block_prompts.replace(
"{{instructions}}",
f"Please output a valid {code_block} object.",
)
),
)
if len(prompt_messages) > 0 and isinstance(prompt_messages[-1], UserPromptMessage):
# add ```JSON\n to the last text message
if isinstance(prompt_messages[-1].content, str):
prompt_messages[-1].content += f"\n```{code_block}\n"
elif isinstance(prompt_messages[-1].content, list):
for i in range(len(prompt_messages[-1].content) - 1, -1, -1):
if prompt_messages[-1].content[i].type == PromptMessageContentType.TEXT:
prompt_messages[-1].content[i].data += f"\n```{code_block}\n"
break
else:
# append a user message
prompt_messages.append(UserPromptMessage(content=f"```{code_block}\n"))
response = self._invoke(
model=model,
credentials=credentials,
prompt_messages=prompt_messages,
model_parameters=model_parameters,
tools=tools,
stop=stop,
stream=stream,
user=user,
)
if isinstance(response, Generator):
first_chunk = next(response)
def new_generator():
yield first_chunk
yield from response
if (
first_chunk.delta.message.content
and isinstance(first_chunk.delta.message.content, str)
and first_chunk.delta.message.content.startswith("`")
):
return self._code_block_mode_stream_processor_with_backtick(
model=model,
prompt_messages=prompt_messages,
input_generator=new_generator(),
)
else:
return self._code_block_mode_stream_processor(
model=model,
prompt_messages=prompt_messages,
input_generator=new_generator(),
)
return response
def _code_block_mode_stream_processor(
self,
model: str,
prompt_messages: list[PromptMessage],
input_generator: Generator[LLMResultChunk, None, None],
) -> Generator[LLMResultChunk, None, None]:
"""
Code block mode stream processor, ensure the response is a code block with output markdown quote
:param model: model name
:param prompt_messages: prompt messages
:param input_generator: input generator
:return: output generator
"""
state = "normal"
backtick_count = 0
for piece in input_generator:
if piece.delta.message.content:
content = piece.delta.message.content
piece.delta.message.content = ""
yield piece
piece = content
else:
yield piece
continue
new_piece: str = ""
for char in piece:
char = str(char)
if state == "normal":
if char == "`":
state = "in_backticks"
backtick_count = 1
else:
new_piece += char
elif state == "in_backticks":
if char == "`":
backtick_count += 1
if backtick_count == 3:
state = "skip_content"
backtick_count = 0
else:
new_piece += "`" * backtick_count + char
state = "normal"
backtick_count = 0
elif state == "skip_content" and char.isspace():
state = "normal"
if new_piece:
yield LLMResultChunk(
model=model,
delta=LLMResultChunkDelta(
index=0,
message=AssistantPromptMessage(content=new_piece, tool_calls=[]),
),
)
def _code_block_mode_stream_processor_with_backtick(
self,
model: str,
prompt_messages: list,
input_generator: Generator[LLMResultChunk, None, None],
) -> Generator[LLMResultChunk, None, None]:
"""
Code block mode stream processor, ensure the response is a code block with output markdown quote.
This version skips the language identifier that follows the opening triple backticks.
:param model: model name
:param prompt_messages: prompt messages
:param input_generator: input generator
:return: output generator
"""
state = "search_start"
backtick_count = 0
for piece in input_generator:
if piece.delta.message.content:
content = piece.delta.message.content
# Reset content to ensure we're only processing and yielding the relevant parts
piece.delta.message.content = ""
# Yield a piece with cleared content before processing it to maintain the generator structure
yield piece
piece = content
else:
# Yield pieces without content directly
yield piece
continue
if state == "done":
continue
new_piece: str = ""
for char in piece:
if state == "search_start":
if char == "`":
backtick_count += 1
if backtick_count == 3:
state = "skip_language"
backtick_count = 0
else:
backtick_count = 0
elif state == "skip_language":
# Skip everything until the first newline, marking the end of the language identifier
if char == "\n":
state = "in_code_block"
elif state == "in_code_block":
if char == "`":
backtick_count += 1
if backtick_count == 3:
state = "done"
break
else:
if backtick_count > 0:
# If backticks were counted but we're still collecting content, it was a false start
new_piece += "`" * backtick_count
backtick_count = 0
new_piece += str(char)
elif state == "done":
break
if new_piece:
# Only yield content collected within the code block
yield LLMResultChunk(
model=model,
delta=LLMResultChunkDelta(
index=0,
message=AssistantPromptMessage(content=new_piece, tool_calls=[]),
),
)
def _wrap_thinking_by_reasoning_content(self, delta: dict, is_reasoning: bool) -> tuple[str, bool]:
"""
If the reasoning response is from delta.get("reasoning_content"), we wrap
it with HTML think tag.
:param delta: delta dictionary from LLM streaming response
:param is_reasoning: is reasoning
:return: tuple of (processed_content, is_reasoning)
"""
content = delta.get("content") or ""
reasoning_content = delta.get("reasoning_content")
output = content
if reasoning_content:
if not is_reasoning:
output = "<think>\n" + reasoning_content
is_reasoning = True
else:
output = reasoning_content
else:
if is_reasoning:
is_reasoning = False
if not reasoning_content:
output = "\n</think>"
if content:
output += content
return output, is_reasoning
############################################################
# For executor use only #
############################################################
def invoke(
self,
model: str,
credentials: dict,
prompt_messages: list[PromptMessage],
model_parameters: dict | None = None,
tools: list[PromptMessageTool] | None = None,
stop: list[str] | None = None,
stream: bool = True,
user: str | None = None,
) -> Generator[LLMResultChunk, None, None]:
"""
Invoke large language model
:param model: model name
:param credentials: model credentials
:param prompt_messages: prompt messages
:param model_parameters: model parameters
:param tools: tools for tool calling
:param stop: stop words
:param stream: is stream response
:param user: unique user id
:param callbacks: callbacks
:return: full response or stream response chunk generator result
"""
# validate and filter model parameters
if model_parameters is None:
model_parameters = {}
model_parameters = self._validate_and_filter_model_parameters(model, model_parameters, credentials)
with self.timing_context():
try:
if "response_format" in model_parameters and model_parameters["response_format"] in {"JSON", "XML"}:
result = self._code_block_mode_wrapper(
model=model,
credentials=credentials,
prompt_messages=prompt_messages,
model_parameters=model_parameters,
tools=tools,
stop=stop,
stream=stream,
user=user,
)
else:
result = self._invoke(
model,
credentials,
prompt_messages,
model_parameters,
tools,
stop,
stream,
user,
)
except Exception as e:
raise self._transform_invoke_error(e) from e
if isinstance(result, LLMResult):
yield result.to_llm_result_chunk()
else:
# NOTE: `yield from` cannot been replaced by `return` because of `timing_context`
yield from result