Update src/txagent/txagent.py
Browse files- src/txagent/txagent.py +470 -204
src/txagent/txagent.py
CHANGED
@@ -12,22 +12,23 @@ from tooluniverse import ToolUniverse
|
|
12 |
from gradio import ChatMessage
|
13 |
from .toolrag import ToolRAGModel
|
14 |
import torch
|
|
|
15 |
import logging
|
16 |
-
|
17 |
logger = logging.getLogger(__name__)
|
18 |
logging.basicConfig(level=logging.INFO)
|
19 |
|
20 |
from .utils import NoRepeatSentenceProcessor, ReasoningTraceChecker, tool_result_format
|
21 |
|
|
|
22 |
class TxAgent:
|
23 |
def __init__(self, model_name,
|
24 |
rag_model_name,
|
25 |
-
tool_files_dict=None,
|
26 |
-
enable_finish=
|
27 |
-
enable_rag=
|
28 |
enable_summary=False,
|
29 |
init_rag_num=0,
|
30 |
-
step_rag_num=
|
31 |
summary_mode='step',
|
32 |
summary_skip_last_k=0,
|
33 |
summary_context_length=None,
|
@@ -44,9 +45,10 @@ class TxAgent:
|
|
44 |
self.rag_model_name = rag_model_name
|
45 |
self.tool_files_dict = tool_files_dict
|
46 |
self.model = None
|
47 |
-
self.rag_model = ToolRAGModel(rag_model_name)
|
48 |
self.tooluniverse = None
|
49 |
-
self.
|
|
|
50 |
self.self_prompt = "Strictly follow the instruction."
|
51 |
self.chat_prompt = "You are helpful assistant to chat with the user."
|
52 |
self.enable_finish = enable_finish
|
@@ -66,28 +68,26 @@ class TxAgent:
|
|
66 |
|
67 |
def init_model(self):
|
68 |
self.load_models()
|
69 |
-
|
70 |
-
|
71 |
-
self.load_tool_desc_embedding()
|
72 |
|
73 |
def print_self_values(self):
|
74 |
for attr, value in self.__dict__.items():
|
75 |
-
|
76 |
|
77 |
def load_models(self, model_name=None):
|
78 |
if model_name is not None:
|
79 |
if model_name == self.model_name:
|
80 |
return f"The model {model_name} is already loaded."
|
81 |
self.model_name = model_name
|
82 |
-
|
|
|
83 |
self.chat_template = Template(self.model.get_tokenizer().chat_template)
|
84 |
self.tokenizer = self.model.get_tokenizer()
|
|
|
85 |
return f"Model {model_name} loaded successfully."
|
86 |
|
87 |
def load_tooluniverse(self):
|
88 |
-
if self.tool_files_dict is None and not self.enable_rag:
|
89 |
-
logger.info("Skipping tool universe loading: RAG disabled and no tool files.")
|
90 |
-
return
|
91 |
self.tooluniverse = ToolUniverse(tool_files=self.tool_files_dict)
|
92 |
self.tooluniverse.load_tools()
|
93 |
special_tools = self.tooluniverse.prepare_tool_prompts(
|
@@ -95,24 +95,20 @@ class TxAgent:
|
|
95 |
self.special_tools_name = [tool['name'] for tool in special_tools]
|
96 |
|
97 |
def load_tool_desc_embedding(self):
|
98 |
-
|
99 |
-
self.rag_model.load_tool_desc_embedding(self.tooluniverse)
|
100 |
|
101 |
def rag_infer(self, query, top_k=5):
|
102 |
-
if not self.enable_rag or not self.rag_model:
|
103 |
-
return []
|
104 |
return self.rag_model.rag_infer(query, top_k)
|
105 |
|
106 |
def initialize_tools_prompt(self, call_agent, call_agent_level, message):
|
107 |
picked_tools_prompt = []
|
108 |
-
if not self.enable_rag:
|
109 |
-
return picked_tools_prompt, call_agent_level
|
110 |
picked_tools_prompt = self.add_special_tools(
|
111 |
picked_tools_prompt, call_agent=call_agent)
|
112 |
if call_agent:
|
113 |
call_agent_level += 1
|
114 |
if call_agent_level >= 2:
|
115 |
call_agent = False
|
|
|
116 |
if not call_agent:
|
117 |
picked_tools_prompt += self.tool_RAG(
|
118 |
message=message, rag_num=self.init_rag_num)
|
@@ -121,12 +117,13 @@ class TxAgent:
|
|
121 |
def initialize_conversation(self, message, conversation=None, history=None):
|
122 |
if conversation is None:
|
123 |
conversation = []
|
|
|
124 |
conversation = self.set_system_prompt(
|
125 |
conversation, self.prompt_multi_step)
|
126 |
if history is not None:
|
127 |
if len(history) == 0:
|
128 |
conversation = []
|
129 |
-
|
130 |
else:
|
131 |
for i in range(len(history)):
|
132 |
if history[i]['role'] == 'user':
|
@@ -138,7 +135,9 @@ class TxAgent:
|
|
138 |
if i == len(history)-1 and history[i]['role'] == 'assistant':
|
139 |
conversation.append(
|
140 |
{"role": "assistant", "content": history[i]['content']})
|
|
|
141 |
conversation.append({"role": "user", "content": message})
|
|
|
142 |
return conversation
|
143 |
|
144 |
def tool_RAG(self, message=None,
|
@@ -146,52 +145,60 @@ class TxAgent:
|
|
146 |
existing_tools_prompt=[],
|
147 |
rag_num=5,
|
148 |
return_call_result=False):
|
149 |
-
|
150 |
-
return [] if not return_call_result else ([], [])
|
151 |
-
extra_factor = 30
|
152 |
if picked_tool_names is None:
|
153 |
assert picked_tool_names is not None or message is not None
|
154 |
picked_tool_names = self.rag_infer(
|
155 |
message, top_k=rag_num*extra_factor)
|
156 |
-
|
|
|
|
|
|
|
|
|
157 |
picked_tool_names_no_special = picked_tool_names_no_special[:rag_num]
|
158 |
picked_tool_names = picked_tool_names_no_special[:rag_num]
|
|
|
159 |
picked_tools = self.tooluniverse.get_tool_by_name(picked_tool_names)
|
160 |
-
picked_tools_prompt = self.tooluniverse.prepare_tool_prompts(
|
|
|
161 |
if return_call_result:
|
162 |
return picked_tools_prompt, picked_tool_names
|
163 |
return picked_tools_prompt
|
164 |
|
165 |
def add_special_tools(self, tools, call_agent=False):
|
166 |
-
if
|
167 |
-
|
168 |
-
|
169 |
-
|
170 |
-
|
171 |
-
|
172 |
-
|
173 |
-
|
174 |
-
|
175 |
-
|
176 |
-
|
177 |
-
|
178 |
-
|
179 |
-
|
180 |
-
|
181 |
-
|
182 |
-
|
|
|
|
|
|
|
|
|
183 |
return tools
|
184 |
|
185 |
def add_finish_tools(self, tools):
|
186 |
-
|
187 |
-
|
188 |
-
|
189 |
-
logger.info("Finish tool is added")
|
190 |
return tools
|
191 |
|
192 |
def set_system_prompt(self, conversation, sys_prompt):
|
193 |
if len(conversation) == 0:
|
194 |
-
conversation.append(
|
|
|
195 |
else:
|
196 |
conversation[0] = {"role": "system", "content": sys_prompt}
|
197 |
return conversation
|
@@ -203,15 +210,15 @@ class TxAgent:
|
|
203 |
call_agent=False,
|
204 |
call_agent_level=None,
|
205 |
temperature=None):
|
206 |
-
|
207 |
-
|
208 |
-
|
209 |
call_results = []
|
210 |
special_tool_call = ''
|
211 |
if function_call_json is not None:
|
212 |
if isinstance(function_call_json, list):
|
213 |
for i in range(len(function_call_json)):
|
214 |
-
|
215 |
if function_call_json[i]["name"] == 'Finish':
|
216 |
special_tool_call = 'Finish'
|
217 |
break
|
@@ -239,12 +246,14 @@ class TxAgent:
|
|
239 |
else:
|
240 |
call_result = call_result.split('[FinalAnswer]')[-1].strip()
|
241 |
else:
|
242 |
-
call_result = "Error: The CallAgent has been disabled."
|
243 |
else:
|
244 |
-
call_result = self.tooluniverse.run_one_function(
|
|
|
|
|
245 |
call_id = self.tooluniverse.call_id_gen()
|
246 |
function_call_json[i]["call_id"] = call_id
|
247 |
-
|
248 |
call_results.append({
|
249 |
"role": "tool",
|
250 |
"content": json.dumps({"tool_name": function_call_json[i]["name"], "content": call_result, "call_id": call_id})
|
@@ -252,30 +261,33 @@ class TxAgent:
|
|
252 |
else:
|
253 |
call_results.append({
|
254 |
"role": "tool",
|
255 |
-
"content": json.dumps({"content": "Not a valid function call."})
|
256 |
})
|
|
|
257 |
revised_messages = [{
|
258 |
"role": "assistant",
|
259 |
"content": message.strip(),
|
260 |
"tool_calls": json.dumps(function_call_json)
|
261 |
}] + call_results
|
|
|
|
|
262 |
return revised_messages, existing_tools_prompt, special_tool_call
|
263 |
|
264 |
def run_function_call_stream(self, fcall_str,
|
265 |
-
|
266 |
-
|
267 |
-
|
268 |
-
|
269 |
-
|
270 |
-
|
271 |
-
|
272 |
-
|
273 |
-
|
274 |
-
|
275 |
-
function_call_json, message = self.tooluniverse.extract_function_call_json(fcall_str, return_message=return_message, verbose=False)
|
276 |
call_results = []
|
277 |
special_tool_call = ''
|
278 |
-
|
|
|
279 |
if function_call_json is not None:
|
280 |
if isinstance(function_call_json, list):
|
281 |
for i in range(len(function_call_json)):
|
@@ -303,21 +315,25 @@ class TxAgent:
|
|
303 |
"\nYou must follow the following plan to answer the question: " +
|
304 |
str(solution_plan)
|
305 |
)
|
306 |
-
sub_agent_task = "Sub TxAgent plan: " +
|
|
|
307 |
call_result = yield from self.run_gradio_chat(
|
308 |
full_message, history=[], temperature=temperature,
|
309 |
max_new_tokens=1024, max_token=99999,
|
310 |
call_agent=False, call_agent_level=call_agent_level,
|
311 |
conversation=None,
|
312 |
sub_agent_task=sub_agent_task)
|
|
|
313 |
if call_result is not None and isinstance(call_result, str):
|
314 |
call_result = call_result.split('[FinalAnswer]')[-1]
|
315 |
else:
|
316 |
call_result = "⚠️ No content returned from sub-agent."
|
317 |
else:
|
318 |
-
call_result = "Error: The CallAgent has been disabled."
|
319 |
else:
|
320 |
-
call_result = self.tooluniverse.run_one_function(
|
|
|
|
|
321 |
call_id = self.tooluniverse.call_id_gen()
|
322 |
function_call_json[i]["call_id"] = call_id
|
323 |
call_results.append({
|
@@ -334,25 +350,34 @@ class TxAgent:
|
|
334 |
else:
|
335 |
call_results.append({
|
336 |
"role": "tool",
|
337 |
-
"content": json.dumps({"content": "Not a valid function call."})
|
338 |
})
|
|
|
339 |
revised_messages = [{
|
340 |
"role": "assistant",
|
341 |
"content": message.strip(),
|
342 |
"tool_calls": json.dumps(function_call_json)
|
343 |
}] + call_results
|
344 |
-
|
|
|
|
|
|
|
|
|
|
|
345 |
|
346 |
def get_answer_based_on_unfinished_reasoning(self, conversation, temperature, max_new_tokens, max_token, outputs=None):
|
347 |
-
if conversation[-1]['role'] == '
|
348 |
-
conversation.append(
|
|
|
|
|
|
|
349 |
last_outputs_str = self.llm_infer(messages=conversation,
|
350 |
-
|
351 |
-
|
352 |
-
|
353 |
-
|
354 |
-
|
355 |
-
|
356 |
return last_outputs_str
|
357 |
|
358 |
def run_multistep_agent(self, message: str,
|
@@ -362,10 +387,20 @@ class TxAgent:
|
|
362 |
max_round: int = 20,
|
363 |
call_agent=False,
|
364 |
call_agent_level=0) -> str:
|
365 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
366 |
picked_tools_prompt, call_agent_level = self.initialize_tools_prompt(
|
367 |
call_agent, call_agent_level, message)
|
368 |
conversation = self.initialize_conversation(message)
|
|
|
369 |
outputs = []
|
370 |
last_outputs = []
|
371 |
next_round = True
|
@@ -374,6 +409,7 @@ class TxAgent:
|
|
374 |
token_overflow = False
|
375 |
enable_summary = False
|
376 |
last_status = {}
|
|
|
377 |
if self.enable_checker:
|
378 |
checker = ReasoningTraceChecker(message, conversation)
|
379 |
try:
|
@@ -387,60 +423,73 @@ class TxAgent:
|
|
387 |
call_agent=call_agent,
|
388 |
call_agent_level=call_agent_level,
|
389 |
temperature=temperature)
|
|
|
390 |
if special_tool_call == 'Finish':
|
391 |
next_round = False
|
392 |
conversation.extend(function_call_messages)
|
|
|
|
|
|
|
393 |
content = function_call_messages[0]['content']
|
394 |
if content is None:
|
395 |
-
return "❌ No content returned after Finish."
|
396 |
return content.split('[FinalAnswer]')[-1]
|
|
|
397 |
if (self.enable_summary or token_overflow) and not call_agent:
|
|
|
|
|
398 |
enable_summary = True
|
399 |
last_status = self.function_result_summary(
|
400 |
conversation, status=last_status, enable_summary=enable_summary)
|
401 |
-
|
|
|
402 |
conversation.extend(function_call_messages)
|
403 |
-
outputs.append(tool_result_format(
|
|
|
404 |
else:
|
405 |
next_round = False
|
406 |
-
conversation.extend(
|
|
|
407 |
return ''.join(last_outputs).replace("</s>", "")
|
408 |
if self.enable_checker:
|
409 |
good_status, wrong_info = checker.check_conversation()
|
410 |
if not good_status:
|
411 |
-
|
|
|
|
|
412 |
break
|
413 |
last_outputs = []
|
414 |
outputs.append("### TxAgent:\n")
|
415 |
-
last_outputs_str, token_overflow = self.llm_infer(
|
416 |
-
|
417 |
-
|
418 |
-
|
419 |
-
|
420 |
-
|
421 |
-
max_token=max_token,
|
422 |
-
check_token_status=True)
|
423 |
if last_outputs_str is None:
|
424 |
-
|
425 |
if self.force_finish:
|
426 |
-
return self.get_answer_based_on_unfinished_reasoning(
|
427 |
-
|
428 |
-
|
429 |
-
|
|
|
430 |
if max_round == current_round:
|
431 |
-
|
432 |
if self.force_finish:
|
433 |
-
return self.get_answer_based_on_unfinished_reasoning(
|
434 |
-
|
435 |
-
|
|
|
436 |
except Exception as e:
|
437 |
-
|
438 |
if self.force_finish:
|
439 |
-
return self.get_answer_based_on_unfinished_reasoning(
|
440 |
-
|
441 |
-
|
442 |
|
443 |
def build_logits_processor(self, messages, llm):
|
|
|
444 |
tokenizer = llm.get_tokenizer()
|
445 |
if self.avoid_repeat and len(messages) > 2:
|
446 |
assistant_messages = []
|
@@ -449,49 +498,66 @@ class TxAgent:
|
|
449 |
assistant_messages.append(messages[-i]['content'])
|
450 |
if len(assistant_messages) == 2:
|
451 |
break
|
452 |
-
forbidden_ids = [tokenizer.encode(
|
453 |
-
|
454 |
-
|
|
|
|
|
455 |
|
456 |
def llm_infer(self, messages, temperature=0.1, tools=None,
|
457 |
output_begin_string=None, max_new_tokens=2048,
|
458 |
max_token=None, skip_special_tokens=True,
|
459 |
-
model=None, tokenizer=None, terminators=None, seed=None,
|
460 |
-
|
461 |
if model is None:
|
462 |
model = self.model
|
|
|
463 |
logits_processor = self.build_logits_processor(messages, model)
|
464 |
sampling_params = SamplingParams(
|
465 |
temperature=temperature,
|
466 |
max_tokens=max_new_tokens,
|
|
|
467 |
seed=seed if seed is not None else self.seed,
|
468 |
)
|
|
|
469 |
prompt = self.chat_template.render(
|
470 |
messages=messages, tools=tools, add_generation_prompt=True)
|
471 |
if output_begin_string is not None:
|
472 |
prompt += output_begin_string
|
|
|
473 |
if check_token_status and max_token is not None:
|
474 |
token_overflow = False
|
475 |
-
num_input_tokens = len(self.tokenizer.encode(
|
476 |
-
|
477 |
-
|
478 |
-
|
479 |
-
|
480 |
-
|
481 |
-
|
482 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
483 |
output = output[0].outputs[0].text
|
484 |
-
|
485 |
if check_token_status and max_token is not None:
|
486 |
return output, token_overflow
|
|
|
487 |
return output
|
488 |
|
489 |
def run_self_agent(self, message: str,
|
490 |
-
|
491 |
-
|
492 |
-
|
493 |
-
|
494 |
-
|
|
|
|
|
495 |
conversation.append({"role": "user", "content": message})
|
496 |
return self.llm_infer(messages=conversation,
|
497 |
temperature=temperature,
|
@@ -502,8 +568,10 @@ class TxAgent:
|
|
502 |
temperature: float,
|
503 |
max_new_tokens: int,
|
504 |
max_token: int) -> str:
|
505 |
-
|
506 |
-
|
|
|
|
|
507 |
conversation.append({"role": "user", "content": message})
|
508 |
return self.llm_infer(messages=conversation,
|
509 |
temperature=temperature,
|
@@ -511,106 +579,155 @@ class TxAgent:
|
|
511 |
max_new_tokens=max_new_tokens, max_token=max_token)
|
512 |
|
513 |
def run_format_agent(self, message: str,
|
514 |
-
|
515 |
-
|
516 |
-
|
517 |
-
|
518 |
-
|
|
|
519 |
if '[FinalAnswer]' in answer:
|
520 |
possible_final_answer = answer.split("[FinalAnswer]")[-1]
|
521 |
elif "\n\n" in answer:
|
522 |
possible_final_answer = answer.split("\n\n")[-1]
|
523 |
else:
|
524 |
possible_final_answer = answer.strip()
|
525 |
-
if len(possible_final_answer) == 1
|
526 |
-
|
527 |
-
|
528 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
529 |
conversation = []
|
530 |
-
format_prompt = "You are helpful assistant to transform the answer to 'A', 'B', 'C', 'D'."
|
531 |
conversation = self.set_system_prompt(conversation, format_prompt)
|
532 |
-
conversation.append({"role": "user", "content": message +
|
|
|
533 |
return self.llm_infer(messages=conversation,
|
534 |
temperature=temperature,
|
535 |
tools=None,
|
536 |
max_new_tokens=max_new_tokens, max_token=max_token)
|
537 |
|
538 |
def run_summary_agent(self, thought_calls: str,
|
539 |
-
|
540 |
-
|
541 |
-
|
542 |
-
|
543 |
-
|
544 |
generate_tool_result_summary_training_prompt = """Thought and function calls:
|
545 |
{thought_calls}
|
546 |
Function calls' responses:
|
547 |
\"\"\"
|
548 |
{function_response}
|
549 |
\"\"\"
|
550 |
-
|
551 |
-
|
552 |
-
|
|
|
|
|
|
|
|
|
553 |
output = self.llm_infer(messages=conversation,
|
554 |
temperature=temperature,
|
555 |
tools=None,
|
556 |
max_new_tokens=max_new_tokens, max_token=max_token)
|
|
|
557 |
if '[' in output:
|
558 |
output = output.split('[')[0]
|
559 |
return output
|
560 |
|
561 |
def function_result_summary(self, input_list, status, enable_summary):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
562 |
if 'tool_call_step' not in status:
|
563 |
status['tool_call_step'] = 0
|
|
|
564 |
for idx in range(len(input_list)):
|
565 |
pos_id = len(input_list)-idx-1
|
566 |
-
if input_list[pos_id]['role'] == 'assistant'
|
567 |
-
if '
|
568 |
-
|
|
|
569 |
break
|
570 |
-
|
|
|
|
|
|
|
|
|
|
|
571 |
if not enable_summary:
|
572 |
return status
|
|
|
573 |
if 'summarized_index' not in status:
|
574 |
status['summarized_index'] = 0
|
|
|
575 |
if 'summarized_step' not in status:
|
576 |
status['summarized_step'] = 0
|
|
|
577 |
if 'previous_length' not in status:
|
578 |
status['previous_length'] = 0
|
|
|
579 |
if 'history' not in status:
|
580 |
status['history'] = []
|
|
|
581 |
function_response = ''
|
582 |
-
idx =
|
583 |
current_summarized_index = status['summarized_index']
|
|
|
584 |
status['history'].append(self.summary_mode == 'step' and status['summarized_step']
|
585 |
< status['step']-status['tool_call_step']-self.summary_skip_last_k)
|
|
|
|
|
586 |
while idx < len(input_list):
|
587 |
if (self.summary_mode == 'step' and status['summarized_step'] < status['step']-status['tool_call_step']-self.summary_skip_last_k) or (self.summary_mode == 'length' and status['previous_length'] > self.summary_context_length):
|
|
|
588 |
if input_list[idx]['role'] == 'assistant':
|
589 |
if 'Tool_RAG' in str(input_list[idx]['tool_calls']):
|
590 |
this_thought_calls = None
|
591 |
else:
|
592 |
if len(function_response) != 0:
|
593 |
-
|
594 |
status['summarized_step'] += 1
|
595 |
result_summary = self.run_summary_agent(
|
596 |
thought_calls=this_thought_calls,
|
597 |
function_response=function_response,
|
598 |
temperature=0.1,
|
599 |
max_new_tokens=1024,
|
600 |
-
max_token=99999
|
601 |
-
|
|
|
|
|
|
|
602 |
status['summarized_index'] = last_call_idx + 2
|
603 |
idx += 1
|
|
|
604 |
last_call_idx = idx
|
605 |
-
this_thought_calls = input_list[idx]['content'] +
|
|
|
606 |
function_response = ''
|
|
|
607 |
elif input_list[idx]['role'] == 'tool' and this_thought_calls is not None:
|
608 |
function_response += input_list[idx]['content']
|
609 |
del input_list[idx]
|
610 |
idx -= 1
|
|
|
611 |
else:
|
612 |
break
|
613 |
idx += 1
|
|
|
614 |
if len(function_response) != 0:
|
615 |
status['summarized_step'] += 1
|
616 |
result_summary = self.run_summary_agent(
|
@@ -618,20 +735,30 @@ Generate one summarized sentence about "function calls' responses" with necessar
|
|
618 |
function_response=function_response,
|
619 |
temperature=0.1,
|
620 |
max_new_tokens=1024,
|
621 |
-
max_token=99999
|
|
|
|
|
622 |
tool_calls = json.loads(input_list[last_call_idx]['tool_calls'])
|
623 |
for tool_call in tool_calls:
|
624 |
del tool_call['call_id']
|
625 |
input_list[last_call_idx]['tool_calls'] = json.dumps(tool_calls)
|
626 |
-
input_list.insert(
|
|
|
627 |
status['summarized_index'] = last_call_idx + 2
|
|
|
628 |
return status
|
629 |
|
|
|
|
|
|
|
630 |
def update_parameters(self, **kwargs):
|
631 |
for key, value in kwargs.items():
|
632 |
if hasattr(self, key):
|
633 |
setattr(self, key, value)
|
634 |
-
|
|
|
|
|
|
|
635 |
return updated_attributes
|
636 |
|
637 |
def run_gradio_chat(self, message: str,
|
@@ -645,53 +772,192 @@ Generate one summarized sentence about "function calls' responses" with necessar
|
|
645 |
seed: int = None,
|
646 |
call_agent_level: int = 0,
|
647 |
sub_agent_task: str = None,
|
648 |
-
uploaded_files: list = None):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
649 |
logger.debug(f"[TxAgent] Chat started, message: {message[:100]}...")
|
650 |
print("\033[1;32;40m[TxAgent] Chat started\033[0m")
|
|
|
651 |
if not message or len(message.strip()) < 5:
|
652 |
yield "Please provide a valid message or upload files to analyze."
|
653 |
return "Invalid input."
|
|
|
654 |
if message.startswith("[\U0001f9f0 Tool_RAG") or message.startswith("⚒️"):
|
655 |
return ""
|
656 |
-
|
657 |
-
|
658 |
-
|
659 |
-
|
660 |
-
|
661 |
-
|
662 |
-
|
663 |
-
|
664 |
-
|
665 |
-
|
666 |
-
|
667 |
-
|
668 |
-
|
669 |
-
|
670 |
-
|
671 |
-
|
672 |
-
|
673 |
-
|
674 |
-
|
675 |
-
|
676 |
-
|
677 |
-
|
678 |
-
|
679 |
-
|
680 |
-
|
681 |
-
|
682 |
-
|
683 |
-
|
684 |
-
|
685 |
-
|
686 |
-
|
687 |
-
|
688 |
-
|
689 |
-
|
690 |
-
|
691 |
-
|
692 |
-
|
693 |
-
|
694 |
-
|
695 |
-
|
696 |
-
|
697 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
12 |
from gradio import ChatMessage
|
13 |
from .toolrag import ToolRAGModel
|
14 |
import torch
|
15 |
+
# near the top of txagent.py
|
16 |
import logging
|
|
|
17 |
logger = logging.getLogger(__name__)
|
18 |
logging.basicConfig(level=logging.INFO)
|
19 |
|
20 |
from .utils import NoRepeatSentenceProcessor, ReasoningTraceChecker, tool_result_format
|
21 |
|
22 |
+
|
23 |
class TxAgent:
|
24 |
def __init__(self, model_name,
|
25 |
rag_model_name,
|
26 |
+
tool_files_dict=None, # None leads to the default tool files in ToolUniverse
|
27 |
+
enable_finish=True,
|
28 |
+
enable_rag=True,
|
29 |
enable_summary=False,
|
30 |
init_rag_num=0,
|
31 |
+
step_rag_num=10,
|
32 |
summary_mode='step',
|
33 |
summary_skip_last_k=0,
|
34 |
summary_context_length=None,
|
|
|
45 |
self.rag_model_name = rag_model_name
|
46 |
self.tool_files_dict = tool_files_dict
|
47 |
self.model = None
|
48 |
+
self.rag_model = ToolRAGModel(rag_model_name)
|
49 |
self.tooluniverse = None
|
50 |
+
# self.tool_desc = None
|
51 |
+
self.prompt_multi_step = "You are a helpful assistant that will solve problems through detailed, step-by-step reasoning and actions based on your reasoning. Typically, your actions will use the provided functions. You have access to the following functions."
|
52 |
self.self_prompt = "Strictly follow the instruction."
|
53 |
self.chat_prompt = "You are helpful assistant to chat with the user."
|
54 |
self.enable_finish = enable_finish
|
|
|
68 |
|
69 |
def init_model(self):
|
70 |
self.load_models()
|
71 |
+
self.load_tooluniverse()
|
72 |
+
self.load_tool_desc_embedding()
|
|
|
73 |
|
74 |
def print_self_values(self):
|
75 |
for attr, value in self.__dict__.items():
|
76 |
+
print(f"{attr}: {value}")
|
77 |
|
78 |
def load_models(self, model_name=None):
|
79 |
if model_name is not None:
|
80 |
if model_name == self.model_name:
|
81 |
return f"The model {model_name} is already loaded."
|
82 |
self.model_name = model_name
|
83 |
+
|
84 |
+
self.model = LLM(model=self.model_name)
|
85 |
self.chat_template = Template(self.model.get_tokenizer().chat_template)
|
86 |
self.tokenizer = self.model.get_tokenizer()
|
87 |
+
|
88 |
return f"Model {model_name} loaded successfully."
|
89 |
|
90 |
def load_tooluniverse(self):
|
|
|
|
|
|
|
91 |
self.tooluniverse = ToolUniverse(tool_files=self.tool_files_dict)
|
92 |
self.tooluniverse.load_tools()
|
93 |
special_tools = self.tooluniverse.prepare_tool_prompts(
|
|
|
95 |
self.special_tools_name = [tool['name'] for tool in special_tools]
|
96 |
|
97 |
def load_tool_desc_embedding(self):
|
98 |
+
self.rag_model.load_tool_desc_embedding(self.tooluniverse)
|
|
|
99 |
|
100 |
def rag_infer(self, query, top_k=5):
|
|
|
|
|
101 |
return self.rag_model.rag_infer(query, top_k)
|
102 |
|
103 |
def initialize_tools_prompt(self, call_agent, call_agent_level, message):
|
104 |
picked_tools_prompt = []
|
|
|
|
|
105 |
picked_tools_prompt = self.add_special_tools(
|
106 |
picked_tools_prompt, call_agent=call_agent)
|
107 |
if call_agent:
|
108 |
call_agent_level += 1
|
109 |
if call_agent_level >= 2:
|
110 |
call_agent = False
|
111 |
+
|
112 |
if not call_agent:
|
113 |
picked_tools_prompt += self.tool_RAG(
|
114 |
message=message, rag_num=self.init_rag_num)
|
|
|
117 |
def initialize_conversation(self, message, conversation=None, history=None):
|
118 |
if conversation is None:
|
119 |
conversation = []
|
120 |
+
|
121 |
conversation = self.set_system_prompt(
|
122 |
conversation, self.prompt_multi_step)
|
123 |
if history is not None:
|
124 |
if len(history) == 0:
|
125 |
conversation = []
|
126 |
+
print("clear conversation successfully")
|
127 |
else:
|
128 |
for i in range(len(history)):
|
129 |
if history[i]['role'] == 'user':
|
|
|
135 |
if i == len(history)-1 and history[i]['role'] == 'assistant':
|
136 |
conversation.append(
|
137 |
{"role": "assistant", "content": history[i]['content']})
|
138 |
+
|
139 |
conversation.append({"role": "user", "content": message})
|
140 |
+
|
141 |
return conversation
|
142 |
|
143 |
def tool_RAG(self, message=None,
|
|
|
145 |
existing_tools_prompt=[],
|
146 |
rag_num=5,
|
147 |
return_call_result=False):
|
148 |
+
extra_factor = 30 # Factor to retrieve more than rag_num
|
|
|
|
|
149 |
if picked_tool_names is None:
|
150 |
assert picked_tool_names is not None or message is not None
|
151 |
picked_tool_names = self.rag_infer(
|
152 |
message, top_k=rag_num*extra_factor)
|
153 |
+
|
154 |
+
picked_tool_names_no_special = []
|
155 |
+
for tool in picked_tool_names:
|
156 |
+
if tool not in self.special_tools_name:
|
157 |
+
picked_tool_names_no_special.append(tool)
|
158 |
picked_tool_names_no_special = picked_tool_names_no_special[:rag_num]
|
159 |
picked_tool_names = picked_tool_names_no_special[:rag_num]
|
160 |
+
|
161 |
picked_tools = self.tooluniverse.get_tool_by_name(picked_tool_names)
|
162 |
+
picked_tools_prompt = self.tooluniverse.prepare_tool_prompts(
|
163 |
+
picked_tools)
|
164 |
if return_call_result:
|
165 |
return picked_tools_prompt, picked_tool_names
|
166 |
return picked_tools_prompt
|
167 |
|
168 |
def add_special_tools(self, tools, call_agent=False):
|
169 |
+
if self.enable_finish:
|
170 |
+
tools.append(self.tooluniverse.get_one_tool_by_one_name(
|
171 |
+
'Finish', return_prompt=True))
|
172 |
+
print("Finish tool is added")
|
173 |
+
if call_agent:
|
174 |
+
tools.append(self.tooluniverse.get_one_tool_by_one_name(
|
175 |
+
'CallAgent', return_prompt=True))
|
176 |
+
print("CallAgent tool is added")
|
177 |
+
else:
|
178 |
+
if self.enable_rag:
|
179 |
+
tools.append(self.tooluniverse.get_one_tool_by_one_name(
|
180 |
+
'Tool_RAG', return_prompt=True))
|
181 |
+
print("Tool_RAG tool is added")
|
182 |
+
|
183 |
+
if self.additional_default_tools is not None:
|
184 |
+
for each_tool_name in self.additional_default_tools:
|
185 |
+
tool_prompt = self.tooluniverse.get_one_tool_by_one_name(
|
186 |
+
each_tool_name, return_prompt=True)
|
187 |
+
if tool_prompt is not None:
|
188 |
+
print(f"{each_tool_name} tool is added")
|
189 |
+
tools.append(tool_prompt)
|
190 |
return tools
|
191 |
|
192 |
def add_finish_tools(self, tools):
|
193 |
+
tools.append(self.tooluniverse.get_one_tool_by_one_name(
|
194 |
+
'Finish', return_prompt=True))
|
195 |
+
print("Finish tool is added")
|
|
|
196 |
return tools
|
197 |
|
198 |
def set_system_prompt(self, conversation, sys_prompt):
|
199 |
if len(conversation) == 0:
|
200 |
+
conversation.append(
|
201 |
+
{"role": "system", "content": sys_prompt})
|
202 |
else:
|
203 |
conversation[0] = {"role": "system", "content": sys_prompt}
|
204 |
return conversation
|
|
|
210 |
call_agent=False,
|
211 |
call_agent_level=None,
|
212 |
temperature=None):
|
213 |
+
|
214 |
+
function_call_json, message = self.tooluniverse.extract_function_call_json(
|
215 |
+
fcall_str, return_message=return_message, verbose=False)
|
216 |
call_results = []
|
217 |
special_tool_call = ''
|
218 |
if function_call_json is not None:
|
219 |
if isinstance(function_call_json, list):
|
220 |
for i in range(len(function_call_json)):
|
221 |
+
print("\033[94mTool Call:\033[0m", function_call_json[i])
|
222 |
if function_call_json[i]["name"] == 'Finish':
|
223 |
special_tool_call = 'Finish'
|
224 |
break
|
|
|
246 |
else:
|
247 |
call_result = call_result.split('[FinalAnswer]')[-1].strip()
|
248 |
else:
|
249 |
+
call_result = "Error: The CallAgent has been disabled. Please proceed with your reasoning process to solve this question."
|
250 |
else:
|
251 |
+
call_result = self.tooluniverse.run_one_function(
|
252 |
+
function_call_json[i])
|
253 |
+
|
254 |
call_id = self.tooluniverse.call_id_gen()
|
255 |
function_call_json[i]["call_id"] = call_id
|
256 |
+
print("\033[94mTool Call Result:\033[0m", call_result)
|
257 |
call_results.append({
|
258 |
"role": "tool",
|
259 |
"content": json.dumps({"tool_name": function_call_json[i]["name"], "content": call_result, "call_id": call_id})
|
|
|
261 |
else:
|
262 |
call_results.append({
|
263 |
"role": "tool",
|
264 |
+
"content": json.dumps({"content": "Not a valid function call, please check the function call format."})
|
265 |
})
|
266 |
+
|
267 |
revised_messages = [{
|
268 |
"role": "assistant",
|
269 |
"content": message.strip(),
|
270 |
"tool_calls": json.dumps(function_call_json)
|
271 |
}] + call_results
|
272 |
+
|
273 |
+
# Yield the final result.
|
274 |
return revised_messages, existing_tools_prompt, special_tool_call
|
275 |
|
276 |
def run_function_call_stream(self, fcall_str,
|
277 |
+
return_message=False,
|
278 |
+
existing_tools_prompt=None,
|
279 |
+
message_for_call_agent=None,
|
280 |
+
call_agent=False,
|
281 |
+
call_agent_level=None,
|
282 |
+
temperature=None,
|
283 |
+
return_gradio_history=True):
|
284 |
+
|
285 |
+
function_call_json, message = self.tooluniverse.extract_function_call_json(
|
286 |
+
fcall_str, return_message=return_message, verbose=False)
|
|
|
287 |
call_results = []
|
288 |
special_tool_call = ''
|
289 |
+
if return_gradio_history:
|
290 |
+
gradio_history = []
|
291 |
if function_call_json is not None:
|
292 |
if isinstance(function_call_json, list):
|
293 |
for i in range(len(function_call_json)):
|
|
|
315 |
"\nYou must follow the following plan to answer the question: " +
|
316 |
str(solution_plan)
|
317 |
)
|
318 |
+
sub_agent_task = "Sub TxAgent plan: " + \
|
319 |
+
str(solution_plan)
|
320 |
call_result = yield from self.run_gradio_chat(
|
321 |
full_message, history=[], temperature=temperature,
|
322 |
max_new_tokens=1024, max_token=99999,
|
323 |
call_agent=False, call_agent_level=call_agent_level,
|
324 |
conversation=None,
|
325 |
sub_agent_task=sub_agent_task)
|
326 |
+
|
327 |
if call_result is not None and isinstance(call_result, str):
|
328 |
call_result = call_result.split('[FinalAnswer]')[-1]
|
329 |
else:
|
330 |
call_result = "⚠️ No content returned from sub-agent."
|
331 |
else:
|
332 |
+
call_result = "Error: The CallAgent has been disabled. Please proceed with your reasoning process to solve this question."
|
333 |
else:
|
334 |
+
call_result = self.tooluniverse.run_one_function(
|
335 |
+
function_call_json[i])
|
336 |
+
|
337 |
call_id = self.tooluniverse.call_id_gen()
|
338 |
function_call_json[i]["call_id"] = call_id
|
339 |
call_results.append({
|
|
|
350 |
else:
|
351 |
call_results.append({
|
352 |
"role": "tool",
|
353 |
+
"content": json.dumps({"content": "Not a valid function call, please check the function call format."})
|
354 |
})
|
355 |
+
|
356 |
revised_messages = [{
|
357 |
"role": "assistant",
|
358 |
"content": message.strip(),
|
359 |
"tool_calls": json.dumps(function_call_json)
|
360 |
}] + call_results
|
361 |
+
|
362 |
+
if return_gradio_history:
|
363 |
+
return revised_messages, existing_tools_prompt, special_tool_call, gradio_history
|
364 |
+
else:
|
365 |
+
return revised_messages, existing_tools_prompt, special_tool_call
|
366 |
+
|
367 |
|
368 |
def get_answer_based_on_unfinished_reasoning(self, conversation, temperature, max_new_tokens, max_token, outputs=None):
|
369 |
+
if conversation[-1]['role'] == 'assisant':
|
370 |
+
conversation.append(
|
371 |
+
{'role': 'tool', 'content': 'Errors happen during the function call, please come up with the final answer with the current information.'})
|
372 |
+
finish_tools_prompt = self.add_finish_tools([])
|
373 |
+
|
374 |
last_outputs_str = self.llm_infer(messages=conversation,
|
375 |
+
temperature=temperature,
|
376 |
+
tools=finish_tools_prompt,
|
377 |
+
output_begin_string='Since I cannot continue reasoning, I will provide the final answer based on the current information and general knowledge.\n\n[FinalAnswer]',
|
378 |
+
skip_special_tokens=True,
|
379 |
+
max_new_tokens=max_new_tokens, max_token=max_token)
|
380 |
+
print(last_outputs_str)
|
381 |
return last_outputs_str
|
382 |
|
383 |
def run_multistep_agent(self, message: str,
|
|
|
387 |
max_round: int = 20,
|
388 |
call_agent=False,
|
389 |
call_agent_level=0) -> str:
|
390 |
+
"""
|
391 |
+
Generate a streaming response using the llama3-8b model.
|
392 |
+
Args:
|
393 |
+
message (str): The input message.
|
394 |
+
temperature (float): The temperature for generating the response.
|
395 |
+
max_new_tokens (int): The maximum number of new tokens to generate.
|
396 |
+
Returns:
|
397 |
+
str: The generated response.
|
398 |
+
"""
|
399 |
+
print("\033[1;32;40mstart\033[0m")
|
400 |
picked_tools_prompt, call_agent_level = self.initialize_tools_prompt(
|
401 |
call_agent, call_agent_level, message)
|
402 |
conversation = self.initialize_conversation(message)
|
403 |
+
|
404 |
outputs = []
|
405 |
last_outputs = []
|
406 |
next_round = True
|
|
|
409 |
token_overflow = False
|
410 |
enable_summary = False
|
411 |
last_status = {}
|
412 |
+
|
413 |
if self.enable_checker:
|
414 |
checker = ReasoningTraceChecker(message, conversation)
|
415 |
try:
|
|
|
423 |
call_agent=call_agent,
|
424 |
call_agent_level=call_agent_level,
|
425 |
temperature=temperature)
|
426 |
+
|
427 |
if special_tool_call == 'Finish':
|
428 |
next_round = False
|
429 |
conversation.extend(function_call_messages)
|
430 |
+
if isinstance(function_call_messages[0]['content'], types.GeneratorType):
|
431 |
+
function_call_messages[0]['content'] = next(
|
432 |
+
function_call_messages[0]['content'])
|
433 |
content = function_call_messages[0]['content']
|
434 |
if content is None:
|
435 |
+
return "❌ No content returned after Finish tool call."
|
436 |
return content.split('[FinalAnswer]')[-1]
|
437 |
+
|
438 |
if (self.enable_summary or token_overflow) and not call_agent:
|
439 |
+
if token_overflow:
|
440 |
+
print("token_overflow, using summary")
|
441 |
enable_summary = True
|
442 |
last_status = self.function_result_summary(
|
443 |
conversation, status=last_status, enable_summary=enable_summary)
|
444 |
+
|
445 |
+
if function_call_messages is not None:
|
446 |
conversation.extend(function_call_messages)
|
447 |
+
outputs.append(tool_result_format(
|
448 |
+
function_call_messages))
|
449 |
else:
|
450 |
next_round = False
|
451 |
+
conversation.extend(
|
452 |
+
[{"role": "assistant", "content": ''.join(last_outputs)}])
|
453 |
return ''.join(last_outputs).replace("</s>", "")
|
454 |
if self.enable_checker:
|
455 |
good_status, wrong_info = checker.check_conversation()
|
456 |
if not good_status:
|
457 |
+
next_round = False
|
458 |
+
print(
|
459 |
+
"Internal error in reasoning: " + wrong_info)
|
460 |
break
|
461 |
last_outputs = []
|
462 |
outputs.append("### TxAgent:\n")
|
463 |
+
last_outputs_str, token_overflow = self.llm_infer(messages=conversation,
|
464 |
+
temperature=temperature,
|
465 |
+
tools=picked_tools_prompt,
|
466 |
+
skip_special_tokens=False,
|
467 |
+
max_new_tokens=max_new_tokens, max_token=max_token,
|
468 |
+
check_token_status=True)
|
|
|
|
|
469 |
if last_outputs_str is None:
|
470 |
+
print("The number of tokens exceeds the maximum limit.")
|
471 |
if self.force_finish:
|
472 |
+
return self.get_answer_based_on_unfinished_reasoning(conversation, temperature, max_new_tokens, max_token)
|
473 |
+
else:
|
474 |
+
return "❌ Token limit exceeded — no further steps possible."
|
475 |
+
else:
|
476 |
+
last_outputs.append(last_outputs_str)
|
477 |
if max_round == current_round:
|
478 |
+
print("The number of rounds exceeds the maximum limit!")
|
479 |
if self.force_finish:
|
480 |
+
return self.get_answer_based_on_unfinished_reasoning(conversation, temperature, max_new_tokens, max_token)
|
481 |
+
else:
|
482 |
+
return None
|
483 |
+
|
484 |
except Exception as e:
|
485 |
+
print(f"Error: {e}")
|
486 |
if self.force_finish:
|
487 |
+
return self.get_answer_based_on_unfinished_reasoning(conversation, temperature, max_new_tokens, max_token)
|
488 |
+
else:
|
489 |
+
return None
|
490 |
|
491 |
def build_logits_processor(self, messages, llm):
|
492 |
+
# Use the tokenizer from the LLM instance.
|
493 |
tokenizer = llm.get_tokenizer()
|
494 |
if self.avoid_repeat and len(messages) > 2:
|
495 |
assistant_messages = []
|
|
|
498 |
assistant_messages.append(messages[-i]['content'])
|
499 |
if len(assistant_messages) == 2:
|
500 |
break
|
501 |
+
forbidden_ids = [tokenizer.encode(
|
502 |
+
msg, add_special_tokens=False) for msg in assistant_messages]
|
503 |
+
return [NoRepeatSentenceProcessor(forbidden_ids, 5)]
|
504 |
+
else:
|
505 |
+
return None
|
506 |
|
507 |
def llm_infer(self, messages, temperature=0.1, tools=None,
|
508 |
output_begin_string=None, max_new_tokens=2048,
|
509 |
max_token=None, skip_special_tokens=True,
|
510 |
+
model=None, tokenizer=None, terminators=None, seed=None, check_token_status=False):
|
511 |
+
|
512 |
if model is None:
|
513 |
model = self.model
|
514 |
+
|
515 |
logits_processor = self.build_logits_processor(messages, model)
|
516 |
sampling_params = SamplingParams(
|
517 |
temperature=temperature,
|
518 |
max_tokens=max_new_tokens,
|
519 |
+
|
520 |
seed=seed if seed is not None else self.seed,
|
521 |
)
|
522 |
+
|
523 |
prompt = self.chat_template.render(
|
524 |
messages=messages, tools=tools, add_generation_prompt=True)
|
525 |
if output_begin_string is not None:
|
526 |
prompt += output_begin_string
|
527 |
+
|
528 |
if check_token_status and max_token is not None:
|
529 |
token_overflow = False
|
530 |
+
num_input_tokens = len(self.tokenizer.encode(
|
531 |
+
prompt, return_tensors="pt")[0])
|
532 |
+
if max_token is not None:
|
533 |
+
if num_input_tokens > max_token:
|
534 |
+
torch.cuda.empty_cache()
|
535 |
+
gc.collect()
|
536 |
+
print("Number of input tokens before inference:",
|
537 |
+
num_input_tokens)
|
538 |
+
logger.info(
|
539 |
+
"The number of tokens exceeds the maximum limit!!!!")
|
540 |
+
token_overflow = True
|
541 |
+
return None, token_overflow
|
542 |
+
output = model.generate(
|
543 |
+
prompt,
|
544 |
+
sampling_params=sampling_params,
|
545 |
+
)
|
546 |
output = output[0].outputs[0].text
|
547 |
+
print("\033[92m" + output + "\033[0m")
|
548 |
if check_token_status and max_token is not None:
|
549 |
return output, token_overflow
|
550 |
+
|
551 |
return output
|
552 |
|
553 |
def run_self_agent(self, message: str,
|
554 |
+
temperature: float,
|
555 |
+
max_new_tokens: int,
|
556 |
+
max_token: int) -> str:
|
557 |
+
|
558 |
+
print("\033[1;32;40mstart self agent\033[0m")
|
559 |
+
conversation = []
|
560 |
+
conversation = self.set_system_prompt(conversation, self.self_prompt)
|
561 |
conversation.append({"role": "user", "content": message})
|
562 |
return self.llm_infer(messages=conversation,
|
563 |
temperature=temperature,
|
|
|
568 |
temperature: float,
|
569 |
max_new_tokens: int,
|
570 |
max_token: int) -> str:
|
571 |
+
|
572 |
+
print("\033[1;32;40mstart chat agent\033[0m")
|
573 |
+
conversation = []
|
574 |
+
conversation = self.set_system_prompt(conversation, self.chat_prompt)
|
575 |
conversation.append({"role": "user", "content": message})
|
576 |
return self.llm_infer(messages=conversation,
|
577 |
temperature=temperature,
|
|
|
579 |
max_new_tokens=max_new_tokens, max_token=max_token)
|
580 |
|
581 |
def run_format_agent(self, message: str,
|
582 |
+
answer: str,
|
583 |
+
temperature: float,
|
584 |
+
max_new_tokens: int,
|
585 |
+
max_token: int) -> str:
|
586 |
+
|
587 |
+
print("\033[1;32;40mstart format agent\033[0m")
|
588 |
if '[FinalAnswer]' in answer:
|
589 |
possible_final_answer = answer.split("[FinalAnswer]")[-1]
|
590 |
elif "\n\n" in answer:
|
591 |
possible_final_answer = answer.split("\n\n")[-1]
|
592 |
else:
|
593 |
possible_final_answer = answer.strip()
|
594 |
+
if len(possible_final_answer) == 1:
|
595 |
+
choice = possible_final_answer[0]
|
596 |
+
if choice in ['A', 'B', 'C', 'D', 'E']:
|
597 |
+
return choice
|
598 |
+
elif len(possible_final_answer) > 1:
|
599 |
+
if possible_final_answer[1] == ':':
|
600 |
+
choice = possible_final_answer[0]
|
601 |
+
if choice in ['A', 'B', 'C', 'D', 'E']:
|
602 |
+
print("choice", choice)
|
603 |
+
return choice
|
604 |
+
|
605 |
conversation = []
|
606 |
+
format_prompt = f"You are helpful assistant to transform the answer of agent to the final answer of 'A', 'B', 'C', 'D'."
|
607 |
conversation = self.set_system_prompt(conversation, format_prompt)
|
608 |
+
conversation.append({"role": "user", "content": message +
|
609 |
+
"\nThe final answer of agent:" + answer + "\n The answer is (must be a letter):"})
|
610 |
return self.llm_infer(messages=conversation,
|
611 |
temperature=temperature,
|
612 |
tools=None,
|
613 |
max_new_tokens=max_new_tokens, max_token=max_token)
|
614 |
|
615 |
def run_summary_agent(self, thought_calls: str,
|
616 |
+
function_response: str,
|
617 |
+
temperature: float,
|
618 |
+
max_new_tokens: int,
|
619 |
+
max_token: int) -> str:
|
620 |
+
print("\033[1;32;40mSummarized Tool Result:\033[0m")
|
621 |
generate_tool_result_summary_training_prompt = """Thought and function calls:
|
622 |
{thought_calls}
|
623 |
Function calls' responses:
|
624 |
\"\"\"
|
625 |
{function_response}
|
626 |
\"\"\"
|
627 |
+
Based on the Thought and function calls, and the function calls' responses, you need to generate a summary of the function calls' responses that fulfills the requirements of the thought. The summary MUST BE ONE sentence and include all necessary information.
|
628 |
+
Directly respond with the summarized sentence of the function calls' responses only.
|
629 |
+
Generate **one summarized sentence** about "function calls' responses" with necessary information, and respond with a string:
|
630 |
+
""".format(thought_calls=thought_calls, function_response=function_response)
|
631 |
+
conversation = []
|
632 |
+
conversation.append(
|
633 |
+
{"role": "user", "content": generate_tool_result_summary_training_prompt})
|
634 |
output = self.llm_infer(messages=conversation,
|
635 |
temperature=temperature,
|
636 |
tools=None,
|
637 |
max_new_tokens=max_new_tokens, max_token=max_token)
|
638 |
+
|
639 |
if '[' in output:
|
640 |
output = output.split('[')[0]
|
641 |
return output
|
642 |
|
643 |
def function_result_summary(self, input_list, status, enable_summary):
|
644 |
+
"""
|
645 |
+
Processes the input list, extracting information from sequences of 'user', 'tool', 'assistant' roles.
|
646 |
+
Supports 'length' and 'step' modes, and skips the last 'k' groups.
|
647 |
+
Parameters:
|
648 |
+
input_list (list): A list of dictionaries containing role and other information.
|
649 |
+
summary_skip_last_k (int): Number of groups to skip from the end. Defaults to 0.
|
650 |
+
summary_context_length (int): The context length threshold for the 'length' mode.
|
651 |
+
last_processed_index (tuple or int): The last processed index.
|
652 |
+
Returns:
|
653 |
+
list: A list of extracted information from valid sequences.
|
654 |
+
"""
|
655 |
if 'tool_call_step' not in status:
|
656 |
status['tool_call_step'] = 0
|
657 |
+
|
658 |
for idx in range(len(input_list)):
|
659 |
pos_id = len(input_list)-idx-1
|
660 |
+
if input_list[pos_id]['role'] == 'assistant':
|
661 |
+
if 'tool_calls' in input_list[pos_id]:
|
662 |
+
if 'Tool_RAG' in str(input_list[pos_id]['tool_calls']):
|
663 |
+
status['tool_call_step'] += 1
|
664 |
break
|
665 |
+
|
666 |
+
if 'step' in status:
|
667 |
+
status['step'] += 1
|
668 |
+
else:
|
669 |
+
status['step'] = 0
|
670 |
+
|
671 |
if not enable_summary:
|
672 |
return status
|
673 |
+
|
674 |
if 'summarized_index' not in status:
|
675 |
status['summarized_index'] = 0
|
676 |
+
|
677 |
if 'summarized_step' not in status:
|
678 |
status['summarized_step'] = 0
|
679 |
+
|
680 |
if 'previous_length' not in status:
|
681 |
status['previous_length'] = 0
|
682 |
+
|
683 |
if 'history' not in status:
|
684 |
status['history'] = []
|
685 |
+
|
686 |
function_response = ''
|
687 |
+
idx = 0
|
688 |
current_summarized_index = status['summarized_index']
|
689 |
+
|
690 |
status['history'].append(self.summary_mode == 'step' and status['summarized_step']
|
691 |
< status['step']-status['tool_call_step']-self.summary_skip_last_k)
|
692 |
+
|
693 |
+
idx = current_summarized_index
|
694 |
while idx < len(input_list):
|
695 |
if (self.summary_mode == 'step' and status['summarized_step'] < status['step']-status['tool_call_step']-self.summary_skip_last_k) or (self.summary_mode == 'length' and status['previous_length'] > self.summary_context_length):
|
696 |
+
|
697 |
if input_list[idx]['role'] == 'assistant':
|
698 |
if 'Tool_RAG' in str(input_list[idx]['tool_calls']):
|
699 |
this_thought_calls = None
|
700 |
else:
|
701 |
if len(function_response) != 0:
|
702 |
+
print("internal summary")
|
703 |
status['summarized_step'] += 1
|
704 |
result_summary = self.run_summary_agent(
|
705 |
thought_calls=this_thought_calls,
|
706 |
function_response=function_response,
|
707 |
temperature=0.1,
|
708 |
max_new_tokens=1024,
|
709 |
+
max_token=99999
|
710 |
+
)
|
711 |
+
|
712 |
+
input_list.insert(
|
713 |
+
last_call_idx+1, {'role': 'tool', 'content': result_summary})
|
714 |
status['summarized_index'] = last_call_idx + 2
|
715 |
idx += 1
|
716 |
+
|
717 |
last_call_idx = idx
|
718 |
+
this_thought_calls = input_list[idx]['content'] + \
|
719 |
+
input_list[idx]['tool_calls']
|
720 |
function_response = ''
|
721 |
+
|
722 |
elif input_list[idx]['role'] == 'tool' and this_thought_calls is not None:
|
723 |
function_response += input_list[idx]['content']
|
724 |
del input_list[idx]
|
725 |
idx -= 1
|
726 |
+
|
727 |
else:
|
728 |
break
|
729 |
idx += 1
|
730 |
+
|
731 |
if len(function_response) != 0:
|
732 |
status['summarized_step'] += 1
|
733 |
result_summary = self.run_summary_agent(
|
|
|
735 |
function_response=function_response,
|
736 |
temperature=0.1,
|
737 |
max_new_tokens=1024,
|
738 |
+
max_token=99999
|
739 |
+
)
|
740 |
+
|
741 |
tool_calls = json.loads(input_list[last_call_idx]['tool_calls'])
|
742 |
for tool_call in tool_calls:
|
743 |
del tool_call['call_id']
|
744 |
input_list[last_call_idx]['tool_calls'] = json.dumps(tool_calls)
|
745 |
+
input_list.insert(
|
746 |
+
last_call_idx+1, {'role': 'tool', 'content': result_summary})
|
747 |
status['summarized_index'] = last_call_idx + 2
|
748 |
+
|
749 |
return status
|
750 |
|
751 |
+
# Following are Gradio related functions
|
752 |
+
|
753 |
+
# General update method that accepts any new arguments through kwargs
|
754 |
def update_parameters(self, **kwargs):
|
755 |
for key, value in kwargs.items():
|
756 |
if hasattr(self, key):
|
757 |
setattr(self, key, value)
|
758 |
+
|
759 |
+
# Return the updated attributes
|
760 |
+
updated_attributes = {key: value for key,
|
761 |
+
value in kwargs.items() if hasattr(self, key)}
|
762 |
return updated_attributes
|
763 |
|
764 |
def run_gradio_chat(self, message: str,
|
|
|
772 |
seed: int = None,
|
773 |
call_agent_level: int = 0,
|
774 |
sub_agent_task: str = None,
|
775 |
+
uploaded_files: list = None) -> str:
|
776 |
+
"""
|
777 |
+
Generate a streaming response using the loaded model.
|
778 |
+
Args:
|
779 |
+
message (str): The input message (with file content if uploaded).
|
780 |
+
history (list): The conversation history used by ChatInterface.
|
781 |
+
temperature (float): Sampling temperature.
|
782 |
+
max_new_tokens (int): Max new tokens.
|
783 |
+
max_token (int): Max total tokens allowed.
|
784 |
+
Returns:
|
785 |
+
str: Final assistant message.
|
786 |
+
"""
|
787 |
logger.debug(f"[TxAgent] Chat started, message: {message[:100]}...")
|
788 |
print("\033[1;32;40m[TxAgent] Chat started\033[0m")
|
789 |
+
|
790 |
if not message or len(message.strip()) < 5:
|
791 |
yield "Please provide a valid message or upload files to analyze."
|
792 |
return "Invalid input."
|
793 |
+
|
794 |
if message.startswith("[\U0001f9f0 Tool_RAG") or message.startswith("⚒️"):
|
795 |
return ""
|
796 |
+
|
797 |
+
outputs = []
|
798 |
+
outputs_str = ''
|
799 |
+
last_outputs = []
|
800 |
+
|
801 |
+
picked_tools_prompt, call_agent_level = self.initialize_tools_prompt(
|
802 |
+
call_agent,
|
803 |
+
call_agent_level,
|
804 |
+
message)
|
805 |
+
|
806 |
+
conversation = self.initialize_conversation(
|
807 |
+
message,
|
808 |
+
conversation=conversation,
|
809 |
+
history=history)
|
810 |
+
history = []
|
811 |
+
|
812 |
+
next_round = True
|
813 |
+
function_call_messages = []
|
814 |
+
current_round = 0
|
815 |
+
enable_summary = False
|
816 |
+
last_status = {}
|
817 |
+
token_overflow = False
|
818 |
+
|
819 |
+
if self.enable_checker:
|
820 |
+
checker = ReasoningTraceChecker(
|
821 |
+
message, conversation, init_index=len(conversation))
|
822 |
+
|
823 |
+
try:
|
824 |
+
while next_round and current_round < max_round:
|
825 |
+
current_round += 1
|
826 |
+
logger.debug(f"Round {current_round}, conversation length: {len(conversation)}")
|
827 |
+
|
828 |
+
if last_outputs:
|
829 |
+
function_call_messages, picked_tools_prompt, special_tool_call, current_gradio_history = yield from self.run_function_call_stream(
|
830 |
+
last_outputs, return_message=True,
|
831 |
+
existing_tools_prompt=picked_tools_prompt,
|
832 |
+
message_for_call_agent=message,
|
833 |
+
call_agent=call_agent,
|
834 |
+
call_agent_level=call_agent_level,
|
835 |
+
temperature=temperature)
|
836 |
+
|
837 |
+
history.extend(current_gradio_history)
|
838 |
+
|
839 |
+
if special_tool_call == 'Finish' and function_call_messages:
|
840 |
+
yield history
|
841 |
+
next_round = False
|
842 |
+
conversation.extend(function_call_messages)
|
843 |
+
return function_call_messages[0]['content']
|
844 |
+
|
845 |
+
elif special_tool_call in ['RequireClarification', 'DirectResponse']:
|
846 |
+
last_msg = history[-1] if history else ChatMessage(role="assistant", content="Response needed.")
|
847 |
+
history.append(ChatMessage(role="assistant", content=last_msg.content))
|
848 |
+
yield history
|
849 |
+
next_round = False
|
850 |
+
return last_msg.content
|
851 |
+
|
852 |
+
if (self.enable_summary or token_overflow) and not call_agent:
|
853 |
+
enable_summary = True
|
854 |
+
|
855 |
+
last_status = self.function_result_summary(
|
856 |
+
conversation, status=last_status,
|
857 |
+
enable_summary=enable_summary)
|
858 |
+
|
859 |
+
if function_call_messages:
|
860 |
+
conversation.extend(function_call_messages)
|
861 |
+
yield history
|
862 |
+
else:
|
863 |
+
next_round = False
|
864 |
+
conversation.append({"role": "assistant", "content": ''.join(last_outputs)})
|
865 |
+
return ''.join(last_outputs).replace("</s>", "")
|
866 |
+
|
867 |
+
if self.enable_checker:
|
868 |
+
good_status, wrong_info = checker.check_conversation()
|
869 |
+
if not good_status:
|
870 |
+
print("Checker flagged reasoning error: ", wrong_info)
|
871 |
+
break
|
872 |
+
|
873 |
+
last_outputs = []
|
874 |
+
last_outputs_str, token_overflow = self.llm_infer(
|
875 |
+
messages=conversation,
|
876 |
+
temperature=temperature,
|
877 |
+
tools=picked_tools_prompt,
|
878 |
+
skip_special_tokens=False,
|
879 |
+
max_new_tokens=max_new_tokens,
|
880 |
+
max_token=max_token,
|
881 |
+
seed=seed,
|
882 |
+
check_token_status=True)
|
883 |
+
|
884 |
+
logger.debug(f"llm_infer output: {last_outputs_str[:100] if last_outputs_str else None}, token_overflow: {token_overflow}")
|
885 |
+
|
886 |
+
if last_outputs_str is None:
|
887 |
+
logger.warning("llm_infer returned None due to token overflow")
|
888 |
+
if self.force_finish:
|
889 |
+
last_outputs_str = self.get_answer_based_on_unfinished_reasoning(
|
890 |
+
conversation, temperature, max_new_tokens, max_token)
|
891 |
+
history.append(ChatMessage(role="assistant", content=last_outputs_str.strip()))
|
892 |
+
yield history
|
893 |
+
return last_outputs_str
|
894 |
+
else:
|
895 |
+
error_msg = "Token limit exceeded. Please reduce input size or increase max_token."
|
896 |
+
history.append(ChatMessage(role="assistant", content=error_msg))
|
897 |
+
yield history
|
898 |
+
return error_msg
|
899 |
+
|
900 |
+
last_thought = last_outputs_str.split("[TOOL_CALLS]")[0]
|
901 |
+
|
902 |
+
for msg in history:
|
903 |
+
if msg.metadata is not None:
|
904 |
+
msg.metadata['status'] = 'done'
|
905 |
+
|
906 |
+
if '[FinalAnswer]' in last_thought:
|
907 |
+
parts = last_thought.split('[FinalAnswer]', 1)
|
908 |
+
if len(parts) == 2:
|
909 |
+
final_thought, final_answer = parts
|
910 |
+
else:
|
911 |
+
final_thought, final_answer = last_thought, ""
|
912 |
+
history.append(ChatMessage(role="assistant", content=final_thought.strip()))
|
913 |
+
yield history
|
914 |
+
history.append(ChatMessage(role="assistant", content="**🧠 Final Analysis:**\n" + final_answer.strip()))
|
915 |
+
yield history
|
916 |
+
else:
|
917 |
+
history.append(ChatMessage(role="assistant", content=last_thought))
|
918 |
+
yield history
|
919 |
+
|
920 |
+
last_outputs.append(last_outputs_str)
|
921 |
+
|
922 |
+
if next_round:
|
923 |
+
if self.force_finish:
|
924 |
+
last_outputs_str = self.get_answer_based_on_unfinished_reasoning(
|
925 |
+
conversation, temperature, max_new_tokens, max_token)
|
926 |
+
if '[FinalAnswer]' in last_outputs_str:
|
927 |
+
parts = last_outputs_str.split('[FinalAnswer]', 1)
|
928 |
+
if len(parts) == 2:
|
929 |
+
final_thought, final_answer = parts
|
930 |
+
else:
|
931 |
+
final_thought, final_answer = last_outputs_str, ""
|
932 |
+
history.append(ChatMessage(role="assistant", content=final_thought.strip()))
|
933 |
+
yield history
|
934 |
+
history.append(ChatMessage(role="assistant", content="**🧠 Final Analysis:**\n" + final_answer.strip()))
|
935 |
+
yield history
|
936 |
+
else:
|
937 |
+
history.append(ChatMessage(role="assistant", content=last_outputs_str.strip()))
|
938 |
+
yield history
|
939 |
+
else:
|
940 |
+
yield "The number of reasoning rounds exceeded the limit."
|
941 |
+
|
942 |
+
except Exception as e:
|
943 |
+
logger.error(f"Exception in run_gradio_chat: {e}", exc_info=True)
|
944 |
+
error_msg = f"An error occurred: {e}"
|
945 |
+
history.append(ChatMessage(role="assistant", content=error_msg))
|
946 |
+
yield history
|
947 |
+
if self.force_finish:
|
948 |
+
last_outputs_str = self.get_answer_based_on_unfinished_reasoning(
|
949 |
+
conversation, temperature, max_new_tokens, max_token)
|
950 |
+
if '[FinalAnswer]' in last_outputs_str:
|
951 |
+
parts = last_outputs_str.split('[FinalAnswer]', 1)
|
952 |
+
if len(parts) == 2:
|
953 |
+
final_thought, final_answer = parts
|
954 |
+
else:
|
955 |
+
final_thought, final_answer = last_outputs_str, ""
|
956 |
+
history.append(ChatMessage(role="assistant", content=final_thought.strip()))
|
957 |
+
yield history
|
958 |
+
history.append(ChatMessage(role="assistant", content="**🧠 Final Analysis:**\n" + final_answer.strip()))
|
959 |
+
yield history
|
960 |
+
else:
|
961 |
+
history.append(ChatMessage(role="assistant", content=last_outputs_str.strip()))
|
962 |
+
yield history
|
963 |
+
return error_msg
|