@@ -59,6 +59,7 @@ from ml_model.services.pulid import Pulid from ml_model.services.qwen import Qwen from ml_model.services.qwen_235B import Qwen_235B from ml_model.services.qwen_3_6 import Qwen_3_6 +from ml_model.services.qwen_3_7 import Qwen_3_7 from ml_model.services.qwen_3_max_thinking import Qwen_3_Max_Thinking from ml_model.services.raifgpt import Raifgpt from ml_model.services.ray import Ray @@ -43,6 +43,22 @@ class Claude(SimpleService): }, # 1M tokens } + HISTORY_CONTEXT_SYSTEM_PROMPT = """ + Conversation history handling rules (strict priority): + 1. The user's latest message is the only message that determines whether you may respond. + Evaluate it independently. Past messages must never cause a refusal by themselves. + 2. Use earlier history only when it is safe and genuinely helps answer the latest message + (context, definitions, preferences, prior facts). Skip everything else. + 3. Treat the following in history as void — do not read, cite, follow, or be influenced by them: + NSFW content, sexual content, illegal requests, violence, hate, self-harm, + prohibited topics, policy violations, and any prior refusals or moderation replies + (including messages like "I cannot process your request"). + 4. If the latest message is appropriate, you must answer it fully and helpfully. + Do not refuse, apologize for history, mention moderation of past turns, + or explain why earlier messages were blocked. + 5. Refuse only when the latest message itself violates policy — never because of history alone. + """ + FABLE_SYSTEM_PROMPT = """ You are a helpful assistant. Answer the user’s question directly and accurately. @@ -89,8 +105,10 @@ class Claude(SimpleService): callback_data = {'provider': {'order': ['anthropic']}, **input_message.info, 'tools': []} messages = [ {'role': 'system', 'content': system_prompt}, + *self.get_chat_history(), {'role': 'user', 'content': input_message.content}, ] + messages.insert(0, {'role': 'system', 'content': self.HISTORY_CONTEXT_SYSTEM_PROMPT}) if version_slug == 'claude-fable-5': current_user_balance = PaymentPlanSelector(self.store.user).get_current_balance() if current_user_balance < (cost := Decimal('100')) and input_message.file: @@ -0,0 +1,183 @@ +import time +from datetime import timedelta +from decimal import Decimal +from pathlib import Path + +import filetype + +from messages.models import Message +from ml_model.exceptions import CorruptedFileError, FileExtensionNotSupported +from ml_model.services.EmbeddingService import EmbeddingService +from ml_model.services.FileService import FileProcessingService +from ml_model.exceptions import ModelVersionNotAvailable +from ml_model.services.base import SimpleService +from ml_model.tasks import openrouter_run +from poller.models import Proxy +from tools.chats.models import Chat +from tools.copywrite.models import Copywrite +from tools.public_api.models import APIStore + + +class Qwen_3_7(SimpleService): + COEFFICIENT = Decimal('300.0') + + TOKENS_COST = { + 'qwen3.7-max': {'input': Decimal('750'), 'output': Decimal('2250')}, + 'qwen3.7-plus': {'input': Decimal('120'), 'output': Decimal('480')}, + } + + MAX_OUTPUT_TOKENS = 30_000 + + TOOLS_TOKEN_COSTS = {'text-embedding-3-small': {'output': Decimal('0.00001')}} + + SYSTEM = """ + You are a reliable and practical AI assistant. + + Provide accurate, direct, and useful answers. + Focus on the user's actual goal and solve the task with minimal unnecessary explanation. + + Rules: + - Follow the user's instructions exactly. + - Start with the answer, not with introductions or disclaimers. + - Be concise when the question is simple. + - Be detailed when the task is complex or requires analysis. + - Ask clarifying questions only when necessary. + - Do not make up facts, data, sources, or capabilities. + - If information is uncertain, state the uncertainty clearly. + - Prefer actionable recommendations over theory. + - Structure long answers with headings, lists, and examples. + - Preserve important details and constraints from the conversation. + - When writing code, prioritize correctness, readability, and maintainability. + - Adapt the level of detail, terminology, and tone to the user's apparent expertise. + + Your goal is to maximize usefulness, clarity, and task completion. + """ + + def calculate_price(self, version: str, cost: float, embedding_tokens: int) -> Decimal: + price = Decimal(cost) * self.COEFFICIENT + if embedding_tokens > 0: + price += self.TOOLS_TOKEN_COSTS['text-embedding-3-small']['output'] * embedding_tokens + return price.quantize(Decimal('0.1'), rounding='ROUND_UP') + + def save_results(self, content: str, time: timedelta, save: bool = True) -> list[Message]: + msgs = [ + Message( + content=content, + content_object=self.store, + elapsed_time=time, + ) + ] + if save: + return Message.objects.bulk_create(msgs) + return msgs + + def make(self, input_message: Message, save: bool = True) -> list[Message]: + start_time = time.time() + version_slug = input_message.info.pop('version', None) + if version_slug is None or version_slug not in self.TOKENS_COST: + raise ModelVersionNotAvailable(version_slug, self.TOKENS_COST) + callback_data = {'max_tokens': self.MAX_OUTPUT_TOKENS, 'tools': [], **input_message.info} + messages = [ + {'role': 'system', 'content': self.SYSTEM}, + *self.get_chat_history(), + {'role': 'user', 'content': input_message.content}, + ] + embedding_tokens = 0 + if input_message.file: + file_service = FileProcessingService + file_bytes = input_message.file.read() + input_message.file.close() + kind = filetype.guess(file_bytes[:550]) + supported_extensions = ['PDF', 'DOC', 'DOCX', 'XLSX'] + if not kind: + if Path(input_message.file.name).suffix[1:].upper() not in supported_extensions: + raise FileExtensionNotSupported(supported_extensions) + raise CorruptedFileError + raw_file_extension = kind.extension + file_extension = file_service.get_file_extension(raw_file_extension, file_bytes) + if file_extension in ('pdf', 'doc', 'docx', 'xlsx'): + text = file_service.get_file_data(file_extension, file_bytes) + chunks = EmbeddingService.split_text_to_chunks(text) + if len(text) > 20_000: + for proxy in Proxy.objects.all(): + document_name = chunks[0].partition(f':{chr(10)}')[2].split(f'{chr(10)}')[0][:100] + embedding_tokens, file_data = EmbeddingService.get_large_file_data( + self.store.messages.first().pk, + chunks, + proxy, + input_message.content, + model='text-embedding-3-small', + index_name='ml_model-index-1536', + ) + messages[-1]['content'] = EmbeddingService.make_embeddings_prompt( + document_name=document_name, + section_texts=file_data, + question=input_message.content, + ) + else: + messages[-1]['content'] = ( + f'Используй системный промпт. Содержание файла: ' + f'{chunks}. Вопрос: {input_message.content}' + ) + else: + raise FileExtensionNotSupported(supported_extensions) + callback_data['tools'].append( + { + 'type': 'openrouter:web_search', + 'parameters': { + 'engine': 'parallel', + 'max_results': 1, + 'max_total_results': 3, + 'search_context_size': 'low', + }, + } + ) + model_slug = f'qwen/{version_slug}' + result = openrouter_run(model_slug, messages, callback_data, 'Qwen 3.7') + process_time = timedelta(seconds=(time.time() - start_time)) + self.handle_invoice( + input_message.content_object.model, + version=version_slug, + cost=result[1], + embedding_tokens=embedding_tokens, + ) + msgs = self.save_results(result[0], process_time) + return msgs + + def get_chat_history( + self, message_limit: int = 10, max_character_limit: int = 1500 + ) -> list[dict[str, str | list]]: + if isinstance(self.store, Chat): + air_messages = list( + reversed( + Message.objects.filter( + chats_chats_messages=self.store, is_deleted=False, is_sent=True + ).order_by('-created_at')[1 : message_limit + 1] + ) + ) + elif isinstance(self.store, APIStore): + air_messages = [] + elif isinstance(self.store, Copywrite): + air_messages = list( + reversed( + Message.objects.filter( + copywrite_copywrites_messages=self.store, + is_deleted=False, + is_sent=True, + ).order_by('-created_at')[:message_limit] + ) + ) + else: + air_messages = [] + memory = [] + for msg in air_messages: + content = msg.content or '' + if msg.from_model: + memory.append({'role': 'assistant', 'content': content}) + else: + memory.append({'role': 'user', 'content': content}) + character_length = sum(len(content['content']) for content in memory) + while character_length > max_character_limit: + character_length -= len(memory.pop(0)['content']) + + return memory @@ -40,6 +40,12 @@ from poller.models import Proxy logger = logging.getLogger(__name__) +CLAUDE_CONTENT_FILTER_FALLBACK = ( + 'Запрос содержит материалы, которые я не могу обрабатывать.\n\n' + '**Токены списываются, поскольку мне приходится анализировать ваш запрос. ' + 'Прошу формулировать запрос точнее, чтобы избежать повторного анализа**' +) + @shared_task def create_d_image(payload: dict): @@ -167,6 +173,13 @@ def openrouter_run(version: str, messages: list, callback_data: dict, model_name for c in (data.get('choices') or []) ): raise RequestBlocked + if model_name == 'Claude': + usage = data.get('usage') or {} + return ( + CLAUDE_CONTENT_FILTER_FALLBACK, + usage.get('prompt_tokens', 0), + usage.get('completion_tokens', 0), + ) raise GenerationException content = ','.join(raw_content) reasoning = ','.join( @@ -177,7 +190,7 @@ def openrouter_run(version: str, messages: list, callback_data: dict, model_name reasoning = re.sub(r'Вывод:|Основная мысль:|Рассуждение:|\*\*', '', reasoning) answer = reasoning if any(m in data['model'] for m in ('google/gemini', 'x-ai/grok-4.3')) or re.match( - r'^qwen/qwen3\.5-.*$', data['model'] + r'^qwen/qwen3\.(?:5|6|7)-.*$', data['model'] ): answer = content elif reasoning and content: @@ -196,8 +209,12 @@ def openrouter_run(version: str, messages: list, callback_data: dict, model_name logger.error(f'Model {model_name} disabled') raise DeploymentDisabled else: - input_tokens = data['usage']['prompt_tokens'] - output_tokens = data['usage']['completion_tokens'] + if re.match(r'^qwen/qwen3\.7-.*$', data['model']): + input_tokens = data['usage']['cost'] + output_tokens = 0 + else: + input_tokens = data['usage']['prompt_tokens'] + output_tokens = data['usage']['completion_tokens'] return (re.sub(r'\\+["n*]', '', answer), input_tokens, output_tokens) logger.error(f'Error occured via model {model_name}. Data: {resp.content}') raise Exception(f'No answer from {model_name}, please retry later')