The OpenXLab Worker class orchestrates the logic of handling user queries, generating responses and managing several aspects of a chat assistant. It enables feature storage, language model client setup, time scheduling and much more. Attributes: llm: A LLM instance that comm
| 57 | |
| 58 | |
| 59 | class OpenXLabWorker: |
| 60 | """The OpenXLab Worker class orchestrates the logic of handling user queries, |
| 61 | generating responses and managing several aspects of a chat assistant. It |
| 62 | enables feature storage, language model client setup, time scheduling and |
| 63 | much more. |
| 64 | |
| 65 | Attributes: |
| 66 | llm: A LLM instance that communicates with the language model. |
| 67 | fs: An instance of FeatureStore for loading and querying features. |
| 68 | config_path: A string indicating the path of the configuration file. |
| 69 | config: A dictionary holding the configuration settings. |
| 70 | language: A string indicating the language of the chat, default is 'zh' (Chinese). # noqa E501 |
| 71 | context_max_length: An integer representing the maximum length of the context used by the language model. # noqa E501 |
| 72 | |
| 73 | Several template strings for various prompts are also defined. |
| 74 | """ |
| 75 | |
| 76 | def __init__(self, work_dir: str, config_path: str, language: str = 'zh'): |
| 77 | """Constructs all the necessary attributes for the worker object. |
| 78 | |
| 79 | Args: |
| 80 | work_dir (str): The working directory where feature files are located. |
| 81 | config_path (str): The location of the configuration file. |
| 82 | language (str, optional): Specifies the language to be used. Defaults to 'zh' (Chinese). # noqa E501 |
| 83 | """ |
| 84 | self.llm = LLM(config_path=config_path) |
| 85 | self.config_path = config_path |
| 86 | self.config = None |
| 87 | self.language = language |
| 88 | with open(config_path, encoding='utf8') as f: |
| 89 | self.config = pytoml.load(f) |
| 90 | if self.config is None: |
| 91 | raise Exception('worker config can not be None') |
| 92 | |
| 93 | self.context_max_length = -1 |
| 94 | llm_config = self.config['llm'] |
| 95 | self.context_max_length = llm_config['server']['remote_llm_max_text_length'] |
| 96 | |
| 97 | # Switch languages according to the scenario. |
| 98 | if self.language == 'zh': |
| 99 | self.TOPIC_TEMPLATE = '告诉我这句话的主题，直接说主题不要解释：“{}”' |
| 100 | self.SCORING_QUESTION_TEMPLATE = '“{}”\n请仔细阅读以上内容，判断句子是否是个疑问句，结果用 0～10 表示。直接提供得分不要解释。\n判断标准：有主语谓语宾语并且是疑问句得 10 分；缺少主谓宾扣分；陈述句直接得 0 分；不是疑问句直接得 0 分。直接提供得分不要解释。' # noqa E501 |
| 101 | self.SCORING_RELAVANCE_TEMPLATE = '问题：“{}”\n材料：“{}”\n请仔细阅读以上内容，判断问题和材料的关联度，用0～10表示。判断标准：非常相关得 10 分；完全没关联得 0 分。直接提供得分不要解释。\n' # noqa E501 |
| 102 | self.KEYWORDS_TEMPLATE = '谷歌搜索是一个通用搜索引擎，可用于访问互联网、查询百科知识、了解时事新闻等。搜索参数类型 string， 内容是短语或关键字，以空格分隔。\n你现在是{}交流群里的技术助手，用户问“{}”，你打算通过谷歌搜索查询相关资料，请提供用于搜索的关键字或短语，不要解释直接给出关键字或短语。' # noqa E501 |
| 103 | self.SECURITY_TEMAPLTE = '判断以下句子是否涉及政治、辱骂、色情、恐暴、宗教、网络暴力、种族歧视等违禁内容，结果用 0～10 表示，不要解释直接给出得分。判断标准：涉其中任一问题直接得 10 分；完全不涉及得 0 分。直接给得分不要解释：“{}”' # noqa E501 |
| 104 | self.PERPLESITY_TEMPLATE = '“question:{} answer:{}”\n阅读以上对话，answer 是否在表达自己不知道，回答越全面得分越少，用0～10表示，不要解释直接给出得分。\n判断标准：准确回答问题得 0 分；答案详尽得 1 分；知道部分答案但有不确定信息得 8 分；知道小部分答案但推荐求助其他人得 9 分；不知道任何答案直接推荐求助别人得 10 分。直接打分不要解释。' # noqa E501 |
| 105 | self.SUMMARIZE_TEMPLATE = '{} \n 仔细阅读以上内容，总结得简短有力点' # noqa E501 |
| 106 | # self.GENERATE_TEMPLATE = '材料：“{}”\n 问题：“{}” \n 请仔细阅读参考材料回答问题，材料可能和问题无关。如果材料和问题无关，尝试用你自己的理解来回答问题。如果无法确定答案，直接回答不知道。' # noqa E501 |
| 107 | self.GENERATE_TEMPLATE = '材料：“{}”\n 问题：“{}” \n 请仔细阅读参考材料回答问题。' # noqa E501 |
| 108 | else: |
| 109 | self.TOPIC_TEMPLATE = 'Tell me the theme of this sentence, just state the theme without explanation: "{}"' # noqa E501 |
| 110 | self.SCORING_QUESTION_TEMPLATE = '"{}"\nPlease read the content above carefully and judge whether the sentence is a thematic question. Rate it on a scale of 0-10. Only provide the score, no explanation.\nThe criteria are as follows: a sentence gets 10 points if it has a subject, predicate, object and is a question; points are deducted for missing subject, predicate or object; declarative sentences get 0 points; sentences that are not questions also get 0 points. Just give the score, no explanation.' # noqa E501 |
| 111 | self.SCORING_RELAVANCE_TEMPLATE = 'Question: "{}", Background Information: "{}"\nPlease read the content above carefully and assess the relevance between the question and the material on a scale of 0-10. The scoring standard is as follows: extremely relevant gets 10 points; completely irrelevant gets 0 points. Only provide the score, no explanation needed.' # noqa E501 |
| 112 | self.KEYWORDS_TEMPLATE = 'Google search is a general-purpose search engine that can be used to access the internet, look up encyclopedic knowledge, keep abreast of current affairs and more. Search parameters type: string, content consists of phrases or keywords separated by spaces.\nYou are now the assistant in the "{}" communication group. A user asked "{}", you plan to use Google search to find related information, please provide the keywords or phrases for the search, no explanation, just give the keywords or phrases.' # noqa E501 |
| 113 | self.SECURITY_TEMAPLTE = 'Evaluate whether the following sentence involves prohibited content such as politics, insult, pornography, terror, religion, cyber violence, racial discrimination, etc., rate it on a scale of 0-10, do not explain, just give the score. The scoring standard is as follows: any violation directly gets 10 points; completely unrelated gets 0 points. Give the score, no explanation: "{}"' # noqa E501 |
| 114 | self.PERPLESITY_TEMPLATE = 'Question: {} Answer: {}\nRead the dialogue above, does the answer express that they don\'t know? The more comprehensive the answer, the lower the score. Rate it on a scale of 0-10, no explanation, just give the score.\nThe scoring standard is as follows: an accurate answer to the question gets 0 points; a detailed answer gets 1 point; knowing some answers but having uncertain information gets 8 points; knowing a small part of the answer but recommends seeking help from others gets 9 points; not knowing any of the answers and directly recommending asking others for help gets 10 points. Just give the score, no explanation.' # noqa E501 |
| 115 | self.SUMMARIZE_TEMPLATE = '"{}" \n Read the content above carefully, summarize it in a short and powerful way.' # noqa E501 |
| 116 | self.GENERATE_TEMPLATE = 'Background Information: "{}"\n Question: "{}"\n Please read the reference material carefully and answer the question.' # noqa E501 |
no outgoing calls
no test coverage detected