{"id":1470,"date":"2024-05-15T01:33:02","date_gmt":"2024-05-15T01:33:02","guid":{"rendered":"https:\/\/www.nicekj.com\/?p=1470"},"modified":"2024-05-15T01:33:02","modified_gmt":"2024-05-15T01:33:02","slug":"ruhezailangchainzhongduidamoxingdeshuchujinxinggeshihua","status":"publish","type":"post","link":"https:\/\/www.nicekj.com\/ruhezailangchainzhongduidamoxingdeshuchujinxinggeshihua.html","title":{"rendered":"\u5982\u4f55\u5728langchain\u4e2d\u5bf9\u5927\u6a21\u578b\u7684\u8f93\u51fa\u8fdb\u884c\u683c\u5f0f\u5316"},"content":{"rendered":"

\u7b80\u4ecb<\/h1>\n

\u6211\u4eec\u77e5\u9053\u5728\u5927\u8bed\u8a00\u6a21\u578b\u4e2d, \u4e0d\u7ba1\u6a21\u578b\u7684\u80fd\u529b\u6709\u591a\u5f3a\u5927\uff0c\u4ed6\u7684\u8f93\u5165\u548c\u8f93\u51fa\u57fa\u672c\u4e0a\u90fd\u662f\u6587\u672c\u683c\u5f0f\u7684\uff0c\u6587\u672c\u683c\u5f0f\u7684\u8f93\u5165\u8f93\u51fa\u867d\u7136\u5bf9\u4eba\u6765\u8bf4\u975e\u5e38\u7684\u53cb\u597d\uff0c\u4f46\u662f\u5982\u679c\u6211\u4eec\u60f3\u8981\u8fdb\u884c\u4e00\u4e9b\u7ed3\u6784\u5316\u5904\u7406\u7684\u8bdd\u8fd8\u662f\u4f1a\u6709\u4e00\u70b9\u70b9\u7684\u4e0d\u65b9\u4fbf\u3002<\/p>\n

\u4e0d\u7528\u62c5\u5fc3\uff0clangchain\u5df2\u7ecf\u4e3a\u6211\u4eec\u60f3\u5230\u4e86\u8fd9\u4e2a\u95ee\u9898\uff0c\u5e76\u4e14\u63d0\u51fa\u4e86\u5b8c\u6ee1\u7684\u89e3\u51b3\u65b9\u6848\u3002<\/p>\n

langchain\u4e2d\u7684output parsers<\/h1>\n

langchain\u4e2d\u6240\u6709\u7684output parsers\u90fd\u662f\u7ee7\u627f\u81eaBaseOutputParser\u3002\u8fd9\u4e2a\u57fa\u7840\u7c7b\u63d0\u4f9b\u4e86\u5bf9LLM\u5927\u6a21\u578b\u8f93\u51fa\u7684\u683c\u5f0f\u5316\u65b9\u6cd5\uff0c\u662f\u4e00\u4e2a\u4f18\u79c0\u7684\u5de5\u5177\u7c7b\u3002<\/p>\n

\u6211\u4eec\u5148\u6765\u770b\u4e0b\u4ed6\u7684\u5b9e\u73b0\uff1a<\/p>\n

<\/div>
python<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>class<\/span> BaseOutputParser<\/span>(BaseModel, ABC, Generic<\/span>[T]):<\/span>\n<\/span>\n    @abstractmethod<\/span><\/span>\n    def<\/span> parse<\/span>(self, text: str<\/span><\/span>) -> T:<\/span>\n        \"\"\"Parse the output of an LLM call.<\/span><\/span>\n<\/span>\n        A method which takes in a string (assumed output of a language model )<\/span>\n        and parses it into some structure.<\/span>\n<\/span>\n        Args:<\/span>\n            text: output of language model<\/span>\n<\/span>\n        Returns:<\/span>\n            structured output<\/span>\n        \"\"\"<\/span>\n<\/span>\n    def<\/span> parse_with_prompt<\/span>(self, completion: str<\/span>, prompt: PromptValue<\/span>) -> Any<\/span>:<\/span>\n        \"\"\"Optional method to parse the output of an LLM call with a prompt.<\/span><\/span>\n<\/span>\n        The prompt is largely provided in the event the OutputParser wants<\/span>\n        to retry or fix the output in some way, and needs information from<\/span>\n        the prompt to do so.<\/span>\n<\/span>\n        Args:<\/span>\n            completion: output of language model<\/span>\n            prompt: prompt value<\/span>\n<\/span>\n        Returns:<\/span>\n            structured output<\/span>\n        \"\"\"<\/span>\n        return<\/span> self.parse(completion)<\/span>\n<\/span>\n    def<\/span> get_format_instructions<\/span>(self<\/span>) -> str<\/span>:<\/span>\n        \"\"\"Instructions on how the LLM output should be formatted.\"\"\"<\/span><\/span>\n        raise<\/span> NotImplementedError<\/span>\n<\/span>\n    @property<\/span><\/span>\n    def<\/span> _type<\/span>(self<\/span>) -> str<\/span>:<\/span>\n        \"\"\"Return the type key.\"\"\"<\/span><\/span>\n        raise<\/span> NotImplementedError(<\/span>\n            f\"_type property is not implemented in class {self.__class__.__name__}<\/span>.\"<\/span><\/span>\n            \" This is required for serialization.\"<\/span><\/span>\n        )<\/span>\n<\/span>\n    def<\/span> dict<\/span>(self, **kwargs: Any<\/span><\/span>) -> Dict<\/span>:<\/span>\n        \"\"\"Return dictionary representation of output parser.\"\"\"<\/span><\/span>\n        output_parser_dict = super<\/span>().dict<\/span>()<\/span>\n        output_parser_dict[\"_type\"<\/span>] = self._type<\/span><\/span>\n        return<\/span> output_parser_dict<\/span>\n<\/code><\/pre>\n
BaseOutputParser \u662f\u4e00\u4e2a\u57fa\u7840\u7684\u7c7b\uff0c\u53ef\u80fd\u88ab\u5176\u4ed6\u7279\u5b9a\u7684\u8f93\u51fa\u89e3\u6790\u5668\u7ee7\u627f\uff0c\u4ee5\u5b9e\u73b0\u7279\u5b9a\u8bed\u8a00\u6a21\u578b\u7684\u8f93\u51fa\u89e3\u6790\u3002<\/p>\n
\u8fd9\u4e2a\u7c7b\u4f7f\u7528\u4e86Python\u7684ABC\u6a21\u5757\uff0c\u8868\u660e\u5b83\u662f\u4e00\u4e2a\u62bd\u8c61\u57fa\u7c7b\uff08Abstract Base Class\uff09\uff0c\u4e0d\u80fd\u88ab\u76f4\u63a5\u5b9e\u4f8b\u5316\uff0c\u800c\u662f\u9700\u8981\u5b50\u7c7b\u7ee7\u627f\u5e76\u5b9e\u73b0\u62bd\u8c61\u65b9\u6cd5\u3002<\/p>\n
Generic[T] \u8868\u793a\u8fd9\u4e2a\u7c7b\u662f\u4e00\u4e2a\u6cdb\u578b\u7c7b\uff0c\u5176\u4e2dT \u662f\u4e00\u4e2a\u7c7b\u578b\u53d8\u91cf\uff0c\u5b83\u8868\u793a\u89e3\u6790\u540e\u7684\u8f93\u51fa\u6570\u636e\u7684\u7c7b\u578b\u3002<\/p>\n
@abstractmethod \u88c5\u9970\u5668\u6807\u8bb0\u4e86 parse \u65b9\u6cd5\uff0c\u8bf4\u660e\u5b83\u662f\u4e00\u4e2a\u62bd\u8c61\u65b9\u6cd5\uff0c\u5fc5\u987b\u5728\u5b50\u7c7b\u4e2d\u5b9e\u73b0\u3002parse \u65b9\u6cd5\u63a5\u53d7\u4e00\u4e2a\u5b57\u7b26\u4e32\u53c2\u6570 text\uff0c\u901a\u5e38\u662f\u8bed\u8a00\u6a21\u578b\u7684\u8f93\u51fa\u6587\u672c\uff0c\u7136\u540e\u5c06\u5176\u89e3\u6790\u6210\u7279\u5b9a\u7684\u6570\u636e\u7ed3\u6784\uff0c\u5e76\u8fd4\u56de\u3002<\/p>\n
parse_with_prompt \u65b9\u6cd5\u4e5f\u662f\u4e00\u4e2a\u62bd\u8c61\u65b9\u6cd5\uff0c\u63a5\u53d7\u4e24\u4e2a\u53c2\u6570\uff0ccompletion \u662f\u8bed\u8a00\u6a21\u578b\u7684\u8f93\u51fa\uff0cprompt \u662f\u4e0e\u8f93\u51fa\u76f8\u5173\u7684\u63d0\u793a\u4fe1\u606f\u3002\u8fd9\u4e2a\u65b9\u6cd5\u662f\u53ef\u9009\u7684\uff0c\u53ef\u4ee5\u7528\u4e8e\u5728\u9700\u8981\u65f6\u89e3\u6790\u8f93\u51fa\uff0c\u53ef\u80fd\u6839\u636e\u63d0\u793a\u4fe1\u606f\u6765\u8c03\u6574\u8f93\u51fa\u3002<\/p>\n
get_format_instructions \u65b9\u6cd5\u8fd4\u56de\u5173\u4e8e\u5982\u4f55\u683c\u5f0f\u5316\u8bed\u8a00\u6a21\u578b\u8f93\u51fa\u7684\u8bf4\u660e\u3002\u8fd9\u4e2a\u65b9\u6cd5\u53ef\u4ee5\u7528\u4e8e\u63d0\u4f9b\u89e3\u6790\u540e\u6570\u636e\u7684\u683c\u5f0f\u5316\u4fe1\u606f\u3002<\/p>\n
_type \u662f\u4e00\u4e2a\u5c5e\u6027\uff0c\u53ef\u80fd\u7528\u4e8e\u6807\u8bc6\u8fd9\u4e2a\u89e3\u6790\u5668\u7684\u7c7b\u578b\uff0c\u7528\u4e8e\u540e\u7eed\u7684\u5e8f\u5217\u5316\u6216\u5176\u4ed6\u64cd\u4f5c\u3002<\/p>\n
dict \u65b9\u6cd5\u8fd4\u56de\u4e00\u4e2a\u5305\u542b\u8f93\u51fa\u89e3\u6790\u5668\u4fe1\u606f\u7684\u5b57\u5178\uff0c\u8fd9\u4e2a\u5b57\u5178\u53ef\u4ee5\u7528\u4e8e\u5e8f\u5217\u5316\u6216\u5176\u4ed6\u64cd\u4f5c\u3002<\/p>\n
\u5176\u4e2d\u5b50\u7c7b\u5fc5\u987b\u8981\u5b9e\u73b0\u7684\u65b9\u6cd5\u5c31\u662fparse\u3002\u5176\u4ed6\u7684\u90fd\u505a\u4e3a\u8f85\u52a9\u4f5c\u7528\u3002<\/p>\n
langchain\u4e2d\u6709\u54ea\u4e9bOutput Parser<\/h1>\n
\u90a3\u4e48langchain\u4e2d\u6709\u54ea\u4e9bOutput Parser\u7684\u5177\u4f53\u5b9e\u73b0\u5462\uff1f\u5177\u4f53\u5bf9\u5e94\u6211\u4eec\u5e94\u7528\u4e2d\u7684\u4ec0\u4e48\u573a\u666f\u5462\uff1f<\/p>\n
\u63a5\u4e0b\u6765\u6211\u4eec\u5c06\u4f1a\u4e00\u4e00\u9053\u6765\u3002<\/p>\n
List parser<\/h2>\n
ListOutputParser\u7684\u4f5c\u7528\u5c31\u662f\u628aLLM\u7684\u8f93\u51fa\u8f6c\u6210\u4e00\u4e2alist\u3002ListOutputParser\u4e5f\u662f\u4e00\u4e2a\u57fa\u7c7b\uff0c\u6211\u4eec\u5177\u4f53\u4f7f\u7528\u7684\u662f\u4ed6\u7684\u5b50\u7c7b\uff1aCommaSeparatedListOutputParser\u3002<\/p>\n
\u770b\u4e00\u4e0b\u4ed6\u7684parse\u65b9\u6cd5\uff1a<\/p>\n
<\/div>
python<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>    def<\/span> parse<\/span>(self, text: str<\/span><\/span>) -> List<\/span>[str<\/span>]:<\/span>\n        \"\"\"Parse the output of an LLM call.\"\"\"<\/span><\/span>\n        return<\/span> text.strip().split(\", \"<\/span>)<\/span>\n<\/code><\/pre>\n
\u8fd8\u6709\u4e00\u4e2aget_format_instructions:<\/p>\n
<\/div>
python<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>    def<\/span> get_format_instructions<\/span>(self<\/span>) -> str<\/span>:<\/span>\n        return<\/span> (<\/span>\n            \"Your response should be a list of comma separated values, \"<\/span><\/span>\n            \"eg: `foo, bar, baz`\"<\/span><\/span>\n        )<\/span>\n<\/code><\/pre>\n
get_format_instructions\u662f\u544a\u8bc9LLM\u4ee5\u4ec0\u4e48\u6837\u7684\u683c\u5f0f\u8fdb\u884c\u6570\u636e\u7684\u8fd4\u56de\u3002<\/p>\n
\u5c31\u662f\u628aLLM\u7684\u8f93\u51fa\u7528\u9017\u53f7\u8fdb\u884c\u5206\u5272\u3002<\/p>\n
\u4e0b\u9762\u662f\u4e00\u4e2a\u57fa\u672c\u7684\u4f7f\u7528\u4f8b\u5b50\uff1a<\/p>\n
<\/div>
ini<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>output_parser<\/span> = CommaSeparatedListOutputParser()<\/span>\n<\/span>\nformat_instructions<\/span> = output_parser.get_format_instructions()<\/span>\nprompt<\/span> = PromptTemplate(<\/span>\n    template<\/span>=\"\u5217\u51fa\u51e0\u79cd{subject}.\\n{format_instructions}\"<\/span>,<\/span>\n    input_variables<\/span>=[\"subject\"<\/span>],<\/span>\n    partial_variables<\/span>={\"format_instructions\"<\/span>: format_instructions}<\/span>\n)<\/span>\n<\/span>\n_input<\/span> = prompt.format(subject=\"\u6c34\u679c\"<\/span>)<\/span>\noutput<\/span> = model(_input)<\/span>\nprint(output)<\/span>\nprint(output_parser.parse(output))<\/span>\n<\/code><\/pre>\n
\u6211\u4eec\u53ef\u4ee5\u5f97\u5230\u4e0b\u9762\u7684\u8f93\u51fa\uff1a<\/p>\n
<\/div>
css<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>Apple, Orange, Banana, Grape, Watermelon, Strawberry, Pineapple, Peach, Mango, Cherry<\/span>\n['Apple'<\/span>, 'Orange'<\/span>, 'Banana'<\/span>, 'Grape'<\/span>, 'Watermelon'<\/span>, 'Strawberry'<\/span>, 'Pineapple'<\/span>, 'Peach'<\/span>, 'Mango'<\/span>, 'Cherry'<\/span>]<\/span><\/span>\n<\/code><\/pre>\n
\u770b\u5230\u8fd9\u91cc\uff0c\u5927\u5bb6\u53ef\u80fd\u6709\u7591\u95ee\u4e86\uff0c \u4e3a\u4ec0\u4e48\u6211\u4eec\u95ee\u7684\u662f\u4e2d\u6587\uff0c\u8fd4\u56de\u7684\u5374\u662f\u56e0\u4e3a\u5462\uff1f<\/p>\n
\u8fd9\u662f\u56e0\u4e3aoutput_parser.get_format_instructions\u5c31\u662f\u7528\u82f1\u6587\u63cf\u8ff0\u7684\uff0c\u6240\u4ee5LLM\u4f1a\u81ea\u7136\u7684\u7528\u82f1\u6587\u6765\u56de\u7b54\u3002<\/p>\n
\u522b\u6025\uff0c\u6211\u4eec\u53ef\u4ee5\u7a0d\u5fae\u4fee\u6539\u4e0b\u8fd0\u884c\u4ee3\u7801\uff0c\u5982\u4e0b\uff1a<\/p>\n
<\/div>
ini<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>output_parser<\/span> = CommaSeparatedListOutputParser()<\/span>\n<\/span>\nformat_instructions<\/span> = output_parser.get_format_instructions()<\/span>\nprompt<\/span> = PromptTemplate(<\/span>\n    template<\/span>=\"\u5217\u51fa\u51e0\u79cd{subject}.\\n{format_instructions}\"<\/span>,<\/span>\n    input_variables<\/span>=[\"subject\"<\/span>],<\/span>\n    partial_variables<\/span>={\"format_instructions\"<\/span>: format_instructions + \"\u7528\u4e2d\u6587\u56de\u7b54\"<\/span>}<\/span>\n)<\/span>\n<\/span>\n_input<\/span> = prompt.format(subject=\"\u6c34\u679c\"<\/span>)<\/span>\noutput<\/span> = model(_input)<\/span>\nprint(output)<\/span>\nprint(output_parser.parse(output))<\/span>\n<\/code><\/pre>\n
\u6211\u4eec\u5728format_instructions\u4e4b\u540e\uff0c\u63d0\u793aLLM\u9700\u8981\u7528\u4e2d\u6587\u6765\u56de\u7b54\u95ee\u9898\u3002\u8fd9\u6837\u6211\u4eec\u5c31\u53ef\u4ee5\u5f97\u5230\u4e0b\u9762\u7684\u7ed3\u679c\uff1a<\/p>\n
<\/div>
css<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>\u82f9\u679c,\u6a58\u5b50,\u9999\u8549,\u68a8,\u8461\u8404,\u8292\u679c,\u67e0\u6aac,\u6843<\/span>\n['\u82f9\u679c,\u6a58\u5b50,\u9999\u8549,\u68a8,\u8461\u8404,\u8292\u679c,\u67e0\u6aac,\u6843'<\/span>]<\/span><\/span>\n<\/code><\/pre>\n
\u662f\u4e0d\u662f\u5f88\u68d2\uff1f<\/p>\n
Datetime parser<\/h2>\n
DatetimeOutputParser\u7528\u6765\u5c06LLM\u7684\u8f93\u51fa\u8fdb\u884c\u65f6\u95f4\u7684\u683c\u5f0f\u5316\u3002<\/p>\n
<\/div>
python<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>class<\/span> DatetimeOutputParser<\/span>(BaseOutputParser[datetime]):<\/span>\n    format<\/span>: str<\/span> = \"%Y-%m-%dT%H:%M:%S.%fZ\"<\/span><\/span>\n<\/span>\n    def<\/span> get_format_instructions<\/span>(self<\/span>) -> str<\/span>:<\/span>\n        examples = comma_list(_generate_random_datetime_strings(self.format<\/span>))<\/span>\n        return<\/span> f\"\"\"Write a datetime string that matches the <\/span><\/span>\n            following pattern: \"{self.format<\/span>}<\/span>\". Examples: {examples}<\/span>\"\"\"<\/span>\n<\/span>\n    def<\/span> parse<\/span>(self, response: str<\/span><\/span>) -> datetime:<\/span>\n        try<\/span>:<\/span>\n            return<\/span> datetime.strptime(response.strip(), self.format<\/span>)<\/span>\n        except<\/span> ValueError as<\/span> e:<\/span>\n            raise<\/span> OutputParserException(<\/span>\n                f\"Could not parse datetime string: {response}<\/span>\"<\/span><\/span>\n            ) from<\/span> e<\/span>\n<\/span>\n    @property<\/span><\/span>\n    def<\/span> _type<\/span>(self<\/span>) -> str<\/span>:<\/span>\n        return<\/span> \"datetime\"<\/span><\/span>\n<\/code><\/pre>\n
\u5728get_format_instructions\u4e2d\uff0c\u4ed6\u544a\u8bc9LLM\u8fd4\u56de\u7684\u7ed3\u679c\u662f\u4e00\u4e2a\u65e5\u671f\u7684\u5b57\u7b26\u4e32\u3002<\/p>\n
\u7136\u540e\u5728parse\u65b9\u6cd5\u4e2d\u5bf9\u8fd9\u4e2aLLM\u7684\u8f93\u51fa\u8fdb\u884c\u683c\u5f0f\u5316\uff0c\u6700\u540e\u8fd4\u56dedatetime\u3002<\/p>\n
\u6211\u4eec\u770b\u4e0b\u5177\u4f53\u7684\u5e94\u7528\uff1a<\/p>\n
<\/div>
ini<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>output_parser<\/span> = DatetimeOutputParser()<\/span>\ntemplate<\/span> = \"\"\"\u56de\u7b54\u4e0b\u9762\u95ee\u9898:<\/span><\/span>\n{question}<\/span>\n{format_instructions}\"\"\"<\/span>\nprompt<\/span> = PromptTemplate.from_template(<\/span>\n    template,<\/span>\n    partial_variables<\/span>={\"format_instructions\"<\/span>: output_parser.get_format_instructions()},<\/span>\n)<\/span>\nchain<\/span> = LLMChain(prompt=prompt, llm=model)<\/span>\noutput<\/span> = chain.run(\"\u4e2d\u534e\u4eba\u6c11\u5171\u548c\u56fd\u662f\u4ec0\u4e48\u65f6\u5019\u6210\u7acb\u7684?\"<\/span>)<\/span>\nprint(output)<\/span>\nprint(output_parser.parse(output))<\/span>\n<\/code><\/pre>\n
<\/div>
yaml<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>1949-10-01T00:00:00.000000Z<\/span><\/span>\n1949-10-01 00:00:00<\/span><\/span>\n<\/code><\/pre>\n
\u56de\u7b54\u7684\u8fd8\u4e0d\u9519\uff0c\u7ed9\u4ed6\u70b9\u4e2a\u8d5e\u3002<\/p>\n
Enum parser<\/h2>\n
\u5982\u679c\u4f60\u6709\u679a\u4e3e\u7684\u7c7b\u578b\uff0c\u90a3\u4e48\u53ef\u4ee5\u5c1d\u8bd5\u4f7f\u7528EnumOutputParser.<\/p>\n
EnumOutputParser\u7684\u6784\u9020\u51fd\u6570\u9700\u8981\u4f20\u5165\u4e00\u4e2aEnum,\u6211\u4eec\u4e3b\u8981\u770b\u4e0b\u4ed6\u7684\u4e24\u4e2a\u65b9\u6cd5\uff1a<\/p>\n
<\/div>
python<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>    @property<\/span><\/span>\n    def<\/span> _valid_values<\/span>(self<\/span>) -> List<\/span>[str<\/span>]:<\/span>\n        return<\/span> [e.value for<\/span> e in<\/span> self.enum]<\/span>\n<\/span>\n    def<\/span> parse<\/span>(self, response: str<\/span><\/span>) -> Any<\/span>:<\/span>\n        try<\/span>:<\/span>\n            return<\/span> self.enum(response.strip())<\/span>\n        except<\/span> ValueError:<\/span>\n            raise<\/span> OutputParserException(<\/span>\n                f\"Response '{response}<\/span>' is not one of the \"<\/span><\/span>\n                f\"expected values: {self._valid_values}<\/span>\"<\/span><\/span>\n            )<\/span>\n<\/span>\n    def<\/span> get_format_instructions<\/span>(self<\/span>) -> str<\/span>:<\/span>\n        return<\/span> f\"Select one of the following options: {', '<\/span>.join(self._valid_values)}<\/span>\"<\/span><\/span>\n<\/code><\/pre>\n
parse\u65b9\u6cd5\u63a5\u6536\u4e00\u4e2a\u5b57\u7b26\u4e32 response\uff0c\u5c1d\u8bd5\u5c06\u5176\u89e3\u6790\u4e3a\u679a\u4e3e\u7c7b\u578b\u7684\u4e00\u4e2a\u6210\u5458\u3002\u5982\u679c\u89e3\u6790\u6210\u529f\uff0c\u5b83\u4f1a\u8fd4\u56de\u8be5\u679a\u4e3e\u6210\u5458\uff1b\u5982\u679c\u89e3\u6790\u5931\u8d25\uff0c\u5b83\u4f1a\u629b\u51fa\u4e00\u4e2a OutputParserException \u5f02\u5e38\uff0c\u5f02\u5e38\u4fe1\u606f\u4e2d\u5305\u542b\u4e86\u6240\u6709\u6709\u6548\u503c\u7684\u5217\u8868\u3002<\/p>\n
get_format_instructions\u544a\u8bc9LLM\u9700\u8981\u4eceEnum\u7684\u6709\u6548value\u4e2d\u9009\u62e9\u4e00\u4e2a\u8f93\u51fa\u3002\u8fd9\u6837parse\u624d\u80fd\u63a5\u53d7\u5230\u6b63\u786e\u7684\u8f93\u5165\u503c\u3002<\/p>\n
\u5177\u4f53\u4f7f\u7528\u7684\u4f8b\u5b50\u53ef\u4ee5\u53c2\u8003\u524d\u9762\u4e24\u4e2aparser\u7684\u7528\u6cd5\u3002\u7bc7\u5e45\u8d77\u89c1\uff0c\u8fd9\u91cc\u5c31\u4e0d\u5217\u4e86\u3002<\/p>\n
Pydantic (JSON) parser<\/h2>\n
JSON\u53ef\u80fd\u662f\u6211\u4eec\u5728\u65e5\u5e38\u4ee3\u7801\u4e2d\u6700\u5e38\u7528\u7684\u6570\u636e\u7ed3\u6784\u4e86\uff0c\u8fd9\u4e2a\u6570\u636e\u7ed3\u6784\u5f88\u91cd\u8981\u3002<\/p>\n
\u5728langchain\u4e2d\uff0c\u63d0\u4f9b\u7684JSON parser\u53eb\u505a\uff1aPydanticOutputParser\u3002<\/p>\n
\u65e2\u7136\u8981\u8fdb\u884cJSON\u8f6c\u6362\uff0c\u5fc5\u987b\u5f97\u5148\u5b9a\u4e49\u4e00\u4e2aJSON\u7684\u7c7b\u578b\u5bf9\u8c61\uff0c\u7136\u540e\u544a\u8bc9LLM\u5c06\u6587\u672c\u8f93\u51fa\u8f6c\u6362\u6210JSON\u683c\u5f0f\uff0c\u6700\u540e\u8c03\u7528parse\u65b9\u6cd5\u628ajson\u5b57\u7b26\u4e32\u8f6c\u6362\u6210JSON\u5bf9\u8c61\u3002<\/p>\n
\u6211\u4eec\u6765\u770b\u4e00\u4e2a\u4f8b\u5b50\uff1a<\/p>\n
<\/div>
ini<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div><\/span>\nclass Student(BaseModel):<\/span>\n    name: str<\/span> = Field(description=\"\u5b66\u751f\u7684\u59d3\u540d\"<\/span>)<\/span>\n    age: str<\/span> = Field(description=\"\u5b66\u751f\u7684\u5e74\u9f84\"<\/span>)<\/span>\n<\/span>\nstudent_query<\/span> = \"\u544a\u8bc9\u6211\u4e00\u4e2a\u5b66\u751f\u7684\u4fe1\u606f\"<\/span><\/span>\n<\/span>\nparser<\/span> = PydanticOutputParser(pydantic_object=Student)<\/span>\n<\/span>\nprompt<\/span> = PromptTemplate(<\/span>\n    template<\/span>=\"\u56de\u7b54\u4e0b\u9762\u95ee\u9898.\\n{format_instructions}\\n{query}\\n\"<\/span>,<\/span>\n    input_variables<\/span>=[\"query\"<\/span>],<\/span>\n    partial_variables<\/span>={\"format_instructions\"<\/span>: parser.get_format_instructions()+\"\u7528\u4e2d\u6587\u56de\u7b54\"<\/span>},<\/span>\n)<\/span>\n<\/span>\n_input<\/span> = prompt.format_prompt(query=student_query)<\/span>\n<\/span>\noutput<\/span> = model(_input.to_string())<\/span>\nprint(output)<\/span>\nprint(parser.parse(output))<\/span>\n<\/code><\/pre>\n
\u8fd9\u91cc\u6211\u4eec\u5b9a\u4e49\u4e86\u4e00\u4e2aStudent\u7684\u7ed3\u6784\u4f53\uff0c\u7136\u540e\u8ba9LLM\u7ed9\u6211\u4e00\u4e2a\u5b66\u751f\u7684\u4fe1\u606f\uff0c\u5e76\u7528json\u7684\u683c\u5f0f\u8fdb\u884c\u8fd4\u56de\u3002<\/p>\n
\u4e4b\u540e\u6211\u4eec\u4f7f\u7528parser.parse\u6765\u89e3\u6790\u8fd9\u4e2ajson\uff0c\u751f\u6210\u6700\u540e\u7684Student\u4fe1\u606f\u3002<\/p>\n
\u6211\u4eec\u53ef\u4ee5\u5f97\u5230\u4e0b\u9762\u7684\u8f93\u51fa\uff1a<\/p>\n
<\/div>
arduino<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>\u793a\u4f8b\u8f93\u51fa\uff1a{\"name\"<\/span>: \"\u5f20\u4e09\"<\/span>, \"age\"<\/span>: \"18\"<\/span>}<\/span>\nname='\u5f20\u4e09'<\/span> age='18'<\/span><\/span>\n<\/code><\/pre>\n
Structured output parser<\/h2>\n
\u867d\u7136PydanticOutputParser\u975e\u5e38\u5f3a\u5927\uff0c \u4f46\u662f\u6709\u65f6\u5019\u6211\u4eec\u53ea\u662f\u9700\u8981\u4e00\u4e9b\u7b80\u5355\u7684\u7ed3\u6784\u8f93\u51fa\uff0c\u90a3\u4e48\u53ef\u4ee5\u8003\u8651StructuredOutputParser.<\/p>\n
\u6211\u4eec\u770b\u4e00\u4e2a\u5177\u4f53\u7684\u4f8b\u5b50\uff1a<\/p>\n
<\/div>
ini<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>response_schemas<\/span> = [<\/span>\n    ResponseSchema(name=\"name\"<\/span>, description=\"\u5b66\u751f\u7684\u59d3\u540d\"<\/span>),<\/span>\n    ResponseSchema(name=\"age\"<\/span>, description=\"\u5b66\u751f\u7684\u5e74\u9f84\"<\/span>)<\/span>\n]<\/span>\noutput_parser<\/span> = StructuredOutputParser.from_response_schemas(response_schemas)<\/span>\n<\/span>\nformat_instructions<\/span> = output_parser.get_format_instructions()<\/span>\nprompt<\/span> = PromptTemplate(<\/span>\n    template<\/span>=\"\u56de\u7b54\u4e0b\u9762\u95ee\u9898.\\n{format_instructions}\\n{question}\"<\/span>,<\/span>\n    input_variables<\/span>=[\"question\"<\/span>],<\/span>\n    partial_variables<\/span>={\"format_instructions\"<\/span>: format_instructions}<\/span>\n)<\/span>\n<\/span>\n_input<\/span> = prompt.format_prompt(question=\"\u7ed9\u6211\u4e00\u4e2a\u5973\u5b69\u7684\u540d\u5b57?\"<\/span>)<\/span>\noutput<\/span> = model(_input.to_string())<\/span>\nprint(output)<\/span>\nprint(output_parser.parse(output))<\/span>\n<\/code><\/pre>\n
\u8fd9\u4e2a\u4f8b\u5b50\u662f\u4e0a\u9762\u7684PydanticOutputParser\u7684\u6539\u5199\uff0c\u4f46\u662f\u66f4\u52a0\u7b80\u5355\u3002<\/p>\n
\u6211\u4eec\u53ef\u4ee5\u5f97\u5230\u4e0b\u9762\u7684\u7ed3\u679c\uff1a<\/p>\n
<\/div>
arduino<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div> ` ` `json<\/span>\n{<\/span>\n\t\"name\"<\/span>: \"Jane\"<\/span>,<\/span>\n\t\"age\"<\/span>: \"18\"<\/span><\/span>\n}<\/span>\n ` ` `<\/span>\n{'name'<\/span>: 'Jane'<\/span>, 'age'<\/span>: '18'<\/span>}<\/span>\n<\/code><\/pre>\n
output\u8fd4\u56de\u7684\u662f\u4e00\u4e2amarkdown\u683c\u5f0f\u7684json\u5b57\u7b26\u4e32\uff0c\u7136\u540e\u901a\u8fc7output_parser.parse\u5f97\u5230\u6700\u540e\u7684json\u3002<\/p>\n
\u5176\u4ed6\u7684\u4e00\u4e9bparser<\/h2>\n
\u9664\u4e86json\uff0cxml\u683c\u5f0f\u4e5f\u662f\u6bd4\u8f83\u5e38\u7528\u7684\u683c\u5f0f\uff0clangchain\u4e2d\u63d0\u4f9b\u7684XML parser\u53eb\u505aXMLOutputParser\u3002<\/p>\n
\u53e6\u5916\uff0c\u5982\u679c\u6211\u4eec\u5728\u4f7f\u7528parser\u7684\u8fc7\u7a0b\u4e2d\u51fa\u73b0\u4e86\u683c\u5f0f\u95ee\u9898\uff0clangchain\u8fd8\u8d34\u5fc3\u7684\u63d0\u4f9b\u4e86\u4e00\u4e2aOutputFixingParser\u3002\u4e5f\u5c31\u662f\u8bf4\u5f53\u7b2c\u4e00\u4e2aparser\u62a5\u9519\u7684\u65f6\u5019\uff0c\u6216\u8005\u8bf4\u4e0d\u80fd\u89e3\u6790LLM\u8f93\u51fa\u7684\u65f6\u5019\uff0c\u5c31\u4f1a\u6362\u6210OutputFixingParser\u6765\u5c1d\u8bd5\u4fee\u6b63\u683c\u5f0f\u95ee\u9898\uff1a<\/p>\n
<\/div>
ini<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>from langchain.output_parsers import OutputFixingParser<\/span>\n<\/span>\nnew_parser<\/span> = OutputFixingParser.from_llm(parser=parser, llm=ChatOpenAI())<\/span>\n<\/span>\nnew_parser.parse(misformatted)<\/span>\n<\/span>\n<\/code><\/pre>\n
\u5982\u679c\u9519\u8bef\u4e0d\u662f\u56e0\u4e3a\u683c\u5f0f\u5f15\u8d77\u7684\uff0c\u90a3\u4e48langchain\u8fd8\u63d0\u4f9b\u4e86\u4e00\u4e2aRetryOutputParser,\u6765\u5c1d\u8bd5\u91cd\u8bd5\uff1a<\/p>\n
<\/div>
ini<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>from langchain.output_parsers import RetryWithErrorOutputParser<\/span>\n<\/span>\nretry_parser<\/span> = RetryWithErrorOutputParser.from_llm(<\/span>\n    parser<\/span>=parser, llm=OpenAI(temperature=0<\/span>)<\/span>\n)<\/span>\n<\/span>\nretry_parser.parse_with_prompt(bad_response, prompt_value)<\/span>\n<\/code><\/pre>\n
\u8fd9\u51e0\u4e2aparser\u90fd\u975e\u5e38\u6709\u7528\uff0c\u5927\u5bb6\u53ef\u4ee5\u81ea\u884c\u5c1d\u8bd5\u3002<\/p>\n
\u603b\u7ed3<\/h1>\n
\u867d\u7136langchain\u4e2d\u7684\u6709\u4e9bparser\u6211\u4eec\u53ef\u4ee5\u81ea\u884c\u501f\u52a9python\u8bed\u8a00\u7684\u5404\u79cd\u5de5\u5177\u6765\u5b9e\u73b0\u3002\u4f46\u662f\u6709\u4e00\u4e9bparser\u5b9e\u9645\u4e0a\u662f\u8981\u7ed3\u5408LLM\u4e00\u8d77\u6765\u4f7f\u7528\u7684\uff0c\u6bd4\u5982OutputFixingParser\u548cRetryOutputParser\u3002<\/p>\n
\u6240\u4ee5\u5927\u5bb6\u8fd8\u662f\u5c3d\u53ef\u80fd\u7684\u4f7f\u7528langchain\u63d0\u4f9b\u7684parser\u4e3a\u597d\u3002\u6bd5\u7adf\u8f6e\u5b50\u90fd\u7ed9\u4f60\u9020\u597d\u4e86\uff0c\u8fd8\u8981\u5565\u81ea\u884c\u8f66\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
\u6211\u4eec\u77e5\u9053\u5728\u5927\u8bed\u8a00\u6a21\u578b\u4e2d, \u4e0d\u7ba1\u6a21\u578b\u7684\u80fd\u529b\u6709\u591a\u5f3a\u5927\uff0c\u4ed6\u7684\u8f93\u5165\u548c\u8f93\u51fa\u57fa\u672c\u4e0a\u90fd\u662f\u6587\u672c\u683c\u5f0f\u7684\uff0c\u6587\u672c\u683c\u5f0f\u7684\u8f93\u5165\u8f93\u51fa\u867d\u7136\u5bf9\u4eba\u6765\u8bf4\u975e\u5e38\u7684\u53cb\u597d\uff0c\u4f46\u662f\u5982\u679c\u6211\u4eec\u60f3\u8981\u8fdb\u884c\u4e00\u4e9b\u7ed3\u6784\u5316\u5904\u7406\u7684\u8bdd\u8fd8\u662f\u4f1a\u6709\u4e00\u70b9\u70b9\u7684\u4e0d\u65b9\u4fbf\u3002 \u4e0d<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"rank_math_title":"","rank_math_description":"","rank_math_focus_keyword":"","views":"2","footnotes":""},"categories":[3],"tags":[136,126,127,128,129],"collection":[],"class_list":["post-1470","post","type-post","status-publish","format-standard","hentry","category-fenlei2","tag-136","tag-gpt","tag-ai","tag-128","tag-129"],"acf":[],"_links":{"self":[{"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/posts\/1470","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/comments?post=1470"}],"version-history":[{"count":0,"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/posts\/1470\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/media?parent=1470"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/categories?post=1470"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/tags?post=1470"},{"taxonomy":"collection","embeddable":true,"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/collection?post=1470"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}