前言

chatGPT目前的明显问题是不能够获取新知识，也没有办法和外界交互，而plugin就是来解决这个问题的。

chatgpt-retrieval-plugin

插件的基本信息

这是最新开源的一个plugin，里面有几个核心点
第一个是插件的定义

我不逐句翻译了，大概意思就是可以扩展chatGPT，这样就可以试用外部的资料和调用外部的服务了。
然后又说明了一个插件需要的三个部分

1. 需要一个API
2. 需要一个API的定义
3. 需要一个描述这个插件的json文件

插件和chatGPT之间的核心逻辑

这个图其实就是插件与chatGPT的一个核心交互模式。

解释下里面的的核心点：chatGPT是如何和工具进行交互的。
这里就涉及到刚才插件要求的三个要素，Api、Api文档、插件描述。
我们核心看插件描述和API文档

api-plugin.json

{
  "schema_version": "v1",
  "name_for_model": "retrieval",
  "name_for_human": "Retrieval Plugin",
  "description_for_model": "Plugin for searching through the user's documents (such as files, emails, and more) to find answers to questions and retrieve relevant information. Use it whenever a user asks something that might be found in their personal information.",
  "description_for_human": "Search through your documents.",
  "auth": {
    "type": "user_http",
    "authorization_type": "bearer"
  },
  "api": {
    "type": "openapi",
    "url": "https://your-app-url.com/.well-known/openapi.yaml",
    "has_user_authentication": false
  },
  "logo_url": "https://your-app-url.com/.well-known/logo.png",
  "contact_email": "hello@contact.com", 
  "legal_info_url": "hello@legal.com"
}

openapi.yaml

openapi: 3.0.2
info:
  title: Retrieval Plugin API
  description: A retrieval API for querying and filtering documents based on natural language queries and metadata
  version: 1.0.0
  servers:
    - url: https://your-app-url.com
paths:
  /query:
    post:
      summary: Query
      description: Accepts search query objects array each with query and optional filter. Break down complex questions into sub-questions. Refine results by criteria, e.g. time / source, don't do this often. Split queries if ResponseTooLargeError occurs.
      operationId: query_query_post
      requestBody:
        content:
          application/json:
            schema:
              $ref: "#/components/schemas/QueryRequest"
        required: true
      responses:
        "200":
          description: Successful Response
          content:
            application/json:
              schema:
                $ref: "#/components/schemas/QueryResponse"
        "422":
          description: Validation Error
          content:
            application/json:
              schema:
                $ref: "#/components/schemas/HTTPValidationError"
      security:
        - HTTPBearer: []
components:
  schemas:
    DocumentChunkMetadata:
      title: DocumentChunkMetadata
      type: object
      properties:
        source:
          $ref: "#/components/schemas/Source"
        source_id:
          title: Source Id
          type: string
        url:
          title: Url
          type: string
        created_at:
          title: Created At
          type: string
        author:
          title: Author
          type: string
        document_id:
          title: Document Id
          type: string
    DocumentChunkWithScore:
      title: DocumentChunkWithScore
      required:
        - text
        - metadata
        - score
      type: object
      properties:
        id:
          title: Id
          type: string
        text:
          title: Text
          type: string
        metadata:
          $ref: "#/components/schemas/DocumentChunkMetadata"
        embedding:
          title: Embedding
          type: array
          items:
            type: number
        score:
          title: Score
          type: number
    DocumentMetadataFilter:
      title: DocumentMetadataFilter
      type: object
      properties:
        document_id:
          title: Document Id
          type: string
        source:
          $ref: "#/components/schemas/Source"
        source_id:
          title: Source Id
          type: string
        author:
          title: Author
          type: string
        start_date:
          title: Start Date
          type: string
        end_date:
          title: End Date
          type: string
    HTTPValidationError:
      title: HTTPValidationError
      type: object
      properties:
        detail:
          title: Detail
          type: array
          items:
            $ref: "#/components/schemas/ValidationError"
    Query:
      title: Query
      required:
        - query
      type: object
      properties:
        query:
          title: Query
          type: string
        filter:
          $ref: "#/components/schemas/DocumentMetadataFilter"
        top_k:
          title: Top K
          type: integer
          default: 3
    QueryRequest:
      title: QueryRequest
      required:
        - queries
      type: object
      properties:
        queries:
          title: Queries
          type: array
          items:
            $ref: "#/components/schemas/Query"
    QueryResponse:
      title: QueryResponse
      required:
        - results
      type: object
      properties:
        results:
          title: Results
          type: array
          items:
            $ref: "#/components/schemas/QueryResult"
    QueryResult:
      title: QueryResult
      required:
        - query
        - results
      type: object
      properties:
        query:
          title: Query
          type: string
        results:
          title: Results
          type: array
          items:
            $ref: "#/components/schemas/DocumentChunkWithScore"
    Source:
      title: Source
      enum:
        - email
        - file
        - chat
      type: string
      description: An enumeration.
    ValidationError:
      title: ValidationError
      required:
        - loc
        - msg
        - type
      type: object
      properties:
        loc:
          title: Location
          type: array
          items:
            anyOf:
              - type: string
              - type: integer
        msg:
          title: Message
          type: string
        type:
          title: Error Type
          type: string
  securitySchemes:
    HTTPBearer:
      type: http
      scheme: bearer

这两个部分有点长，我摘出重点
在插件描述中：

"name_for_model": "retrieval",
"description_for_model": "Plugin for searching through the user's documents (such as files, emails, and more) to find answers to questions and retrieve relevant information. Use it whenever a user asks something that might be found in their personal information.",

api定义中

 /query:
    post:
      summary: Query
      description: Accepts search query objects array each with query and optional filter. Break down complex questions into sub-questions. Refine results by criteria, e.g. time / source, don't do this often. Split queries if ResponseTooLargeError occurs.
      operationId: query_query_post

这两段逻辑就是ChatGPT可以判断是否要调用这个接口的依据，也就是说这两段会传如到prompt中去。
而具体怎么调用就是通过openapi.yaml 来获取的接口知识，这个对gpt来说实在是太简单。

总结

这个逻辑和LangChain的agent模块基本一致。目前只有一个核心区别，就是这个插件看起来是只作用于ChatGPT，也就是说我们需要登录到chatGPT的网站上通过对话才能使用这个插件。而LangChain不同，他是一个工具，可以开发一个独立与ChatGPT的网站或者工具。

如果我是LangChain，就直接着手开发对接plugin的Agent或者tool咯