Baidu Intelligent Cloud Qianfan AppBuilder (hereinafter referred to as AppBuilder) provides theZero code creation capabilityorCode status development capability, reduce the threshold of AI native application development.
Zero code creation
Zero code creation supports users to complete application settings, capability expansion and other settings through the application configuration interface, and conduct online testing of applications.You can complete the creation of applications by entering instructions, prologue and recommendation questions, selecting components and knowledge base, setting model configuration, question configuration and knowledge base retrieval methods.
At the same time, you can also click the intelligent generation icon in the upper right corner of each function, and AI will automatically complete the text generation and component selection.
essential information
The basic information includes the name, description and application avatar information of the application. You can click the avatar to upload the avatar, or click the AI generation function to automatically generate the application avatar.
Role Instruction
The role instruction determines the role and task of the application you create. You can click the template button to fill in the role instruction based on the template, or you can fill in the role instruction directly.In the role instruction, describe the tasks and objectives expected to be completed by the role, describe the available components and how to use them, and specify the output format, result content, style requirements or word limit of the answer.
assembly
You can select a variety of components to expand the capability boundary of the big model. You can select [Components Provided by the Platform] or [Create Your Own Components] to add custom components for application configuration.
Multiple components can be selected for configuration in each application. It is recommended to select 4 or less components to achieve good results.
The introduction and capabilities of the components provided by the platform can be viewed in details on the "Component Square" page.
At the same time, you can also click the [Create Component] function, refer toWorkflow creation component, complete the creation of customized components through workflow, and set the ability of components by yourself.
knowledge base
The big model will answer questions based on the knowledge document you upload. You can realize the function of knowledge question and answer by referencing the knowledge base file.
Click the "plus sign" next to the knowledge base to pop up the pop-up window of adding knowledge base. Click "Add" to add the created knowledge base. Click "Create Knowledge Base" to jump to the "My Knowledge" page and create a new knowledge base. An application can call up to 5 knowledge bases.Relevant contents of the knowledge base can be viewed in the documentPersonal space - my knowledge。
After selecting the knowledge base, you can click the Advanced Configuration button to adjust the knowledge base retrieval strategy and parameters. Knowledge base retrieval includes three strategies: full-text retrieval, semantic retrieval, and hybrid retrieval:
Full text search: The inverted index strategy is used for retrieval recall, and it is recommended to use it in scenarios where precise keyword matching is required.
Semantic retrieval: returns the content matching the query meaning, not the content matching the query literal meaning.It is recommended to use it in scenarios where context dependency and intention dependency are required.
Hybrid retrieval: Use inverted index and semantic retrieval strategies for recall. It is recommended to use in scenarios where sentence understanding and semantic relevance are required, so the comprehensive effect is better.
In addition to the retrieval strategy, you can determine the results returned to the large model from the knowledge base retrieval by adjusting the parameters of the recall number and matching score:
Recall quantity: represents the number of fragments recalled from the knowledge base that match the input query. The larger the number, the more fragments recalled
Matching score: It is used to calculate the similarity between the input Query and the original fragment of the knowledge base in the retrieval process. The recalled fragments that are higher than or equal to the matching score will be finally input into the big model, and you can set the matching score according to your own needs.For example, properly increasing the matching score can output answers that are more closely related to query, and lowering the matching score can recall more answers, and ultimately the answer of the big model is more flexible.
tips: You can enter the target knowledge base to useHit testThe function is used to test the retrieval results of query in a specific knowledge base corpus, and then modify the "Recall Quantity" and "Match Score" threshold buttons of the application configuration page - knowledge base configuration options according to these retrieval results and the corresponding scores to filter out the retrieval results that are more consistent with the expectations.
data base
Quote structured data, implement table Q&A, and support single table Q&A or multi table joint Q&A.The big model will automatically call data to answer questions according to your questions and data table descriptions. It can be used together with other components and the knowledge base. Each application supports the association of up to one database.
Difference between database and knowledge base:
The structured data stored in the database are data of numerical value, text, integer, decimal, date, time and percentage types. The big model will automatically call the data table and generate code to query, reason and analyze the data table according to your query.
The structured data of knowledge Q&A stored in the knowledge base can only be of text type. According to the uploaded structured data of knowledge Q&A FAQ, the big model will hit the uploaded questions more accurately and generate responses based on the uploaded answers.
Click the "plus" sign next to the database to reference a database that has been created. Click the "Add" button in the pop-up box to add a database. This operation will add all data tables in the database by default. If you have multiple databases in a database and only want to add a few of them, you can expand the database,Check the data table to be added.You can also click "Create Database" to jump to Personal Center - Database to re create a database.
Recommendation question
It supports the configuration of application recommendation questions, up to 3.
Follow up
It supports automatic questioning after the last round of replies based on the user's recent conversation.You can select the default mode or customize the question prompt.
Model configuration
You can select a thinking model and a question and answer model in the model configuration. The thinking model is used for task planning and component selection, and the question and answer model is used to summarize and generate response results.You can refer to the platform resourcesLarge model serviceMake model selection and use.The payment status of the model is turned off by default. After the model is turned on, if you have activated the payment model, the free resources will be automatically switched to the paid resources after consumption to ensure that the application is stable and available. If you need to turn it on, please click the [Modify Configuration] button to go toResource quotaPage.
Upload file
It supports uploading xlsx, jsonl, png, pdf format files in dialog boxes, and can cooperate with code interpreter, image content understanding and other tools to achieve excel/json data analysis, statistical analysis and drawing, insight into conclusions and other capabilities.
App publishing
After completing the application configuration and effect debugging, click the [Publish] button in the upper right corner to conduct multi-channel publishing and support the creation of API call keys.
Multi channel publishing
You can share this link with other developers so that they can immediately experience your AI native application demo on the web experience page;At the same time, the application can be distributed to more Baidu ecosystems through the Holy Matrix, expanding user traffic.
After completing the relevant channel configuration, you can publish the application to WeChat customer service and WeChat public account for users to use.
Please refer toInstructions for AppBuilder SDKConduct code status development.AppBuilder SDK provides a complete AI native application development kit, including rich development components and application sample code.