After creating an application, enter the App Settings Page to set the application model, knowledge base, and output. This document uses the standard mode application as an example to introduce App Settings in detail.
Basic settings
After creating the application, click the app avatar on the upper left corner. You can modify basic information such as the application avatar, application name, and application introduction in the Edit Application pop-up window, or switch application modes. For details about each mode, see Agent Application and Its Three Modes.
After completing the settings, the settings page will differ depending on the application mode. Refer to the table below for details:
Configuration Content
Standard Mode
Single Workflow Mode
Multi-Agent Mode
Basic settings
Application name, avatar, and welcome words remain constant across different modes.
App settings
App settings are independent across different modes, they are not inherited when switching modes.
App settings are independent across different modes, they are not inherited when switching modes.
Prompts, plugins, and other settings are not supported in Single Workflow Mode.
App settings are independent across different modes, they are not inherited when switching modes.
Multi-Agent Mode is basically consistent with Standard Mode in terms of the range of setting items, but the specific options differ.
Once finished setting up the application avatar and name and released the app, they will be displayed in the user interface window.
Model Settings
You can select model and change its configuration in Model Settings. The thinking model is used for user intent recognition (Standard Mode), task planning, and plugin selection (Multi-Agent Mode). The Generative Model is used for reading comprehension, summarization, and response generation. For details on each model, see Model Introduction.
Model Settings related features are as follows:
Model service: New users of Tencent Cloud ADP will automatically obtain a certain amount of free quota. You can select different types of models to perform free application debugging. Based on test results, you can check Purchase Method and use.
Context rounds: Sets the number of dialogue history rounds input as prompt to the model. The more rounds, the higher the correlation of multi-turn dialogue, but also the higher the token consumption.
Parameters:
Temperature: Controls randomness and diversity. A higher value makes outputs more random and creative, suitable for poetry and creative writing. A lower value makes outputs more focused and deterministic, suitable for code generation.
top_p: Controls the diversity of the texts generated by the model. top_p is a nucleus sampling method where the model will consider the most probable words whose cumulative probability reaches the top_p threshold.
Maximum output length: Limits the maximum length of generated text. Helps control API cost and response time, and prevents meaningless long outputs.
Model Comparison Debugging
In Standard Mode and in Multi-Agent Mode with a single Agent, "multi-model comparison debugging" is supported.
You can compare responses of different models to the same question, up to 3 models at the same time. Each model supports parameter tuning. If you find a better-performing model, click Choose to select it. The selected model will take effect after exiting the comparison and saving. If you choose to discard and exit, the selected model will not take effect.
Role Commands
After the user asks a question, the application will respond according to the role defined in Role Commands. You can specify requirements such as response language or tone. The Tencent Cloud Agent Development Platform currently supports both Chinese and English Q&A output.
Template: Predefined prompt format templates are provided. It is recommended to fill in according to the template for better effect. After editing instructions, you can also click Template > Save as Template to save the written command as a template.
AI One-Click Optimization: After initial role settings, click One-Click Optimization to refine the role definition. The model will optimize the input for better execution.
Note:
The AI one-click optimization feature will consume the user's token resources.
Welcome Words
After filling in the welcome words, it will be displayed on the client homepage. Application-level variables can be inserted. You can also use AI one-click optimization to generate the welcome message.
Knowledge
The Knowledge Base shows the default knowledge base and referenced knowledge bases of the application. For each knowledge base, you can set: retrieval and recall rules, retrieval scope, and knowledge base model.
1. Retrieval and recall rules:
Retrieval strategy: Choose between mixed retrieval or semantic retrieval.
Mixed retrieval: Executes keyword and vector retrieval simultaneously. Recommended for scenarios requiring both string and semantic association.
Semantic retrieval: Recommended when overlap between query and text is low and semantic matching is needed.
Excel retrieval enhancement: When enabled, supports natural language queries and calculations on Excel tables. May affect response time.
Reranking: When enabled, you can select a re-ranking model. After retrieval recall, the system reorders results by analyzing the user’s question, so that the most similar content is ranked first. The platform provides 2 preset reranking models, and you can also configure third-party re-ranking models via the Model Marketplace.
File: When enabled, the large model answers based on your constructed file library in the knowledge base. You can directly upload files or upload web pages, which the model will parse and learn. For file-related details, see File Overview.
Recall files: retrieve and return the top N document fragments with the highest matching degree as input for the large model to perform reading comprehension.
File retrieval accuracy: Based on the configured similarity threshold. Lower values recall more chunks, but may reduce accuracy. Content below the threshold will not be recalled.
Q&A: When enabled, the model will answer based on your constructed Q&A library. You can upload files in bulk to import Q&A, manually enter Q&A content, or auto-generate Q&A from files in the document library. For Q&A-related details, see Q&A.
Response from Q&A library: Choose between reply directly and reply after polished.
Reply directly: When the detected similarity exceeds the threshold, the answer is directly returned.
Reply after polished: When an answer is detected, the answer is polished before being returned.
Q&A recalled: The top-N Q&A entries with highest similarity are returned as input for the large model to perform reading comprehension.
Q&A retrieval accuracy: Similar logic as documents. Lower values recall more entries, but may reduce accuracy. Below-threshold entries are not recalled.
Database: When enabled, the model will answer questions based on the integrated third-party database.
3. Knowledge base model: Supports setting the knowledge base Q&A generation model and the knowledge base schema generation model.
Web Search
When Web Search is enabled, the model can combine online information to provide users with more real-time and richer answers.
Workflow
Workflows are used for complex business scenario interactions. On the Workflow Management page, you can control the workflow enable/disable status. For workflow configuration processes and introduction, see What Is a Workflow.
Output Settings
Output method: Supports choosing whether answers are output in streaming form (character-by-character) or non-streaming form (output all at once after generation).
Response settings: Supports choosing whether to use the large model to answer all questions, or to answer conservatively only for questions outside the knowledge base.
Variables and Memory
Variables
Variables are visible across the entire application scope, including system variables, environment variables, API parameters, and application variables. For details, see Variables. Click Variable to access the variable management webpage.
System variable: Runtime variables of the application. Cannot be customized or modified.
Environment variables: Used to store sensitive information such as API keys and user passwords. You can add parameters manually. Click Create to create a new environment variable and provide a default value.
API parameters: Represent variables passed through the custom_variables field when calling the Agent Development Platform API (see Dialog API Documentation (HTTP SSE) and Dialog API Documentation (WebSocket) for details). These can be referenced in workflows to execute subsequent logic. You can add parameters manually with default values. Click Create to create a new API parameter, default value setting is supported.
Application variables: Can be read and modified across the global application scope, and can be passed between workflows and Agents. Users can manually modify them.
Long-Term Memory
Long-Term Memory allows the model to remember information during conversations, enabling personalized dialogue experiences. After enabling long-term memory, the model will capture and store personalized user information during interactions. For details, see Long-Term Memory.
Memory retention period can be set between 1–999 days (default: 30 days). Content beyond the retention period will be deleted.
Memory test content displays all saved memory within the retention range, and supports editing, deleting, and clearing all memory content.
Advanced Settings
Synonym Settings
You can import domain-specific terminology for business scenarios. Synonyms in queries will be unified to the standardized terms in the knowledge base before retrieval, improving retrieval accuracy.
Advanced Options for Conversation Effects
Supports configuring whether the model should rewrite context.
Model input context rewriting
When enabled, the large model combines historical Q&A to complete the current question.
Intent Fulfillment Priority
In most cases, setting intent fulfillment priority is not required.
If there is high similarity between your Q&A and workflow settings, it may affect the model’s accuracy in determining which to call. In this case, you can set reasonable priorities.
Application User Permissions
By default, users can view all knowledge. After setting permissions, different users will only be able to access knowledge within their assigned scope. For details, see Application User Permissions.
