Low-Code Agent Framework Guide Part 4: Best Practices for Coze Model and Plugin Settings

Model Integration and Selection

Coze integrates multiple mainstream large models, including Doubao series, DeepSeek series, and Tongyi series. Some models (e.g., Doubao, Jueci Xingchen) also support visual inputs.

Generation Diversity Parameters

Temperature : Controls randomness; higher values increase diversity, lower values improve precision.

Top P : Cumulative probability sampling; higher values increase diversity but may reduce relevance.

Preset Generation Modes

Precise Mode – Strictly follows instructions, high determinism. Suitable for formal documents, code generation.

Balanced Mode – Balances accuracy and creativity. Suitable for everyday conversation, content recommendation.

Creative Mode – Emphasizes novelty and variety. Suitable for brainstorming, story creation, ad copy.

Custom Mode – Allows manual adjustment of parameters for fine‑grained control.

Input/Output Settings

Context Rounds : Number of dialogue turns considered; more rounds improve coherence but increase compute cost.

Maximum Reply Length : Upper token limit; adjust per task (short for customer service, longer for literary generation).

Output Format : Choose text, JSON, Markdown, etc., according to downstream consumption.

Persona Prompt Structure

A high‑quality prompt consists of four parts: context (role and background), instruction (clear goal), input/output examples (marked), and special definitions (to avoid ambiguity).

Give the model a role and capability
Explain the user's role and situation
Specify the reply language style (optional example)
Specify the input content
Specify the output format (with examples)

Case Study: E‑commerce Customer Service Agent

Model: Qwen Max (strong performance for long, complex dialogues)

Context Rounds: 5

Maximum Reply Length: 4096 tokens

Output Format: Text

Temperature: 0.7 (balanced)

Top P: 0.9 (increase diversity while staying reasonable)

Repetition Penalties: 0.2 for sentences and topics

Persona template example:

# Role: E‑commerce Platform Smart Customer Service

- Ability: Identify user intent, match predefined reply templates, perform context analysis, and assist with order queries while respecting privacy policies.

## Background:
The platform aims to improve service quality and efficiency through an AI assistant.

## Reply Style:
1. Follow service standards and privacy policy.
2. Keep the FAQ database up‑to‑date.
3. Escalate to human agents when needed.

## Example:
- User: "What is my order status?"
  - Reply: "Please wait, I am checking your order. Is your order number [order_id]?"
- User: "How do I return a product?"
  - Reply: "Our return policy is ... Have you submitted a return request?"

Plugin Configuration Guide

Adding Plugins

Directly add a plugin in the agent editor (click the plus icon or let the model recommend one).

Invoke a plugin node within a workflow (configure input/output parameters in the workflow editor).

Plugin Example: Toutiao Search

Details Icon : Shows plugin description and required input parameter q (string query).

Copy Name Icon : Copies the plugin name to the clipboard.

Bind Message Card Icon : Binds structured card data (title, content, image) to the plugin output for user‑friendly display.

Edit Parameters Icon : Reveals input q and output news list definitions; each list item includes title, time, etc.

Delete Icon : Removes the plugin from the agent.

Best‑Practice Tips for Plugins

Focus on Function : Each plugin should serve a single purpose; avoid mixing unrelated tasks.

Simplify Selection : Add only necessary plugins; excess plugins increase system complexity and may destabilize outputs.

Avoid Redundancy : Do not use multiple plugins that perform the same function (e.g., multiple translators).