文档中心

火山方舟

进阶使用

续写模式

调用模型，如果希望控制和引导模型的输出，您可以通过续写模式实现：预填assistant角色的部分内容，模型将基于该内容延续输出，以此达成输出控制目标。您可以将该输出控制机制应用于以下场景：强制输出 JSON/XML 等特定格式、绕过模型最大输出限制生成超长内容、确保大模型在角色扮演场景中保持角色一致性。

模型及 API

支持的模型如下：

文本生成模型中无深度思考能力的模型：具体模型可在文本生成能力筛选。

部分具有深度思考能力的模型：

doubao-seed-evolving

doubao-seed-2-1-pro-260915

doubao-seed-2-1-lite-260915

doubao-seed-2-1-pro-260628

doubao-seed-2-1-turbo-260628

doubao-seed-2-0-lite-260428

doubao-seed-2-0-mini-260428

doubao-seed-2-0-pro-260215

doubao-seed-2-0-lite-260215

doubao-seed-2-0-mini-260215

doubao-seed-2-0-code-preview-260215

doubao-seed-character-260628

glm-5-3-flash-260828

deepseek-v4-1-flash-261017

deepseek-v4-1-flash-260910

deepseek-v4-1-flash-economy-261017

kimi-k3-1-pro-261024

注意

其他深度思考模型不可使用续写模式，即请求时 message 以 assistant role 消息结尾时，thinking 字段会失效，模型保持默认是否深度思考的行为。

API 支持：

Chat API

Responses API

核心配置

方式一

支持模型：除 deepseek-v4-1-flash-261017、deepseek-v4-1-flash-260910、deepseek-v4-1-flash-economy-261017、kimi-k3-1-pro-261024 之外的其他模型。

请将消息列表的最后一条消息的 role 设置为assistant，模型会遵循 assistant 消息的已有格式和内容来续写。

方式二

支持模型：deepseek-v4-1-flash-261017、deepseek-v4-1-flash-260910、deepseek-v4-1-flash-economy-261017、kimi-k3-1-pro-261024

请将消息列表的最后一条消息的 role 设置为assistant，并设置最后一条消息的 prefix 参数为true。模型会遵循 assistant 消息的已有格式和内容来续写。

模式对比

普通模式	续写模式
请求示例 import os # Install SDK: pip install arkruntime from arkruntime import Ark client = Ark( # The base URL for model invocation base_url="https://ark.cn-beijing.volces.com/api/v3", # Get API Key: https://ark.volcengine.com/region:cn-beijing/apikey api_key=os.getenv('ARK_API_KEY'), ) completion = client.chat.completions.create( # Replace with Model ID model = "doubao-seed-2-1-pro-260915", messages=[ {"role": "user", "content": "You are a calculator, please calculate: 1 + 1 "}, ] ) print(completion.choices[0].message.content)	请求示例 import os # Install SDK: pip install arkruntime from arkruntime import Ark client = Ark( # The base URL for model invocation base_url="https://ark.cn-beijing.volces.com/api/v3", # Get API Key: https://ark.volcengine.com/region:cn-beijing/apikey api_key=os.getenv('ARK_API_KEY'), ) completion = client.chat.completions.create( # Replace with Model ID model = "doubao-seed-2-1-pro-260915", messages=[ {"role": "user", "content": "You are a calculator, please calculate: 1 + 1"}, # 额外传入模型信息,模型据此续写 {"role": "assistant", "content": "="} ] ) print(completion.choices[0].message.content)
返回示例 1+1等于2。输出会按照模型根据问题的完整自然语言回复。	返回示例 2 模型根据信息“=”续写，保持对应的格式和行文风格。

接下来，对于几种典型使用场景进行举例说明。

场景：改善输出格式

模型可能会基于自身理解输出内容，导致结果无法直接被其他程序解析。可以通过预填充 { 符号，引导模型跳过一些场景回复，直接输出 JSON 对象，会让回答更加简洁和工整，可以被其他程序更好地解析。

普通模式	续写模式
请求示例 import os # Install SDK: pip install arkruntime from arkruntime import Ark client = Ark( # The base URL for model invocation base_url="https://ark.cn-beijing.volces.com/api/v3", # Get API Key: https://ark.volcengine.com/region:cn-beijing/apikey api_key=os.getenv('ARK_API_KEY'), ) completion = client.chat.completions.create( # Replace with Model ID model = "doubao-seed-2-1-pro-260915", messages=[ {"role": "user", "content": "Use JSON to describe the name and function of the Doubao model"} ] ) print(completion.choices[0].message.content)	请求示例 import os # Install SDK: pip install arkruntime from arkruntime import Ark client = Ark( # The base URL for model invocation base_url="https://ark.cn-beijing.volces.com/api/v3", # Get API Key: https://ark.volcengine.com/region:cn-beijing/apikey api_key=os.getenv('ARK_API_KEY'), ) completion = client.chat.completions.create( # Replace with Model ID model = "doubao-seed-2-1-pro-260915", messages=[ {"role": "user", "content": "Use JSON to describe the name and function of the Doubao model"}, # 额外传入一个 { 的模型信息,模型据此续写 {"role": "assistant", "content": "{"}, ] ) print(completion.choices[0].message.content)
返回示例返回一些说明性语言，无法直接按照 JSON 解析。 Here's a JSON description of the Doubao model, including its name and function: ```json { "model": { "name": "Doubao", "alternative_names": ["doubao"], "type": "large_language_model", "developer": "ByteDance", "primary_function": { "description": "Doubao is an AI assistant designed to understand and generate human-like text responses", "capabilities": [ "natural language understanding", "text generation", "question answering", "translation between languages", "text summarization", "creative writing", "code generation and explanation", "general knowledge provision" ] }, "languages_supported": ["Chinese", "English"], "access_method": "API and web/mobile interfaces", "version": "1.0", "release_date": "2023" } } Note: Some details about Doubao might not be publicly available, as it's a relatively new model from ByteDance. The information here represents common capabilities of similar large language models.	返回示例和前置assistant信息组合成 JSON 返回。 "model_name": "Doubao", "functions": [ { "name": "text_generation", "description": "Generates human-like text based on input prompts, supporting creative writing, summarization, and more." }, { "name": "language_translation", "description": "Translates text between multiple languages with high accuracy and contextual understanding." }, { "name": "sentiment_analysis", "description": "Analyzes the sentiment of input text, determining whether it is positive, negative, or neutral." }, { "name": "question_answering", "description": "Provides accurate answers to questions based on the input context or general knowledge." }, { "name": "code_generation", "description": "Generates code snippets in various programming languages based on natural language descriptions." }, { "name": "conversational_ai", "description": "Engages in natural, context-aware conversations with users, simulating human-like interactions." } ] }

注意事项

方舟已提供结构化输出能力，可让模型输出严格符合格式要求的 JSON 内容，请参见结构化输出(beta)。

由于模型具有一定的随机性，无法 100% 保证回复为标准 JSON 对象。建议您对模型的回复进行校验，下面示例使用 json_repair 库校验返回的是否为合法的 JSON 对象。

import json
# Install json-repair: pip install json-repair
import json_repair
import os
# Install SDK: python -m pip install --upgrade arkruntime
from arkruntime import Ark

PREFILL_PREFIX = "{"
client = Ark(
# The base URL for model invocation
base_url="https://ark.cn-beijing.volces.com/api/v3",
# Get API Key: https://ark.volcengine.com/region:cn-beijing/apikey
api_key=os.getenv('ARK_API_KEY'),
)
completion = client.chat.completions.create(
# Replace with Model ID
model = "doubao-seed-2-1-pro-260915",
messages=[
{"role": "user", "content": "Use JSON to describe the name and function of the Doubao model"},
{"role": "assistant", "content": PREFILL_PREFIX},
]
)
# 拼接 PREFILL_PREFIX 和模型输出
json_string = PREFILL_PREFIX + completion.choices[0].message.content
# 解析 JSON Object
obj = {}
try:
obj = json.loads(json_string)
except json.JSONDecodeError as e:
obj = json_repair.loads(json_string)

print(obj)

场景：输出超过模型输出上限的内容

使用续写模式让大模型根据前轮请求续写内容，并拼接多次输出的内容，组合成超长输出。可以避免因为 max_tokens（最大输出长度）限制，导致无法获取完整内容。

这种方式会将模型输出多次返回给模型，会有更多 token 用量。

注意增加循环次数限制等兜底，防止死循环输出等特殊情况，导致意外花费。

场景：增强角色扮演一致性

我们通常使用系统消息来设置角色的一些信息，但是随着对话轮次的增多，模型可能会跳脱出角色的约束，通过续写模式，提醒模型本次输出扮演的角色，保证对话符合预期。

模型输出示例：

您也可以通过变更messages最后消息，实现角色切换。

模型输出示例：

最近更新时间：2026.10.24 19:43:57

这个页面对您有帮助吗？

有用

无用