续写模式 - 火山方舟文档

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-260910

API 支持：

核心配置#

方式一#

支持模型：除 deepseek-v4-1-flash-260910 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。

1
2
3
4
5messages = [
    {"role": "user", "content": "You are a calculator, please calculate: 1 + 1"},
    # 最后消息为模型消息，role 为 assistant
    {"role": "assistant", "content": "="}
]

方式二#

支持模型：deepseek-v4-1-flash-260910

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。

1
2
3
4messages = [
    {"role": "user", "content": "You are a calculator, please calculate: 1 + 1"},
    {"role": "assistant", "content": "=", "prefix": True}
]

模式对比#

普通模式续写模式

普通模式	续写模式
请求示例 Python 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 Python 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 Shell `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 Shell `2` 模型根据信息“=”续写，保持对应的格式和行文风格。

请求示例

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19import 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

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21import 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

11+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

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

普通模式	续写模式
请求示例 Python 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 Python 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。 Shell 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。 Shell "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." } ] }

请求示例

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19import 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

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21import 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。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29Here'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。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28"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。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32import 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。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44import os
# Install SDK: python -m pip install --upgrade arkruntime
from arkruntime import Ark

messages = [
    {"role": "user", "content": '''Please translate the following text into English: <TEXT_TO_BE_TRANSLATED> '''}, # Replace <TEXT_TO_BE_TRANSLATED> with your content
    {"role": "assistant", "content": ""}
]

# Replace with Model ID
ark_model="doubao-seed-2-1-pro-260915"

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(
    model=ark_model,
    messages=messages
)

loop_count = 0 # 初始化循环计数器
max_loops = 20  # 设置最大循环次数为20

# 触发因为输出长度限制，循环生成 assistant 回复
while completion.choices[0].finish_reason == "length":
    messages[-1]["content"] += completion.choices[0].message.content
    completion = client.chat.completions.create(
        model=ark_model,
        messages=messages
    )
    # 增加计数器
    loop_count += 1
    # 检查是否超过最大循环次数
    if loop_count >= max_loops:
        print(f"Warning: Reached maximum loop count {max_loops}, stopping content generation.")
        break

# 返回最后一条回复
messages[-1]["content"] += completion.choices[0].message.content
print(messages[-1]["content"])

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。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23import os
# Install SDK: python -m pip install --upgrade 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'),
)
# Replace with Model ID
ark_model="doubao-seed-2-1-pro-260915"

messages = [
    {"role": "system", "content": "The following is a scene from Journey to the West. Please converse according to the specified role. For each round of dialogue, only reply with the content of the current role; do not add any additional content.\n\nAfter leaving Wuzhuang Guan, Tang Seng and his disciples traveled for more than a month and arrived at the foot of a great mountain. Wukong looked around and saw lofty mountains, steep ridges, overgrown weeds, and extremely dangerous terrain."},
    {"role": "assistant", "content": "Wukong said:"}
]

completion = client.chat.completions.create(
    model=ark_model,
    messages=messages
)
print(messages[-1]['content']+completion.choices[0].message.content)

模型输出示例：

Wukong said, "Master, this mountain looks very dangerous. I will go explore the way first. You wait here."

您也可以通过变更messages最后消息，实现角色切换。

1
2
3
4
5
6
7...
messages[-1] = {"role": "assistant", "content": "Tang Seng said:"}
completion = client.chat.completions.create(
    model=ark_model,
    messages=messages
)
print(messages[-1]['content']+completion.choices[0].message.content)

模型输出示例：

Tang Seng said: "Wukong, this mountain looks very dangerous, you must be careful to check it out, don't fall into the trap of the monster."