- GRPO(group Relative Policy Optimization)分组奖励估计
- 冷启动采用高质量长思维链的人工生成数据初始化然后开始优化推理能力训练和偏好训练和SFT
kind: app
version: 0.3.0
app: #应用配置的容器节点
description: '应用描述(可选)'
icon: 🤖 #应用图标(可选)
icon_background: '#FFEAD5' #图标背景色(可选)
mode: agent-chat #应用模式(必填) 核心交互方式 advanced-chat workflow chatflow
name: xcode #应用名称(必填)
use_icon_as_answer_icon: false #是否将应用图标用作回答图标(可选) 默认false
dependencies: #依赖项
- current_identifier: null #当前依赖的唯一标识(可选) 区分不同版本的依赖
type: marketplace #依赖类型(必填) marketplace: 市场插件 custom: 自定义依赖
value:
marketplace_plugin_unique_identifier: langgenius/bedrock:0.0.24@d15b21a5f8833d2489cd336798188b929a9f86e1c05f2f4770c05fc3e1e991bf #依赖的具体值(必填) 自定义为api url
- current_identifier: null
type: marketplace
value:
marketplace_plugin_unique_identifier: langgenius/json_process:0.0.2@dde6d7b676ccdcea89206d29232181a840170c19277d3d978e27cd1e3c92c707
- current_identifier: null
type: marketplace
value:
marketplace_plugin_unique_identifier: junjiem/mcp_sse:0.2.1@53cc613667fcf91dd7208dd5f6d2c8df3c7ff0af8b79e8f3c0a430f1b39bda4c
- current_identifier: null
type: marketplace
value:
marketplace_plugin_unique_identifier: langgenius/regex:0.0.3@257eaab07b70ab1f77a881b870eefee93fc8fd0dd13350077410264f31695039
- current_identifier: null
type: marketplace
value:
marketplace_plugin_unique_identifier: yevanchen/markitdown:0.0.4@776b3e2e930e2ffd28a75bb20fecbe7a020849cf754f86e604acacf1258877f6
model_config:
agent_mode:
enabled: true
max_iteration: 5 #最大推理轮次
prompt: "coding assistant" #Agnet角色
strategy: react #function_call采用函数调用策略 react动态规划工具调用路径
tools: #可调用工具列表
- enabled: true
isDeleted: false
notAuthor: false
provider_id: code
provider_name: code
provider_type: builtin
tool_label: 代码解释器
tool_name: simple_code
tool_parameters:
code: ''
language: ''
- enabled: true
isDeleted: false
notAuthor: false
provider_id: junjiem/mcp_sse/mcp_sse
provider_name: junjiem/mcp_sse/mcp_sse
provider_type: builtin
tool_label: 获取 MCP 工具列表
tool_name: mcp_sse_list_tools
tool_parameters:
prompts_as_tools: ''
resources_as_tools: ''
servers_config: ''
- enabled: true
isDeleted: false
notAuthor: false
provider_id: junjiem/mcp_sse/mcp_sse
provider_name: junjiem/mcp_sse/mcp_sse
provider_type: builtin
tool_label: 调用 MCP 工具
tool_name: mcp_sse_call_tool
tool_parameters:
arguments: ''
prompts_as_tools: ''
resources_as_tools: ''
servers_config: ''
tool_name: ''
- enabled: true
isDeleted: false
notAuthor: false
provider_id: langgenius/json_process/json_process
provider_name: langgenius/json_process/json_process
provider_type: builtin
tool_label: JSON 解析
tool_name: parse
tool_parameters:
content: ''
ensure_ascii: ''
json_filter: ''
- enabled: true
isDeleted: false
notAuthor: false
provider_id: langgenius/json_process/json_process
provider_name: langgenius/json_process/json_process
provider_type: builtin
tool_label: JSON 删除
tool_name: json_delete
tool_parameters:
content: ''
ensure_ascii: ''
query: ''
- enabled: true
isDeleted: false
notAuthor: false
provider_id: langgenius/json_process/json_process
provider_name: langgenius/json_process/json_process
provider_type: builtin
tool_label: JSON 替换
tool_name: json_replace
tool_parameters:
content: ''
ensure_ascii: ''
query: ''
replace_model: ''
replace_pattern: ''
replace_value: ''
value_decode: ''
- enabled: true
isDeleted: false
notAuthor: false
provider_id: langgenius/json_process/json_process
provider_name: langgenius/json_process/json_process
provider_type: builtin
tool_label: JSON 插入
tool_name: json_insert
tool_parameters:
content: ''
create_path: ''
ensure_ascii: ''
new_value: ''
query: ''
value_decode: ''
- enabled: true
isDeleted: false
notAuthor: false
provider_id: yevanchen/markitdown/markitdown
provider_name: yevanchen/markitdown/markitdown
provider_type: builtin
tool_label: markitdown
tool_name: markitdown
tool_parameters:
files: ''
- enabled: true
isDeleted: false
notAuthor: false
provider_id: langgenius/regex/regex
provider_name: langgenius/regex/regex
provider_type: builtin
tool_label: 正则表达式内容提取
tool_name: regex_extract
tool_parameters:
content: ''
expression: ''
annotation_reply: #标注回复配置, score_threshold 相似度阈值 embedding_model 向量化模型 storage_mode 存储模式
enabled: false
chat_prompt_config: {} #对话提示词模板 role
completion_prompt_config: {} #定义生成式任务的提示模板和生成规则
dataset_configs: #配置数据集用于RAG
datasets:
datasets: []
reranking_enable: false
retrieval_model: multiple
top_k: 4
dataset_query_variable: '' #数据集查询的动态参数
external_data_tools: [] #定义外部数据工具集成
file_upload:
allowed_file_extensions:
- .JPG
- .JPEG
- .PNG
- .GIF
- .WEBP
- .SVG
- .MP4
- .MOV
- .MPEG
- .WEBM
allowed_file_types:
- image
allowed_file_upload_methods:
- remote_url
- local_file
enabled: true
image:
detail: high
enabled: true
number_limits: 6
transfer_methods:
- remote_url
- local_file
number_limits: 6
model:
completion_params:
cross-region: true # 启用跨区域推理(提升响应速度)
latest_two_messages_cache_checkpoint: false #缓存最近两次用户与系统的对话消息
max_tokens: 32768 # 支持长文本生成(如生成完整报告)
model_name: Claude 4.0 Opus
reasoning_budget: 128000 # 限制LLM推理过程的资源消耗
reasoning_type: false # 推理策略 none step_by_step critical_thinking multi_hop self_consistency
response_format: JSON # Text JSON Markdown HTML JSON Schema
stop: [] # 生成内容遇到什么字符时停止
system_cache_checkpoint: false # 缓存系统级的通用数据 如知识库索引、工具配置、用户偏好
temperature: 0.5 # 降低随机性,输出更稳定, 越小幻觉越少
top_k: 500
top_p: 0.99 # 优先选择高概率词, 提升准确性
mode: chat
name: anthropic claude
provider: langgenius/bedrock/bedrock
more_like_this: #否启用类似内容推荐功能
enabled: false
opening_statement: '' #定义对话开始时的开场白或初始提示
pre_prompt: '' #注入的固定提示 定义角色或全局约束
prompt_type: structured #定义提示模板类型,simple(基础) structured(结构化) custom(自定义）)
retriever_resource: #外部数据检索资源 知识库、数据库
enabled: true
sensitive_word_avoidance: #敏感词过滤
configs: []
enabled: false
type: ''
speech_to_text: #配置语音输入功能
enabled: false
suggested_questions: [] #预定义建议问题列表，引导用户输入
suggested_questions_after_answer: #在生成回答后追加建议问题
enabled: false
text_to_speech: #配置语音输出功能
enabled: false
language: ''
voice: ''
user_input_form: [] #自定义用户输入表单
