19904408334
9d9bd71468
- 流式调用增加可重试错误码检测与指数退避重试 - 业务字段写入改为前置追加，不覆盖已有值 - 请求体按模板元数据递归合并，补充数组/对象字段处理 - 异步任务查询支持请求体映射与占位符替换 - 升级 common 与 gmq 依赖版本，移除本地 replace
167 lines
11 KiB
Go
167 lines
11 KiB
Go
package entity
|
||
|
||
import (
|
||
"model-gateway/consts/model"
|
||
|
||
"gitea.redpowerfuture.com/red-future/common/beans"
|
||
)
|
||
|
||
type modelManageCol struct {
|
||
beans.SQLBaseCol
|
||
ModelSupplier string
|
||
ModelName string
|
||
ModelType string
|
||
BaseURL string
|
||
SystemModel string
|
||
HttpMethod string
|
||
ChatModel string
|
||
ResponseType string
|
||
ApiKey string
|
||
Enabled string
|
||
RequestHeadMapping string
|
||
RequestBodyMapping string
|
||
RequestBusinessFieldMapping string
|
||
ResponseMapping string
|
||
ResponseBodyMapping string
|
||
ResponseBusinessFieldMapping string
|
||
MaxConcurrency string
|
||
TokenPredictPrice string
|
||
TokenPredictPriceUnit string
|
||
PriceConfig string
|
||
MaxTokens string
|
||
MinDuration string
|
||
MaxDuration string
|
||
LastFrame string
|
||
}
|
||
|
||
var ModelManageCol = modelManageCol{
|
||
SQLBaseCol: beans.DefSQLBaseCol,
|
||
ModelSupplier: "model_supplier",
|
||
ModelName: "model_name",
|
||
ModelType: "model_type",
|
||
BaseURL: "base_url",
|
||
SystemModel: "system_model",
|
||
HttpMethod: "http_method",
|
||
ChatModel: "chat_model",
|
||
ResponseType: "response_type",
|
||
ApiKey: "api_key",
|
||
Enabled: "enabled",
|
||
RequestHeadMapping: "request_head_mapping",
|
||
RequestBodyMapping: "request_body_mapping",
|
||
RequestBusinessFieldMapping: "request_business_field_mapping",
|
||
ResponseMapping: "response_mapping",
|
||
ResponseBodyMapping: "response_body_mapping",
|
||
ResponseBusinessFieldMapping: "response_business_field_mapping",
|
||
MaxConcurrency: "max_concurrency",
|
||
TokenPredictPrice: "token_predict_price",
|
||
TokenPredictPriceUnit: "token_predict_price_unit",
|
||
PriceConfig: "price_config",
|
||
MaxTokens: "max_tokens",
|
||
MinDuration: "min_duration",
|
||
MaxDuration: "max_duration",
|
||
LastFrame: "last_frame",
|
||
}
|
||
|
||
type ModelManage struct {
|
||
beans.SQLBaseDO `orm:",inline"`
|
||
ModelSupplier model.SupplierType `orm:"model_supplier" json:"modelSupplier" description:"模型供应商"`
|
||
ModelName string `orm:"model_name" json:"modelName" description:"模型名称"`
|
||
ModelType model.ModelType `orm:"model_type" json:"modelType" description:"模型类型"`
|
||
BaseURL string `orm:"base_url" json:"baseUrl" description:"模型地址"`
|
||
SystemModel *bool `orm:"system_model" json:"systemModel" description:"系统模型"`
|
||
HttpMethod string `orm:"http_method" json:"httpMethod" description:"http方法"`
|
||
ChatModel *bool `orm:"chat_model" json:"chatModel" description:"是否聊天模型"`
|
||
ResponseType model.ResponseType `orm:"response_type" json:"responseType" description:"返回类型：1同步，2异步，3流"`
|
||
ApiKey string `orm:"api_key" json:"apiKey" description:"api key"`
|
||
Enabled *bool `orm:"enabled" json:"enabled" description:"是否启用"`
|
||
RequestHeadMapping map[string]string `orm:"request_head_mapping" json:"requestHeadMapping" description:"请求头映射"`
|
||
RequestBodyMapping map[string]any `orm:"request_body_mapping" json:"requestBodyMapping" description:"请求体映射"`
|
||
RequestBusinessFieldMapping map[string]string `orm:"request_business_field_mapping" json:"requestBusinessFieldMapping" description:"请求业务字段映射"`
|
||
ResponseMapping map[string]any `orm:"response_mapping" json:"responseMapping" description:"响应映射"`
|
||
ResponseBodyMapping map[string]string `orm:"response_body_mapping" json:"responseBodyMapping" description:"响应主体映射"`
|
||
ResponseBusinessFieldMapping map[string]string `orm:"response_business_field_mapping" json:"responseBusinessFieldMapping" description:"响应业务字段映射"`
|
||
MaxConcurrency int `orm:"max_concurrency" json:"maxConcurrency" description:"最大并发数"`
|
||
TokenMapping *TokenMapping `orm:"token_mapping" json:"tokenMapping" description:"token映射"`
|
||
AsyncTaskMapping *AsyncTaskMapping `orm:"async_task_mapping" json:"asyncTaskMapping" description:"异步任务映射"`
|
||
TokenPredictPrice float64 `orm:"token_predict_price" json:"tokenPredictPrice" description:"模型Token预估价格"`
|
||
TokenPredictPriceUnit string `orm:"token_predict_price_unit" json:"tokenPredictPriceUnit" description:"模型token预估价格单位（秒，百万Token，千Token，字数）"`
|
||
PriceConfig *PriceConfig `orm:"price_config" json:"priceConfig" description:"计费规则"`
|
||
MaxTokens int `orm:"max_tokens" json:"maxTokens" description:"最大token数"`
|
||
MinDuration int `orm:"min_duration" json:"minDuration" description:"最小时长（秒）"`
|
||
MaxDuration int `orm:"max_duration" json:"maxDuration" description:"最大时长（秒）"`
|
||
LastFrame string `orm:"last_frame" json:"lastFrame" description:"视频的尾帧图像"`
|
||
ErrorMessageMapping map[string]any `orm:"error_message_mapping" json:"errorMessageMapping" description:"错误消息映射"`
|
||
}
|
||
|
||
type TokenMapping struct {
|
||
PromptTokens string `json:"promptTokens" dc:"输入token"`
|
||
CompletionTokens string `json:"completionTokens" dc:"输出token"`
|
||
TotalTokens string `json:"totalTokens" dc:"总token"`
|
||
}
|
||
|
||
type AsyncTaskMapping struct {
|
||
Url string `json:"url" dc:"url"`
|
||
HttpMethod string `json:"httpMethod" dc:"http方法" d:"POST"`
|
||
RequestBodyMapping map[string]any `json:"requestBodyMapping" description:"请求体映射"`
|
||
RequestHeadMapping map[string]string `json:"requestHeadMapping" description:"请求头映射"`
|
||
ResponseMapping map[string]any `json:"responseMapping" description:"响应映射"`
|
||
TaskId string `json:"taskId" dc:"任务id"`
|
||
TaskStatus string `json:"taskStatus" dc:"任务状态"`
|
||
TaskStatusPending string `json:"taskStatusPending" dc:"任务状态-待处理"`
|
||
TaskStatusRunning string `json:"taskStatusRunning" dc:"任务状态-运行中"`
|
||
TaskStatusSuccess string `json:"taskStatusSuccess" dc:"任务状态-成功"`
|
||
TaskStatusFailed string `json:"taskStatusFailed" dc:"任务状态-失败"`
|
||
TaskStatusCancel string `json:"taskStatusCancel" dc:"任务状态-取消"`
|
||
TaskStatusUnknown string `json:"taskStatusUnknown" dc:"任务状态-未知"`
|
||
}
|
||
|
||
// PriceConfig 模型计费规则（price_config 列，JSONB）。
|
||
// 命中条件为固定字段结构（PriceMatch）：token 档位从调用用量读取，媒体类型由请求体参考媒体字段推导，
|
||
// 全部字段缺省表示无条件命中。
|
||
type PriceConfig struct {
|
||
Currency string `json:"currency" dc:"币种，默认CNY"`
|
||
Unit string `json:"unit" dc:"单价基准：per_1K-千token/per_1M-百万token/per_1-单个"`
|
||
Rules []PriceRule `json:"rules" dc:"定价规则数组，按序首条命中生效"`
|
||
Discount *PriceDiscount `json:"discount" dc:"模型级限时折扣，规则级可覆盖"`
|
||
|
||
// 单规则便捷字段：Rules 为空时的兜底价（等价于一条空 match 规则），全部为 0 时不参与计费
|
||
InputPrice float64 `json:"inputPrice,omitempty" dc:"输入单价"`
|
||
OutputPrice float64 `json:"outputPrice,omitempty" dc:"输出单价"`
|
||
CacheHitPrice float64 `json:"cacheHitPrice,omitempty" dc:"缓存命中单价"`
|
||
CacheStorageHourPrice float64 `json:"cacheStorageHourPrice,omitempty" dc:"缓存存储单价（元/小时）"`
|
||
}
|
||
|
||
// PriceRule 定价规则：match 条件命中后按价格项计价，缺省项视为 0
|
||
type PriceRule struct {
|
||
Name string `json:"name"`
|
||
Match *PriceMatch `json:"match,omitempty" dc:"命中条件（固定字段，见 PriceMatch）；空表示任意调用"`
|
||
Input float64 `json:"input" dc:"输入单价（非音频）"`
|
||
InputAudio float64 `json:"inputAudio,omitempty" dc:"输入单价（音频），请求体 reference_audio 参考媒体字段命中时生效（见 DetectMediaType）"`
|
||
Output float64 `json:"output" dc:"输出单价"`
|
||
CacheHit float64 `json:"cacheHit" dc:"缓存命中单价（非音频）"`
|
||
CacheHitAudio float64 `json:"cacheHitAudio,omitempty" dc:"缓存命中单价（音频），请求体 reference_audio 参考媒体字段命中时生效（见 DetectMediaType）"`
|
||
CacheStorageHour float64 `json:"cacheStorageHour" dc:"缓存存储单价（元/小时）"`
|
||
Discount *PriceDiscount `json:"discount" dc:"规则级折扣，覆盖模型级折扣"`
|
||
}
|
||
|
||
// PriceMatch 命中条件：字段对应调用上下文固定路径，全部缺省（空值）表示无条件命中任意调用。
|
||
// token 档位（InputLength/OutputLength/TotalLength/CachedTokens）从调用用量读取，
|
||
// MediaType 由请求体参考媒体字段推导（见 DetectMediaType）。
|
||
type PriceMatch struct {
|
||
MediaType string `json:"mediaType,omitempty" dc:"输入媒体类型精确值（audio/no_video/has_video，由请求体参考媒体字段推导）"`
|
||
InputLengthMax int64 `json:"inputLengthMax,omitempty" dc:"输入token上限（<=，usage.prompt_tokens）"`
|
||
InputLengthMin int64 `json:"inputLengthMin,omitempty" dc:"输入token下限（>=，usage.prompt_tokens）"`
|
||
OutputLengthMax int64 `json:"outputLengthMax,omitempty" dc:"输出token上限（<=，usage.completion_tokens）"`
|
||
OutputLengthMin int64 `json:"outputLengthMin,omitempty" dc:"输出token下限（>=，usage.completion_tokens）"`
|
||
TotalLengthMax int64 `json:"totalLengthMax,omitempty" dc:"总token上限（<=，usage.total_tokens）"`
|
||
TotalLengthMin int64 `json:"totalLengthMin,omitempty" dc:"总token下限（>=，usage.total_tokens）"`
|
||
CachedTokensMax int64 `json:"cachedTokensMax,omitempty" dc:"缓存命中token上限（<=，usage.cached_tokens）"`
|
||
CachedTokensMin int64 `json:"cachedTokensMin,omitempty" dc:"缓存命中token下限（>=，usage.cached_tokens）"`
|
||
}
|
||
|
||
// PriceDiscount 限时折扣：rate 为折扣率（0.4=4折），effective 为空数组表示长期有效
|
||
type PriceDiscount struct {
|
||
Rate float64 `json:"rate" dc:"折扣率"`
|
||
Effective [2]string `json:"effective" dc:"有效期[起始,截止]，格式YYYY-MM-DD，缺省长期有效"`
|
||
}
|