feat: support third part models of baidu (#1046)

百度千帆平台上的第三方大模型调用

README.md

@@ -67,6 +67,7 @@ _✨ 通过标准的 OpenAI API 格式访问所有的大模型，开箱即用
    + [x] [OpenAI ChatGPT 系列模型](https://platform.openai.com/docs/guides/gpt/chat-completions-api)（支持 [Azure OpenAI API](https://learn.microsoft.com/en-us/azure/ai-services/openai/reference)）
    + [x] [Anthropic Claude 系列模型](https://anthropic.com)
    + [x] [Google PaLM2/Gemini 系列模型](https://developers.generativeai.google)
+   + [x] [Mistral 系列模型](https://mistral.ai/)
    + [x] [百度文心一言系列模型](https://cloud.baidu.com/doc/WENXINWORKSHOP/index.html)
    + [x] [阿里通义千问系列模型](https://help.aliyun.com/document_detail/2400395.html)
    + [x] [讯飞星火认知大模型](https://www.xfyun.cn/doc/spark/Web.html)

common/constants.go

@@ -66,6 +66,7 @@ const (
 	ChannelTypeMoonshot       = 25
 	ChannelTypeBaichuan       = 26
 	ChannelTypeMinimax        = 27
+	ChannelTypeMistral        = 28
 )
 
 var ChannelBaseURLs = []string{
@@ -97,6 +98,7 @@ var ChannelBaseURLs = []string{
 	"https://api.moonshot.cn",                   // 25
 	"https://api.baichuan-ai.com",               // 26
 	"https://api.minimax.chat",                  // 27
+	"https://api.mistral.ai",                    // 28
 }
 
 const (

common/model-ratio.go

@@ -17,7 +17,6 @@ const (
 // https://platform.openai.com/docs/models/model-endpoint-compatibility
 // https://cloud.baidu.com/doc/WENXINWORKSHOP/s/Blfmc9dlf
 // https://openai.com/pricing
-// TODO: when a new api is enabled, check the pricing here
 // 1 === $0.002 / 1K tokens
 // 1 === ￥0.014 / 1k tokens
 var ModelRatio = map[string]float64{
@@ -116,15 +115,29 @@ var ModelRatio = map[string]float64{
 	"abab6-chat":    0.1 * RMB,
 	"abab5.5-chat":  0.015 * RMB,
 	"abab5.5s-chat": 0.005 * RMB,
+	// https://docs.mistral.ai/platform/pricing/
+	"open-mistral-7b":       0.25 / 1000 * USD,
+	"open-mixtral-8x7b":     0.7 / 1000 * USD,
+	"mistral-small-latest":  2.0 / 1000 * USD,
+	"mistral-medium-latest": 2.7 / 1000 * USD,
+	"mistral-large-latest":  8.0 / 1000 * USD,
+	"mistral-embed":         0.1 / 1000 * USD,
 }
 
+var CompletionRatio = map[string]float64{}
+
 var DefaultModelRatio map[string]float64
+var DefaultCompletionRatio map[string]float64
 
 func init() {
 	DefaultModelRatio = make(map[string]float64)
 	for k, v := range ModelRatio {
 		DefaultModelRatio[k] = v
 	}
+	DefaultCompletionRatio = make(map[string]float64)
+	for k, v := range CompletionRatio {
+		DefaultCompletionRatio[k] = v
+	}
 }
 
 func ModelRatio2JSONString() string {
@@ -155,8 +168,6 @@ func GetModelRatio(name string) float64 {
 	return ratio
 }
 
-var CompletionRatio = map[string]float64{}
-
 func CompletionRatio2JSONString() string {
 	jsonBytes, err := json.Marshal(CompletionRatio)
 	if err != nil {
@@ -174,6 +185,9 @@ func GetCompletionRatio(name string) float64 {
 	if ratio, ok := CompletionRatio[name]; ok {
 		return ratio
 	}
+	if ratio, ok := DefaultCompletionRatio[name]; ok {
+		return ratio
+	}
 	if strings.HasPrefix(name, "gpt-3.5") {
 		if strings.HasSuffix(name, "0125") {
 			// https://openai.com/blog/new-embedding-models-and-api-updates
@@ -206,5 +220,8 @@ func GetCompletionRatio(name string) float64 {
 	if strings.HasPrefix(name, "claude-2") {
 		return 2.965517
 	}
+	if strings.HasPrefix(name, "mistral-") {
+		return 3
+	}
 	return 1
 }

common/random.go

@@ -0,0 +1,8 @@
+package common
+
+import "math/rand"
+
+// RandRange returns a random number between min and max (max is not included)
+func RandRange(min, max int) int {
+	return min + rand.Intn(max-min)
+}

controller/channel-test.go

@@ -8,6 +8,7 @@ import (
 	"github.com/songquanpeng/one-api/common"
 	"github.com/songquanpeng/one-api/common/config"
 	"github.com/songquanpeng/one-api/common/logger"
+	"github.com/songquanpeng/one-api/middleware"
 	"github.com/songquanpeng/one-api/model"
 	"github.com/songquanpeng/one-api/relay/constant"
 	"github.com/songquanpeng/one-api/relay/helper"
@@ -18,6 +19,7 @@ import (
 	"net/http/httptest"
 	"net/url"
 	"strconv"
+	"strings"
 	"sync"
 	"time"
 
@@ -51,6 +53,7 @@ func testChannel(channel *model.Channel) (err error, openaiErr *relaymodel.Error
 	c.Request.Header.Set("Content-Type", "application/json")
 	c.Set("channel", channel.Type)
 	c.Set("base_url", channel.GetBaseURL())
+	middleware.SetupContextForSelectedChannel(c, channel, "")
 	meta := util.GetRelayMeta(c)
 	apiType := constant.ChannelType2APIType(channel.Type)
 	adaptor := helper.GetAdaptor(apiType)
@@ -59,6 +62,12 @@ func testChannel(channel *model.Channel) (err error, openaiErr *relaymodel.Error
 	}
 	adaptor.Init(meta)
 	modelName := adaptor.GetModelList()[0]
+	if !strings.Contains(channel.Models, modelName) {
+		modelNames := strings.Split(channel.Models, ",")
+		if len(modelNames) > 0 {
+			modelName = modelNames[0]
+		}
+	}
 	request := buildTestRequest()
 	request.Model = modelName
 	meta.OriginModelName, meta.ActualModelName = modelName, modelName

controller/model.go

@@ -6,6 +6,7 @@ import (
 	"github.com/songquanpeng/one-api/relay/channel/ai360"
 	"github.com/songquanpeng/one-api/relay/channel/baichuan"
 	"github.com/songquanpeng/one-api/relay/channel/minimax"
+	"github.com/songquanpeng/one-api/relay/channel/mistral"
 	"github.com/songquanpeng/one-api/relay/channel/moonshot"
 	"github.com/songquanpeng/one-api/relay/constant"
 	"github.com/songquanpeng/one-api/relay/helper"
@@ -122,6 +123,17 @@ func init() {
 			Parent:     nil,
 		})
 	}
+	for _, modelName := range mistral.ModelList {
+		openAIModels = append(openAIModels, OpenAIModels{
+			Id:         modelName,
+			Object:     "model",
+			Created:    1626777600,
+			OwnedBy:    "mistralai",
+			Permission: permission,
+			Root:       modelName,
+			Parent:     nil,
+		})
+	}
 	openAIModelsMap = make(map[string]OpenAIModels)
 	for _, model := range openAIModels {
 		openAIModelsMap[model.Id] = model

controller/relay.go

@@ -62,7 +62,7 @@ func Relay(c *gin.Context) {
 		retryTimes = 0
 	}
 	for i := retryTimes; i > 0; i-- {
-		channel, err := dbmodel.CacheGetRandomSatisfiedChannel(group, originalModel)
+		channel, err := dbmodel.CacheGetRandomSatisfiedChannel(group, originalModel, i != retryTimes)
 		if err != nil {
 			logger.Errorf(ctx, "CacheGetRandomSatisfiedChannel failed: %w", err)
 			break

middleware/distributor.go

@@ -68,7 +68,7 @@ func Distribute() func(c *gin.Context) {
 				}
 			}
 			requestModel = modelRequest.Model
-			channel, err = model.CacheGetRandomSatisfiedChannel(userGroup, modelRequest.Model)
+			channel, err = model.CacheGetRandomSatisfiedChannel(userGroup, modelRequest.Model, false)
 			if err != nil {
 				message := fmt.Sprintf("当前分组 %s 下对于模型 %s 无可用渠道", userGroup, modelRequest.Model)
 				if channel != nil {

model/cache.go

@@ -191,7 +191,7 @@ func SyncChannelCache(frequency int) {
 	}
 }
 
-func CacheGetRandomSatisfiedChannel(group string, model string) (*Channel, error) {
+func CacheGetRandomSatisfiedChannel(group string, model string, ignoreFirstPriority bool) (*Channel, error) {
 	if !config.MemoryCacheEnabled {
 		return GetRandomSatisfiedChannel(group, model)
 	}
@@ -213,5 +213,10 @@ func CacheGetRandomSatisfiedChannel(group string, model string) (*Channel, error
 		}
 	}
 	idx := rand.Intn(endIdx)
+	if ignoreFirstPriority {
+		if endIdx < len(channels) { // which means there are more than one priority
+			idx = common.RandRange(endIdx, len(channels))
+		}
+	}
 	return channels[idx], nil
 }

relay/channel/ali/main.go

@@ -33,6 +33,9 @@ func ConvertRequest(request model.GeneralOpenAIRequest) *ChatRequest {
 		enableSearch = true
 		aliModel = strings.TrimSuffix(aliModel, EnableSearchModelSuffix)
 	}
+	if request.TopP >= 1 {
+		request.TopP = 0.9999
+	}
 	return &ChatRequest{
 		Model: aliModel,
 		Input: Input{
@@ -42,6 +45,9 @@ func ConvertRequest(request model.GeneralOpenAIRequest) *ChatRequest {
 			EnableSearch:      enableSearch,
 			IncrementalOutput: request.Stream,
 			Seed:              uint64(request.Seed),
+			MaxTokens:         request.MaxTokens,
+			Temperature:       request.Temperature,
+			TopP:              request.TopP,
 		},
 	}
 }

relay/channel/ali/model.go

@@ -16,6 +16,8 @@ type Parameters struct {
 	Seed              uint64  `json:"seed,omitempty"`
 	EnableSearch      bool    `json:"enable_search,omitempty"`
 	IncrementalOutput bool    `json:"incremental_output,omitempty"`
+	MaxTokens         int     `json:"max_tokens,omitempty"`
+	Temperature       float64 `json:"temperature,omitempty"`
 }
 
 type ChatRequest struct {

relay/channel/baidu/adaptor.go

@@ -36,6 +36,8 @@ func (a *Adaptor) GetRequestURL(meta *util.RelayMeta) (string, error) {
 		fullRequestURL = "https://aip.baidubce.com/rpc/2.0/ai_custom/v1/wenxinworkshop/chat/bloomz_7b1"
 	case "Embedding-V1":
 		fullRequestURL = "https://aip.baidubce.com/rpc/2.0/ai_custom/v1/wenxinworkshop/embeddings/embedding-v1"
+	default:
+               fullRequestURL = "https://aip.baidubce.com/rpc/2.0/ai_custom/v1/wenxinworkshop/chat/" + meta.ActualModelName
 	}
 	var accessToken string
 	var err error

relay/channel/gemini/constants.go

@@ -1,6 +1,6 @@
 package gemini
 
 var ModelList = []string{
-	"gemini-pro",
-	"gemini-pro-vision",
+	"gemini-pro", "gemini-1.0-pro-001",
+	"gemini-pro-vision", "gemini-1.0-pro-vision-001",
 }

relay/channel/mistral/constants.go

@@ -0,0 +1,10 @@
+package mistral
+
+var ModelList = []string{
+	"open-mistral-7b",
+	"open-mixtral-8x7b",
+	"mistral-small-latest",
+	"mistral-medium-latest",
+	"mistral-large-latest",
+	"mistral-embed",
+}

relay/channel/openai/adaptor.go

@@ -9,6 +9,7 @@ import (
 	"github.com/songquanpeng/one-api/relay/channel/ai360"
 	"github.com/songquanpeng/one-api/relay/channel/baichuan"
 	"github.com/songquanpeng/one-api/relay/channel/minimax"
+	"github.com/songquanpeng/one-api/relay/channel/mistral"
 	"github.com/songquanpeng/one-api/relay/channel/moonshot"
 	"github.com/songquanpeng/one-api/relay/model"
 	"github.com/songquanpeng/one-api/relay/util"
@@ -94,6 +95,8 @@ func (a *Adaptor) GetModelList() []string {
 		return baichuan.ModelList
 	case common.ChannelTypeMinimax:
 		return minimax.ModelList
+	case common.ChannelTypeMistral:
+		return mistral.ModelList
 	default:
 		return ModelList
 	}
@@ -111,6 +114,8 @@ func (a *Adaptor) GetChannelName() string {
 		return "baichuan"
 	case common.ChannelTypeMinimax:
 		return "minimax"
+	case common.ChannelTypeMistral:
+		return "mistralai"
 	default:
 		return "openai"
 	}

relay/controller/helper.go

@@ -172,10 +172,8 @@ func postConsumeQuota(ctx context.Context, usage *relaymodel.Usage, meta *util.R
 	if err != nil {
 		logger.Error(ctx, "error update user quota cache: "+err.Error())
 	}
-	if quota != 0 {
-		logContent := fmt.Sprintf("模型倍率 %.2f，分组倍率 %.2f，补全倍率 %.2f", modelRatio, groupRatio, completionRatio)
-		model.RecordConsumeLog(ctx, meta.UserId, meta.ChannelId, promptTokens, completionTokens, textRequest.Model, meta.TokenName, quota, logContent)
-		model.UpdateUserUsedQuotaAndRequestCount(meta.UserId, quota)
-		model.UpdateChannelUsedQuota(meta.ChannelId, quota)
-	}
+	logContent := fmt.Sprintf("模型倍率 %.2f，分组倍率 %.2f，补全倍率 %.2f", modelRatio, groupRatio, completionRatio)
+	model.RecordConsumeLog(ctx, meta.UserId, meta.ChannelId, promptTokens, completionTokens, textRequest.Model, meta.TokenName, quota, logContent)
+	model.UpdateUserUsedQuotaAndRequestCount(meta.UserId, quota)
+	model.UpdateChannelUsedQuota(meta.ChannelId, quota)
 }

web/berry/src/constants/ChannelConstants.js

@@ -29,6 +29,12 @@ export const CHANNEL_OPTIONS = {
     value: 24,
     color: 'orange'
   },
+  28: {
+    key: 28,
+    text: 'Mistral AI',
+    value: 28,
+    color: 'orange'
+  },
   15: {
     key: 15,
     text: '百度文心千帆',

web/default/src/constants/channel.constants.js

@@ -4,6 +4,7 @@ export const CHANNEL_OPTIONS = [
   { key: 3, text: 'Azure OpenAI', value: 3, color: 'olive' },
   { key: 11, text: 'Google PaLM2', value: 11, color: 'orange' },
   { key: 24, text: 'Google Gemini', value: 24, color: 'orange' },
+  { key: 28, text: 'Mistral AI', value: 28, color: 'orange' },
   { key: 15, text: '百度文心千帆', value: 15, color: 'blue' },
   { key: 17, text: '阿里通义千问', value: 17, color: 'orange' },
   { key: 18, text: '讯飞星火认知', value: 18, color: 'blue' },