feat(model-router): add keepOriginalModelName option to preserve full model name (#3916)

Signed-off-by: Cai Rui <yangjuan.cr@alibaba-inc.com>

plugins/wasm-go/extensions/model-router/README.md

@@ -9,6 +9,7 @@
 | `addProviderHeader`  | string          | 选填                    | -                        | 从model参数中解析出的provider名字放到哪个请求header中 |
 | `modelToHeader`      | string          | 选填                    | -                        | 直接将model参数放到哪个请求header中                   |
 | `enableOnPathSuffix` | array of string | 选填                    | ["/completions","/embeddings","/images/generations","/audio/speech","/fine_tuning/jobs","/moderations","/image-synthesis","/video-synthesis","/rerank","/messages"] | 只对这些特定路径后缀的请求生效，可以配置为 "*" 以匹配所有路径 |
+| `keepOriginalModelName` | bool         | 选填                    | false                    | 配合 `addProviderHeader` 使用，设为 true 时仍提取 provider 写入 header，但不改写请求体中的 model 字段 |
 | `autoRouting`        | object          | 选填                    | -                        | 自动路由配置，详见下方说明                            |
 
 ### autoRouting 配置
@@ -113,6 +114,21 @@ x-higress-llm-provider: dashscope
 }
 ```
 
+### 保留原始模型名（keepOriginalModelName）
+
+当使用 AI 模型聚合平台（如百炼/DashScope）接入第三方厂商模型时，部分模型名称本身包含 `/`（如 `MiniMax/MiniMax-M2.7`），并非 `provider/model` 格式。此时配合 `addProviderHeader` 使用会导致请求体中的 model 字段被错误改写。
+
+通过设置 `keepOriginalModelName: true`，可以在保留 provider header 提取能力的同时，不改写请求体中的 model 字段：
+
+```yaml
+addProviderHeader: x-higress-llm-provider
+keepOriginalModelName: true
+```
+
+以 model 为 `MiniMax/MiniMax-M2.7` 为例，经过插件后：
+- 请求头 `x-higress-llm-provider` 设置为 `MiniMax`
+- 请求体中的 model 字段保持为 `MiniMax/MiniMax-M2.7`（不改写）
+
 ### 自动路由模式（基于用户消息内容）
 
 当请求中的 model 参数设置为 `higress/auto` 时，插件会自动分析用户消息内容，并根据配置的正则规则选择合适的模型进行路由。

plugins/wasm-go/extensions/model-router/README_EN.md

@@ -9,6 +9,7 @@ The `model-router` plugin implements routing functionality based on the model pa
 | `addProviderHeader`  | string          | Optional                | -                        | Which request header to add the provider name parsed from the model parameter |
 | `modelToHeader`      | string          | Optional                | -                        | Which request header to directly add the model parameter to  |
 | `enableOnPathSuffix` | array of string | Optional                | ["/completions","/embeddings","/images/generations","/audio/speech","/fine_tuning/jobs","/moderations","/image-synthesis","/video-synthesis","/rerank","/messages"] | Only effective for requests with these specific path suffixes, can be configured as "*" to match all paths |
+| `keepOriginalModelName` | bool         | Optional                | false                    | Used with `addProviderHeader`. When set to true, the provider is still extracted into the header, but the model field in the request body is not rewritten |
 
 ## Runtime Properties
 
@@ -95,3 +96,19 @@ The original LLM request body will be changed to:
     "temperature": 0.7,
     "top_p": 0.95
 }
+```
+
+### Preserving Original Model Name (keepOriginalModelName)
+
+When using AI model aggregation platforms (e.g. Bailian/DashScope) with third-party models, some model names inherently contain `/` (e.g. `MiniMax/MiniMax-M2.7`) and are not in `provider/model` format. In this case, using `addProviderHeader` would incorrectly rewrite the model field in the request body.
+
+By setting `keepOriginalModelName: true`, you can keep the provider header extraction while preserving the original model field in the request body:
+
+```yaml
+addProviderHeader: x-higress-llm-provider
+keepOriginalModelName: true
+```
+
+For example, with model `MiniMax/MiniMax-M2.7`, after processing by the plugin:
+- Request header `x-higress-llm-provider` is set to `MiniMax`
+- The model field in the request body remains `MiniMax/MiniMax-M2.7` (not rewritten)

plugins/wasm-go/extensions/model-router/main.go

@@ -47,6 +47,7 @@ type ModelRouterConfig struct {
 	addProviderHeader     string
 	modelToHeader         string
 	enableOnPathSuffix    []string
+	keepOriginalModelName bool
 	// Auto routing configuration
 	enableAutoRouting bool
 	autoRoutingRules  []AutoRoutingRule
@@ -60,6 +61,7 @@ func parseConfig(json gjson.Result, config *ModelRouterConfig) error {
 	}
 	config.addProviderHeader = json.Get("addProviderHeader").String()
 	config.modelToHeader = json.Get("modelToHeader").String()
+	config.keepOriginalModelName = json.Get("keepOriginalModelName").Bool()
 
 	enableOnPathSuffix := json.Get("enableOnPathSuffix")
 	if enableOnPathSuffix.Exists() && enableOnPathSuffix.IsArray() {
@@ -253,12 +255,14 @@ func handleJsonBody(ctx wrapper.HttpContext, config ModelRouterConfig, body []by
 			model := parts[1]
 			_ = proxywasm.ReplaceHttpRequestHeader(config.addProviderHeader, provider)
 
+			if !config.keepOriginalModelName {
 				newBody, err := sjson.SetBytes(body, config.modelKey, model)
 				if err != nil {
 					log.Errorf("failed to update model in json body: %v", err)
 					return types.ActionContinue
 				}
 				_ = proxywasm.ReplaceHttpRequestBody(newBody)
+			}
 			log.Debugf("model route to provider: %s, model: %s", provider, model)
 		} else {
 			log.Debugf("model route to provider not work, model: %s", modelValue)
@@ -319,6 +323,7 @@ func handleMultipartBody(ctx wrapper.HttpContext, config ModelRouterConfig, body
 					model := parts[1]
 					_ = proxywasm.ReplaceHttpRequestHeader(config.addProviderHeader, provider)
 
+					if !config.keepOriginalModelName {
 						// Write modified part
 						h := make(http.Header)
 						for k, v := range part.Header {
@@ -338,6 +343,8 @@ func handleMultipartBody(ctx wrapper.HttpContext, config ModelRouterConfig, body
 						modified = true
 						log.Debugf("model route to provider: %s, model: %s", provider, model)
 						continue
+					}
+					log.Debugf("model route to provider: %s, model kept: %s", provider, modelValue)
 				} else {
 					log.Debugf("model route to provider not work, model: %s", modelValue)
 				}

plugins/wasm-go/extensions/model-router/main_test.go

@@ -288,6 +288,126 @@ func TestOnHttpRequestBody_Multipart(t *testing.T) {
 	})
 }
 
+var keepOriginalModelConfig = func() json.RawMessage {
+	data, _ := json.Marshal(map[string]interface{}{
+		"modelKey":              "model",
+		"addProviderHeader":     "x-provider",
+		"modelToHeader":         "x-model",
+		"keepOriginalModelName": true,
+		"enableOnPathSuffix": []string{
+			"/v1/chat/completions",
+		},
+	})
+	return data
+}()
+
+func TestParseConfigKeepOriginalModelName(t *testing.T) {
+	test.RunGoTest(t, func(t *testing.T) {
+		t.Run("default false", func(t *testing.T) {
+			var cfg ModelRouterConfig
+			err := parseConfig(gjson.ParseBytes(basicConfig), &cfg)
+			require.NoError(t, err)
+			require.False(t, cfg.keepOriginalModelName)
+		})
+
+		t.Run("parse true", func(t *testing.T) {
+			var cfg ModelRouterConfig
+			err := parseConfig(gjson.ParseBytes(keepOriginalModelConfig), &cfg)
+			require.NoError(t, err)
+			require.True(t, cfg.keepOriginalModelName)
+		})
+	})
+}
+
+func TestKeepOriginalModelName(t *testing.T) {
+	test.RunTest(t, func(t *testing.T) {
+		t.Run("json: provider header set but body model preserved", func(t *testing.T) {
+			host, status := test.NewTestHost(keepOriginalModelConfig)
+			defer host.Reset()
+			require.Equal(t, types.OnPluginStartStatusOK, status)
+
+			host.CallOnHttpRequestHeaders([][2]string{
+				{":authority", "example.com"},
+				{":path", "/v1/chat/completions"},
+				{":method", "POST"},
+				{"content-type", "application/json"},
+			})
+
+			origBody := []byte(`{
+				"model": "MiniMax/MiniMax-M2.7",
+				"messages": [{"role": "user", "content": "hello"}]
+			}`)
+			action := host.CallOnHttpRequestBody(origBody)
+			require.Equal(t, types.ActionContinue, action)
+
+			headers := host.GetRequestHeaders()
+			// model header keeps the full name
+			hv, found := getHeader(headers, "x-model")
+			require.True(t, found)
+			require.Equal(t, "MiniMax/MiniMax-M2.7", hv)
+			// provider header IS set (split still extracts provider)
+			pv, found := getHeader(headers, "x-provider")
+			require.True(t, found)
+			require.Equal(t, "MiniMax", pv)
+
+			// body model must remain intact (not rewritten)
+			processed := host.GetRequestBody()
+			require.NotNil(t, processed)
+			require.Equal(t, "MiniMax/MiniMax-M2.7", gjson.GetBytes(processed, "model").String())
+		})
+
+		t.Run("multipart: provider header set but body model preserved", func(t *testing.T) {
+			host, status := test.NewTestHost(keepOriginalModelConfig)
+			defer host.Reset()
+			require.Equal(t, types.OnPluginStartStatusOK, status)
+
+			var buf bytes.Buffer
+			writer := multipart.NewWriter(&buf)
+			modelWriter, err := writer.CreateFormField("model")
+			require.NoError(t, err)
+			_, err = modelWriter.Write([]byte("MiniMax/MiniMax-M2.7"))
+			require.NoError(t, err)
+			err = writer.Close()
+			require.NoError(t, err)
+
+			contentType := "multipart/form-data; boundary=" + writer.Boundary()
+			host.CallOnHttpRequestHeaders([][2]string{
+				{":authority", "example.com"},
+				{":path", "/v1/chat/completions"},
+				{":method", "POST"},
+				{"content-type", contentType},
+			})
+
+			action := host.CallOnHttpRequestBody(buf.Bytes())
+			require.Equal(t, types.ActionContinue, action)
+
+			headers := host.GetRequestHeaders()
+			hv, found := getHeader(headers, "x-model")
+			require.True(t, found)
+			require.Equal(t, "MiniMax/MiniMax-M2.7", hv)
+			// provider header IS set
+			pv, found := getHeader(headers, "x-provider")
+			require.True(t, found)
+			require.Equal(t, "MiniMax", pv)
+
+			// body model should not be rewritten
+			processed := host.GetRequestBody()
+			require.NotNil(t, processed)
+			reader := multipart.NewReader(bytes.NewReader(processed), writer.Boundary())
+			for {
+				part, err := reader.NextPart()
+				if err != nil {
+					break
+				}
+				if part.FormName() == "model" {
+					data, _ := io.ReadAll(part)
+					require.Equal(t, "MiniMax/MiniMax-M2.7", string(data))
+				}
+			}
+		})
+	})
+}
+
 // Auto routing config for tests
 var autoRoutingConfig = func() json.RawMessage {
 	data, _ := json.Marshal(map[string]interface{}{