fix(go-models): record provider token usage (#17423)

## Summary

- Parse chat, embedding, and rerank usage from provider responses
- Record usage with the correct model type even when no usage sink is
provided
- Cover SiliconFlow, Aliyun, Huawei Cloud, Qiniu, and VolcEngine
response formats

internal/entity/models/aliyun.go

@@ -52,6 +52,37 @@ func (a *AliyunModel) Name() string {
 	return "Tongyi-Qianwen"
 }
 
+// AliyunChatResponse mirrors DashScope's OpenAI-compatible chat response.
+type AliyunChatResponse struct {
+	ID      string `json:"id"`
+	Object  string `json:"object"`
+	Created int64  `json:"created"`
+	Model   string `json:"model"`
+	Choices []struct {
+		FinishReason string `json:"finish_reason"`
+		Index        int    `json:"index"`
+		Logprobs     any    `json:"logprobs"`
+		Message      struct {
+			Content          string           `json:"content"`
+			ReasoningContent string           `json:"reasoning_content"`
+			Role             string           `json:"role"`
+			ToolCalls        []map[string]any `json:"tool_calls"`
+		} `json:"message"`
+	} `json:"choices"`
+	SystemFingerprint string `json:"system_fingerprint"`
+	Usage             struct {
+		CompletionTokens        int `json:"completion_tokens"`
+		PromptTokens            int `json:"prompt_tokens"`
+		TotalTokens             int `json:"total_tokens"`
+		CompletionTokensDetails struct {
+			ReasoningTokens int `json:"reasoning_tokens"`
+		} `json:"completion_tokens_details"`
+		PromptTokensDetails struct {
+			CachedTokens int `json:"cached_tokens"`
+		} `json:"prompt_tokens_details"`
+	} `json:"usage"`
+}
+
 func (a *AliyunModel) ChatWithMessages(ctx context.Context, modelName string, messages []Message, apiConfig *APIConfig, chatModelConfig *ChatConfig, modelUsage *common.ModelUsage) (*ChatResponse, error) {
 	if err := a.baseModel.APIConfigCheck(apiConfig); err != nil {
 		return nil, err
@@ -121,52 +152,47 @@ func (a *AliyunModel) ChatWithMessages(ctx context.Context, modelName string, me
 		return nil, fmt.Errorf("API request failed with status %d: %s", resp.StatusCode, string(body))
 	}
 
-	// Parse response
-	var result map[string]interface{}
-	if err = json.Unmarshal(body, &result); err != nil {
-		return nil, fmt.Errorf("failed to parse response: %w", err)
-	}
-
-	choices, ok := result["choices"].([]interface{})
-	if !ok || len(choices) == 0 {
-		return nil, fmt.Errorf("no choices in response")
-	}
-
-	firstChoice, ok := choices[0].(map[string]interface{})
-	if !ok {
-		return nil, fmt.Errorf("invalid choice format")
-	}
-
-	messageMap, ok := firstChoice["message"].(map[string]interface{})
-	if !ok {
-		return nil, fmt.Errorf("invalid message format")
-	}
-
-	answer, hasAnswer := messageMap["content"].(string)
-	toolCalls := extractToolCalls(messageMap)
-	if !hasAnswer && len(toolCalls) == 0 {
-		return nil, fmt.Errorf("response contains neither content nor tool calls")
-	}
-
-	var reasonContent string
-	if chatModelConfig != nil && chatModelConfig.Thinking != nil && *chatModelConfig.Thinking {
-		reasonContent, ok = messageMap["reasoning_content"].(string)
-		if !ok {
-			return nil, fmt.Errorf("invalid content format")
+	return parseChatCompletionResponse(body, chatModelConfig, modelUsage, func(body []byte, chatConfig *ChatConfig) (chatResponseParts, error) {
+		var result AliyunChatResponse
+		if err := json.Unmarshal(body, &result); err != nil {
+			return chatResponseParts{}, fmt.Errorf("failed to parse response: %w", err)
 		}
-		// if first char of reasonContent is \n remove the '\n'
-		if reasonContent != "" && reasonContent[0] == '\n' {
-			reasonContent = reasonContent[1:]
+		if len(result.Choices) == 0 {
+			return chatResponseParts{}, fmt.Errorf("no choices in response")
 		}
-	}
 
-	chatResponse := &ChatResponse{
-		Answer:        &answer,
-		ReasonContent: &reasonContent,
-		ToolCalls:     toolCalls,
-	}
+		choice := &result.Choices[0]
+		content := choice.Message.Content
+		if content == "" && len(choice.Message.ToolCalls) == 0 {
+			return chatResponseParts{}, fmt.Errorf("response contains neither content nor tool calls")
+		}
+		reasonContent := ""
+		if chatConfig != nil && chatConfig.Thinking != nil && *chatConfig.Thinking {
+			reasonContent = choice.Message.ReasoningContent
+			if reasonContent == "" {
+				reasoning, answer := GetThinkingAndAnswer(chatConfig.ModelClass, &content)
+				if reasoning != nil {
+					reasonContent = *reasoning
+					content = *answer
+				}
+			}
+			if reasonContent != "" && reasonContent[0] == '\n' {
+				reasonContent = reasonContent[1:]
+			}
+		}
 
-	return chatResponse, nil
+		return chatResponseParts{
+			RequestID:     result.ID,
+			Content:       &content,
+			ReasonContent: &reasonContent,
+			ToolCalls:     choice.Message.ToolCalls,
+			Usage: &TokenUsage{
+				PromptTokens:     result.Usage.PromptTokens,
+				CompletionTokens: result.Usage.CompletionTokens,
+				TotalTokens:      result.Usage.TotalTokens,
+			},
+		}, nil
+	})
 }
 
 // ChatStreamlyWithSender sends messages and streams response via sender function (best performance, no channel)
@@ -187,6 +213,7 @@ func (a *AliyunModel) ChatStreamlyWithSender(ctx context.Context, modelName stri
 
 	// Build request body with streaming enabled
 	reqBody := buildRequestBody(chatModelConfig, modelName, messages, true)
+	reqBody["stream_options"] = map[string]interface{}{"include_usage": true}
 
 	if chatModelConfig != nil {
 		if chatModelConfig.Stream != nil && !*chatModelConfig.Stream {
@@ -240,6 +267,14 @@ func (a *AliyunModel) ChatStreamlyWithSender(ctx context.Context, modelName stri
 	done, err := ParseSSEStream[map[string]interface{}](resp.Body, func(event map[string]interface{}) error {
 		common.Info(fmt.Sprintf("%v", event))
 
+		tokenUsage, found, usageErr := decodeOpenAICompatibleStreamUsage(event)
+		if usageErr != nil {
+			return usageErr
+		}
+		if found {
+			applyStreamUsage(chatModelConfig, modelUsage, tokenUsage)
+		}
+
 		choices, ok := event["choices"].([]interface{})
 		if !ok || len(choices) == 0 {
 			return nil
@@ -332,11 +367,11 @@ func aliyunToolChoice(modelName string, messages []Message, configured *string)
 }
 
 type aliyunEmbeddingResponse struct {
-	Data   []EmbeddingData `json:"data"`
-	Model  string          `json:"model"`
-	Object string          `json:"object"`
-	Usage  aliyunUsage     `json:"usage"`
-	ID     string          `json:"id"`
+	Data   []aliyunEmbeddingData `json:"data"`
+	Model  string                `json:"model"`
+	Object string                `json:"object"`
+	Usage  aliyunUsage           `json:"usage"`
+	ID     string                `json:"id"`
 }
 
 type aliyunEmbeddingData struct {
@@ -420,6 +455,10 @@ func (a *AliyunModel) Embed(ctx context.Context, modelName *string, texts []stri
 		embeddingData.Index = dataElem.Index
 		embeddings = append(embeddings, embeddingData)
 	}
+	recordResponseUsage(modelUsage, parsed.ID, &TokenUsage{
+		PromptTokens: parsed.Usage.PromptTokens,
+		TotalTokens:  parsed.Usage.TotalTokens,
+	}, "embedding")
 
 	return embeddings, nil
 }
@@ -433,10 +472,14 @@ type aliyunRerankRequest struct {
 }
 
 type aliyunRerankResponse struct {
+	ID      string `json:"id"`
 	Results []struct {
 		Index          int     `json:"index"`
 		RelevanceScore float64 `json:"relevance_score"`
 	} `json:"results"`
+	Usage struct {
+		TotalTokens int `json:"total_tokens"`
+	} `json:"usage"`
 }
 
 func (a *AliyunModel) Rerank(ctx context.Context, modelName *string, query string, documents []string, apiConfig *APIConfig, rerankConfig *RerankConfig, modelUsage *common.ModelUsage) (*RerankResponse, error) {
@@ -459,8 +502,11 @@ func (a *AliyunModel) Rerank(ctx context.Context, modelName *string, query strin
 
 	url := fmt.Sprintf("%s/%s", strings.TrimSuffix(baseURL, "/"), a.baseModel.URLSuffix.Rerank)
 
-	var topN = rerankConfig.TopN
-	if rerankConfig.TopN == 0 {
+	topN := len(documents)
+	if rerankConfig != nil && rerankConfig.TopN > 0 {
+		topN = rerankConfig.TopN
+	}
+	if topN == 0 {
 		topN = len(documents)
 	}
 
@@ -503,12 +549,19 @@ func (a *AliyunModel) Rerank(ctx context.Context, modelName *string, query strin
 		return nil, fmt.Errorf("Aliyun rerank API error: %s, body: %s", resp.Status, string(body))
 	}
 
-	var rerankResponse RerankResponse
-	if err = json.Unmarshal(body, &rerankResponse); err != nil {
+	var parsed aliyunRerankResponse
+	if err = json.Unmarshal(body, &parsed); err != nil {
 		return nil, fmt.Errorf("failed to parse response: %w", err)
 	}
+	rerankResponse := &RerankResponse{Data: make([]RerankResult, 0, len(parsed.Results))}
+	for _, item := range parsed.Results {
+		rerankResponse.Data = append(rerankResponse.Data, RerankResult{Index: item.Index, RelevanceScore: item.RelevanceScore})
+	}
+	recordResponseUsage(modelUsage, parsed.ID, &TokenUsage{
+		TotalTokens: parsed.Usage.TotalTokens,
+	}, "rerank")
 
-	return &rerankResponse, nil
+	return rerankResponse, nil
 }
 
 // TranscribeAudio transcribe audio

internal/entity/models/base_model.go

@@ -63,9 +63,7 @@ func parseChatCompletionResponse(body []byte, chatConfig *ChatConfig, modelUsage
 		return nil, err
 	}
 
-	if err := collectChatModelUsage(modelUsage, parts.RequestID, parts.Usage); err != nil {
-		common.Error("Failed to collect model usage", err)
-	}
+	recordResponseUsage(modelUsage, parts.RequestID, parts.Usage, "chat")
 
 	return &ChatResponse{
 		Answer:        parts.Content,
@@ -75,14 +73,19 @@ func parseChatCompletionResponse(body []byte, chatConfig *ChatConfig, modelUsage
 	}, nil
 }
 
-// collectChatModelUsage records one completed chat response when the caller
-// supplied a usage sink.
-func collectChatModelUsage(modelUsage *common.ModelUsage, requestID string, usage *TokenUsage) error {
+// recordResponseUsage records the request ID and token usage returned by a
+// completed model response.
+func recordResponseUsage(modelUsage *common.ModelUsage, requestID string, usage *TokenUsage, modelType string) {
 	if modelUsage == nil {
-		return nil
+		return
+	}
+	if modelUsage.Type == "" {
+		modelUsage.Type = modelType
 	}
 	modelUsage.RequestID = requestID
-	return collectModelUsage(modelUsage, usage)
+	if err := collectModelUsage(modelUsage, usage); err != nil {
+		common.Error("Failed to collect model usage", err)
+	}
 }
 
 // collectModelUsage records token usage and response time for one model call.
@@ -125,6 +128,9 @@ func applyStreamUsage(chatConfig *ChatConfig, modelUsage *common.ModelUsage, usa
 	if chatConfig != nil {
 		chatConfig.UsageResult = usage
 	}
+	if modelUsage == nil {
+		return
+	}
 	if err := collectModelUsage(modelUsage, usage); err != nil {
 		common.Error("Failed to collect model usage", err)
 	}

internal/entity/models/huaweicloud.go

@@ -50,6 +50,37 @@ func (h *HuaweiCloudModel) Name() string {
 	return "huaweicloud"
 }
 
+// HuaweiCloudChatResponse captures the OpenAI-compatible fields consumed by RAGFlow.
+type HuaweiCloudChatResponse struct {
+	ID      string `json:"id"`
+	Object  string `json:"object"`
+	Created int64  `json:"created"`
+	Model   string `json:"model"`
+	Choices []struct {
+		FinishReason string `json:"finish_reason"`
+		Index        int    `json:"index"`
+		Logprobs     any    `json:"logprobs"`
+		Message      struct {
+			Content          string           `json:"content"`
+			ReasoningContent string           `json:"reasoning_content"`
+			Role             string           `json:"role"`
+			ToolCalls        []map[string]any `json:"tool_calls"`
+		} `json:"message"`
+	} `json:"choices"`
+	SystemFingerprint string `json:"system_fingerprint"`
+	Usage             struct {
+		CompletionTokens        int `json:"completion_tokens"`
+		PromptTokens            int `json:"prompt_tokens"`
+		TotalTokens             int `json:"total_tokens"`
+		CompletionTokensDetails struct {
+			ReasoningTokens int `json:"reasoning_tokens"`
+		} `json:"completion_tokens_details"`
+		PromptTokensDetails struct {
+			CachedTokens int `json:"cached_tokens"`
+		} `json:"prompt_tokens_details"`
+	} `json:"usage"`
+}
+
 func huaweiCloudRegion(api *APIConfig) string {
 	region := "default"
 	if api != nil && api.Region != nil && *api.Region != "" {
@@ -220,44 +251,44 @@ func (h *HuaweiCloudModel) ChatWithMessages(ctx context.Context, modelName strin
 		return nil, fmt.Errorf("Huawei Cloud chat API error: status %d, body: %s", rep.StatusCode, string(body))
 	}
 
-	var result map[string]interface{}
-	if err = json.Unmarshal(body, &result); err != nil {
-		return nil, fmt.Errorf("failed to parse response: %w", err)
-	}
+	return parseChatCompletionResponse(body, chatModelConfig, modelUsage, func(body []byte, chatConfig *ChatConfig) (chatResponseParts, error) {
+		var result HuaweiCloudChatResponse
+		if err := json.Unmarshal(body, &result); err != nil {
+			return chatResponseParts{}, fmt.Errorf("failed to parse response: %w", err)
+		}
+		if len(result.Choices) == 0 {
+			return chatResponseParts{}, fmt.Errorf("no choices in response")
+		}
 
-	choices, ok := result["choices"].([]interface{})
-	if !ok || len(choices) == 0 {
-		return nil, fmt.Errorf("no choices in response")
-	}
-
-	firstChoice, ok := choices[0].(map[string]interface{})
-	if !ok {
-		return nil, fmt.Errorf("invalid choice format")
-	}
-
-	messageMap, ok := firstChoice["message"].(map[string]interface{})
-	if !ok {
-		return nil, fmt.Errorf("invalid message format")
-	}
-	toolCalls := extractToolCalls(messageMap)
-	content, hasContent := messageMap["content"].(string)
-	if !hasContent && len(toolCalls) == 0 {
-		return nil, fmt.Errorf("invalid content format")
-	}
-
-	reasonContent := ""
-	if r, ok := messageMap["reasoning_content"].(string); ok {
-		reasonContent = r
+		choice := &result.Choices[0]
+		content := choice.Message.Content
+		if content == "" && len(choice.Message.ToolCalls) == 0 {
+			return chatResponseParts{}, fmt.Errorf("invalid content format")
+		}
+		reasonContent := choice.Message.ReasoningContent
+		if chatConfig != nil && chatConfig.Thinking != nil && *chatConfig.Thinking && reasonContent == "" {
+			reasoning, answer := GetThinkingAndAnswer(chatConfig.ModelClass, &content)
+			if reasoning != nil {
+				reasonContent = *reasoning
+				content = *answer
+			}
+		}
 		if reasonContent != "" && reasonContent[0] == '\n' {
 			reasonContent = reasonContent[1:]
 		}
-	}
 
-	return &ChatResponse{
-		Answer:        &content,
-		ReasonContent: &reasonContent,
-		ToolCalls:     toolCalls,
-	}, nil
+		return chatResponseParts{
+			RequestID:     result.ID,
+			Content:       &content,
+			ReasonContent: &reasonContent,
+			ToolCalls:     choice.Message.ToolCalls,
+			Usage: &TokenUsage{
+				PromptTokens:     result.Usage.PromptTokens,
+				CompletionTokens: result.Usage.CompletionTokens,
+				TotalTokens:      result.Usage.TotalTokens,
+			},
+		}, nil
+	})
 }
 
 func (h *HuaweiCloudModel) ChatStreamlyWithSender(ctx context.Context, modelName string, messages []Message, apiConfig *APIConfig, chatModelConfig *ChatConfig, modelUsage *common.ModelUsage, sender func(*string, *string) error) error {
@@ -329,6 +360,14 @@ func (h *HuaweiCloudModel) ChatStreamlyWithSender(ctx context.Context, modelName
 			return fmt.Errorf("huaweicloud: upstream stream error: %v", apiErr)
 		}
 
+		tokenUsage, found, usageErr := decodeOpenAICompatibleStreamUsage(event)
+		if usageErr != nil {
+			return usageErr
+		}
+		if found {
+			applyStreamUsage(chatModelConfig, modelUsage, tokenUsage)
+		}
+
 		choices, ok := event["choices"].([]interface{})
 		if !ok || len(choices) == 0 {
 			return nil
@@ -376,10 +415,27 @@ func (h *HuaweiCloudModel) ChatStreamlyWithSender(ctx context.Context, modelName
 }
 
 type huaweiCloudEmbeddingResponse struct {
+	ID   string `json:"id"`
 	Data []struct {
 		Embedding []float64 `json:"embedding"`
 		Index     *int      `json:"index"`
 	} `json:"data"`
+	Usage struct {
+		PromptTokens int `json:"prompt_tokens"`
+		TotalTokens  int `json:"total_tokens"`
+	} `json:"usage"`
+}
+
+type huaweiCloudRerankResponse struct {
+	ID      string `json:"id"`
+	Results []struct {
+		Index          int     `json:"index"`
+		RelevanceScore float64 `json:"relevance_score"`
+	} `json:"results"`
+	Usage struct {
+		PromptTokens int `json:"prompt_tokens"`
+		TotalTokens  int `json:"total_tokens"`
+	} `json:"usage"`
 }
 
 func (h *HuaweiCloudModel) Embed(ctx context.Context, modelName *string, texts []string, apiConfig *APIConfig, embeddingConfig *EmbeddingConfig, modelUsage *common.ModelUsage) ([]EmbeddingData, error) {
@@ -476,6 +532,10 @@ func (h *HuaweiCloudModel) Embed(ctx context.Context, modelName *string, texts [
 			return nil, fmt.Errorf("missing embedding index %d", i)
 		}
 	}
+	recordResponseUsage(modelUsage, parsed.ID, &TokenUsage{
+		PromptTokens: parsed.Usage.PromptTokens,
+		TotalTokens:  parsed.Usage.TotalTokens,
+	}, "embedding")
 
 	return embeddings, nil
 }
@@ -541,12 +601,7 @@ func (h *HuaweiCloudModel) Rerank(ctx context.Context, modelName *string, query
 		return nil, fmt.Errorf("Huawei Cloud rerank API error: status %d, body: %s", resp.StatusCode, string(body))
 	}
 
-	var parsed struct {
-		Results []struct {
-			Index          int     `json:"index"`
-			RelevanceScore float64 `json:"relevance_score"`
-		} `json:"results"`
-	}
+	var parsed huaweiCloudRerankResponse
 	if err = json.Unmarshal(body, &parsed); err != nil {
 		return nil, fmt.Errorf("failed to decode response: %w", err)
 	}
@@ -569,6 +624,10 @@ func (h *HuaweiCloudModel) Rerank(ctx context.Context, modelName *string, query
 			RelevanceScore: item.RelevanceScore,
 		})
 	}
+	recordResponseUsage(modelUsage, parsed.ID, &TokenUsage{
+		PromptTokens: parsed.Usage.PromptTokens,
+		TotalTokens:  parsed.Usage.TotalTokens,
+	}, "rerank")
 
 	return result, nil
 }

internal/entity/models/non_chat_usage_test.go

@@ -0,0 +1,124 @@
+package models
+
+import (
+	"net/http"
+	"net/http/httptest"
+	"ragflow/internal/common"
+	"testing"
+)
+
+func TestProviderEmbeddingAndRerankUsage(t *testing.T) {
+	type providerCase struct {
+		name           string
+		newModel       func(string) ModelDriver
+		embeddingBody  string
+		rerankBody     string
+		embedInput     int
+		embedTotal     int
+		rerankInput    int
+		rerankOutput   int
+		rerankTotal    int
+		supportsRerank bool
+	}
+
+	cases := []providerCase{
+		{
+			name: "siliconflow",
+			newModel: func(baseURL string) ModelDriver {
+				return NewSiliconflowModel(map[string]string{"default": baseURL}, URLSuffix{Embedding: "embedding", Rerank: "rerank"})
+			},
+			embeddingBody: `{"id":"embed-sf","data":[{"embedding":[0.1,0.2],"index":0}],"usage":{"prompt_tokens":7,"total_tokens":7}}`,
+			rerankBody:    `{"id":"rerank-sf","results":[{"index":0,"relevance_score":0.9}],"meta":{"tokens":{"input_tokens":7,"output_tokens":2}}}`,
+			embedInput:    7, embedTotal: 7, rerankInput: 7, rerankOutput: 2, rerankTotal: 9,
+			supportsRerank: true,
+		},
+		{
+			name: "aliyun",
+			newModel: func(baseURL string) ModelDriver {
+				return NewAliyunModel(map[string]string{"default": baseURL}, URLSuffix{Embedding: "embedding", Rerank: "rerank"})
+			},
+			embeddingBody: `{"id":"embed-aliyun","data":[{"embedding":[0.1,0.2],"index":0}],"usage":{"prompt_tokens":7,"total_tokens":7}}`,
+			rerankBody:    `{"id":"rerank-aliyun","results":[{"index":0,"relevance_score":0.9}],"usage":{"total_tokens":9}}`,
+			embedInput:    7, embedTotal: 7, rerankTotal: 9,
+			supportsRerank: true,
+		},
+		{
+			name: "huaweicloud",
+			newModel: func(baseURL string) ModelDriver {
+				return NewHuaweiCloudModel(map[string]string{"default": baseURL}, URLSuffix{Embedding: "embedding", Rerank: "rerank"})
+			},
+			embeddingBody: `{"id":"embed-huawei","data":[{"embedding":[0.1,0.2],"index":0}],"usage":{"prompt_tokens":7,"total_tokens":7}}`,
+			rerankBody:    `{"id":"rerank-huawei","results":[{"index":0,"relevance_score":0.9}],"usage":{"prompt_tokens":9,"total_tokens":9}}`,
+			embedInput:    7, embedTotal: 7, rerankInput: 9, rerankTotal: 9,
+			supportsRerank: true,
+		},
+		{
+			name: "volcengine",
+			newModel: func(baseURL string) ModelDriver {
+				return NewVolcEngine(map[string]string{"default": baseURL}, URLSuffix{Embedding: "embedding"})
+			},
+			embeddingBody: `{"id":"embed-volc","data":{"embedding":[0.1,0.2]},"usage":{"prompt_tokens":7,"total_tokens":7}}`,
+			embedInput:    7, embedTotal: 7,
+		},
+	}
+
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+				w.Header().Set("Content-Type", "application/json")
+				switch r.URL.Path {
+				case "/embedding":
+					_, _ = w.Write([]byte(tc.embeddingBody))
+				case "/rerank":
+					_, _ = w.Write([]byte(tc.rerankBody))
+				default:
+					http.NotFound(w, r)
+				}
+			}))
+			defer server.Close()
+
+			apiKey := "test-key"
+			modelName := "test-model"
+			model := tc.newModel(server.URL)
+
+			embeddingUsage := &common.ModelUsage{}
+			embeddings, err := model.Embed(t.Context(), &modelName, []string{"document"}, &APIConfig{ApiKey: &apiKey}, &EmbeddingConfig{}, embeddingUsage)
+			if err != nil {
+				t.Fatalf("Embed: %v", err)
+			}
+			if len(embeddings) != 1 || len(embeddings[0].Embedding) != 2 {
+				t.Fatalf("embeddings=%#v", embeddings)
+			}
+			assertModelUsage(t, embeddingUsage, tc.embedInput, 0, tc.embedTotal)
+			if embeddingUsage.Type != "embedding" {
+				t.Fatalf("embedding usage type=%q, want embedding", embeddingUsage.Type)
+			}
+			if !tc.supportsRerank {
+				return
+			}
+
+			rerankUsage := &common.ModelUsage{}
+			reranked, err := model.Rerank(t.Context(), &modelName, "query", []string{"document"}, &APIConfig{ApiKey: &apiKey}, &RerankConfig{TopN: 1}, rerankUsage)
+			if err != nil {
+				t.Fatalf("Rerank: %v", err)
+			}
+			if len(reranked.Data) != 1 || reranked.Data[0].Index != 0 {
+				t.Fatalf("reranked=%#v", reranked)
+			}
+			assertModelUsage(t, rerankUsage, tc.rerankInput, tc.rerankOutput, tc.rerankTotal)
+			if rerankUsage.Type != "rerank" {
+				t.Fatalf("rerank usage type=%q, want rerank", rerankUsage.Type)
+			}
+		})
+	}
+}
+
+func assertModelUsage(t *testing.T, usage *common.ModelUsage, input, output, total int) {
+	t.Helper()
+	if usage.RequestID == "" {
+		t.Fatal("usage request ID is empty")
+	}
+	if usage.InputTokens != input || usage.OutputTokens != output || usage.TotalTokens != total {
+		t.Fatalf("usage=(%d,%d,%d), want (%d,%d,%d)", usage.InputTokens, usage.OutputTokens, usage.TotalTokens, input, output, total)
+	}
+}

internal/entity/models/openai_compatible_usage_test.go

@@ -0,0 +1,110 @@
+//
+//  Copyright 2026 The InfiniFlow Authors. All Rights Reserved.
+//
+//  Licensed under the Apache License, Version 2.0 (the "License");
+//  you may not use this file except in compliance with the License.
+//  You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+//  Unless required by applicable law or agreed to in writing, software
+//  distributed under the License is distributed on an "AS IS" BASIS,
+//  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+//  See the License for the specific language governing permissions and
+//  limitations under the License.
+//
+
+package models
+
+import (
+	"encoding/json"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+)
+
+func TestOpenAICompatibleProvidersExtractStreamingUsage(t *testing.T) {
+	providers := []struct {
+		name string
+		new  func(string) ModelDriver
+		path string
+	}{
+		{
+			name: "siliconflow",
+			new: func(baseURL string) ModelDriver {
+				return NewSiliconflowModel(map[string]string{"default": baseURL}, URLSuffix{Chat: "chat/completions"})
+			},
+			path: "/chat/completions",
+		},
+		{
+			name: "aliyun",
+			new: func(baseURL string) ModelDriver {
+				return NewAliyunModel(map[string]string{"default": baseURL}, URLSuffix{Chat: "chat/completions"})
+			},
+			path: "/chat/completions",
+		},
+		{
+			name: "qiniu",
+			new: func(baseURL string) ModelDriver {
+				return NewQiniuModel(map[string]string{"default": baseURL}, URLSuffix{Chat: "chat/completions"})
+			},
+			path: "/chat/completions",
+		},
+		{
+			name: "volcengine",
+			new: func(baseURL string) ModelDriver {
+				return NewVolcEngine(map[string]string{"default": baseURL}, URLSuffix{Chat: "chat/completions"})
+			},
+			path: "/chat/completions",
+		},
+	}
+
+	for _, provider := range providers {
+		provider := provider
+		t.Run(provider.name, func(t *testing.T) {
+			server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+				if r.Method != http.MethodPost {
+					t.Errorf("method=%s, want POST", r.Method)
+				}
+				if got := r.Header.Get("Authorization"); got != "Bearer test-key" {
+					t.Errorf("Authorization=%q, want Bearer test-key", got)
+				}
+				if r.URL.Path != provider.path {
+					t.Errorf("path=%q, want %q", r.URL.Path, provider.path)
+				}
+				var requestBody map[string]any
+				if err := json.NewDecoder(r.Body).Decode(&requestBody); err != nil {
+					t.Fatalf("decode request: %v", err)
+				}
+				streamOptions, ok := requestBody["stream_options"].(map[string]any)
+				if !ok || streamOptions["include_usage"] != true {
+					t.Errorf("stream_options=%#v, want include_usage=true", requestBody["stream_options"])
+				}
+				w.Header().Set("Content-Type", "text/event-stream")
+				_, _ = io.WriteString(w, "data: {\"choices\":[{\"delta\":{\"content\":\"answer\"},\"finish_reason\":null}]}\n\n")
+				_, _ = io.WriteString(w, "data: {\"choices\":[],\"usage\":{\"prompt_tokens\":3,\"completion_tokens\":5,\"total_tokens\":8}}\n\n")
+				_, _ = io.WriteString(w, "data: [DONE]\n\n")
+			}))
+			defer server.Close()
+
+			apiKey := "test-key"
+			config := &ChatConfig{}
+			err := provider.new(server.URL).ChatStreamlyWithSender(
+				t.Context(),
+				"model",
+				[]Message{{Role: "user", Content: "hello"}},
+				&APIConfig{ApiKey: &apiKey},
+				config,
+				nil,
+				func(*string, *string) error { return nil },
+			)
+			if err != nil {
+				t.Fatalf("ChatStreamlyWithSender: %v", err)
+			}
+			if config.UsageResult == nil || config.UsageResult.PromptTokens != 3 || config.UsageResult.CompletionTokens != 5 || config.UsageResult.TotalTokens != 8 {
+				t.Fatalf("UsageResult=%#v, want prompt=3 completion=5 total=8", config.UsageResult)
+			}
+		})
+	}
+}

internal/entity/models/qiniu.go

@@ -50,6 +50,37 @@ func (q *QiniuModel) Name() string {
 	return "qiniu"
 }
 
+// QiniuChatResponse captures the OpenAI-compatible fields consumed by RAGFlow.
+type QiniuChatResponse struct {
+	ID      string `json:"id"`
+	Object  string `json:"object"`
+	Created int64  `json:"created"`
+	Model   string `json:"model"`
+	Choices []struct {
+		FinishReason string `json:"finish_reason"`
+		Index        int    `json:"index"`
+		Logprobs     any    `json:"logprobs"`
+		Message      struct {
+			Content          string           `json:"content"`
+			ReasoningContent string           `json:"reasoning_content"`
+			Role             string           `json:"role"`
+			ToolCalls        []map[string]any `json:"tool_calls"`
+		} `json:"message"`
+	} `json:"choices"`
+	SystemFingerprint string `json:"system_fingerprint"`
+	Usage             struct {
+		CompletionTokens        int `json:"completion_tokens"`
+		PromptTokens            int `json:"prompt_tokens"`
+		TotalTokens             int `json:"total_tokens"`
+		CompletionTokensDetails struct {
+			ReasoningTokens int `json:"reasoning_tokens"`
+		} `json:"completion_tokens_details"`
+		PromptTokensDetails struct {
+			CachedTokens int `json:"cached_tokens"`
+		} `json:"prompt_tokens_details"`
+	} `json:"usage"`
+}
+
 var qiniuQwenThinkingModels = map[string]struct{}{
 	"qwen3-next-80b-a3b-thinking":   {},
 	"qwen3-235b-a22b-thinking-2507": {},
@@ -191,51 +222,47 @@ func (q *QiniuModel) ChatWithMessages(ctx context.Context, modelName string, mes
 		return nil, fmt.Errorf("API request failed with status %d: %s", resp.StatusCode, string(body))
 	}
 
-	var result map[string]interface{}
-	if err := json.Unmarshal(body, &result); err != nil {
-		return nil, fmt.Errorf("failed to parse response: %w", err)
-	}
-
-	choices, ok := result["choices"].([]interface{})
-	if !ok || len(choices) == 0 {
-		return nil, fmt.Errorf("no choices in response")
-	}
-
-	firstChoice, ok := choices[0].(map[string]interface{})
-	if !ok {
-		return nil, fmt.Errorf("invalid choice format")
-	}
-
-	messageMap, ok := firstChoice["message"].(map[string]interface{})
-	if !ok {
-		return nil, fmt.Errorf("invalid message format")
-	}
-
-	toolCalls := extractToolCalls(messageMap)
-	content, hasContent := messageMap["content"].(string)
-	if !hasContent && len(toolCalls) == 0 {
-		return nil, fmt.Errorf("invalid content format")
-	}
-
-	var reasonContent string
-	if chatModelConfig != nil && chatModelConfig.Thinking != nil && *chatModelConfig.Thinking {
-		reasonContent, ok = messageMap["reasoning_content"].(string)
-		if !ok && len(toolCalls) == 0 {
-			return nil, fmt.Errorf("invalid reasoning content format")
+	return parseChatCompletionResponse(body, chatModelConfig, modelUsage, func(body []byte, chatConfig *ChatConfig) (chatResponseParts, error) {
+		var result QiniuChatResponse
+		if err := json.Unmarshal(body, &result); err != nil {
+			return chatResponseParts{}, fmt.Errorf("failed to parse response: %w", err)
+		}
+		if len(result.Choices) == 0 {
+			return chatResponseParts{}, fmt.Errorf("no choices in response")
 		}
 
-		if reasonContent != "" && reasonContent[0] == '\n' {
-			reasonContent = reasonContent[1:]
+		choice := &result.Choices[0]
+		content := choice.Message.Content
+		if content == "" && len(choice.Message.ToolCalls) == 0 {
+			return chatResponseParts{}, fmt.Errorf("invalid content format")
+		}
+		reasonContent := ""
+		if chatConfig != nil && chatConfig.Thinking != nil && *chatConfig.Thinking {
+			reasonContent = choice.Message.ReasoningContent
+			if reasonContent == "" {
+				reasoning, answer := GetThinkingAndAnswer(chatConfig.ModelClass, &content)
+				if reasoning != nil {
+					reasonContent = *reasoning
+					content = *answer
+				}
+			}
+			if reasonContent != "" && reasonContent[0] == '\n' {
+				reasonContent = reasonContent[1:]
+			}
 		}
-	}
 
-	chatResponse := &ChatResponse{
-		Answer:        &content,
-		ReasonContent: &reasonContent,
-		ToolCalls:     toolCalls,
-	}
-
-	return chatResponse, nil
+		return chatResponseParts{
+			RequestID:     result.ID,
+			Content:       &content,
+			ReasonContent: &reasonContent,
+			ToolCalls:     choice.Message.ToolCalls,
+			Usage: &TokenUsage{
+				PromptTokens:     result.Usage.PromptTokens,
+				CompletionTokens: result.Usage.CompletionTokens,
+				TotalTokens:      result.Usage.TotalTokens,
+			},
+		}, nil
+	})
 }
 
 func (q *QiniuModel) ChatStreamlyWithSender(ctx context.Context, modelName string, messages []Message, apiConfig *APIConfig, chatModelConfig *ChatConfig, modelUsage *common.ModelUsage, sender func(*string, *string) error) error {
@@ -254,6 +281,7 @@ func (q *QiniuModel) ChatStreamlyWithSender(ctx context.Context, modelName strin
 	url := fmt.Sprintf("%s/%s", baseURL, q.baseModel.URLSuffix.Chat)
 
 	reqBody := buildRequestBody(chatModelConfig, modelName, messages, true)
+	reqBody["stream_options"] = map[string]interface{}{"include_usage": true}
 
 	if chatModelConfig != nil {
 		applyQiniuThinkingConfig(reqBody, modelName, chatModelConfig)
@@ -291,6 +319,14 @@ func (q *QiniuModel) ChatStreamlyWithSender(ctx context.Context, modelName strin
 	done, err := ParseSSEStream[map[string]interface{}](resp.Body, func(event map[string]interface{}) error {
 		common.Info(fmt.Sprintf("%v", event))
 
+		tokenUsage, found, usageErr := decodeOpenAICompatibleStreamUsage(event)
+		if usageErr != nil {
+			return usageErr
+		}
+		if found {
+			applyStreamUsage(chatModelConfig, modelUsage, tokenUsage)
+		}
+
 		choices, ok := event["choices"].([]interface{})
 		if !ok || len(choices) == 0 {
 			return nil

internal/entity/models/siliconflow.go

@@ -55,6 +55,38 @@ func (s *SiliconflowModel) Name() string {
 	return "SILICONFLOW"
 }
 
+// SiliconflowChatResponse mirrors the response returned by SiliconFlow's
+// POST /chat/completions endpoint. SiliconFlow uses the OpenAI-compatible
+// schema and includes reasoning and cache token details when available.
+type SiliconflowChatResponse struct {
+	ID      string `json:"id"`
+	Object  string `json:"object"`
+	Created int64  `json:"created"`
+	Model   string `json:"model"`
+	Choices []struct {
+		FinishReason string `json:"finish_reason"`
+		Index        int    `json:"index"`
+		Message      struct {
+			Content          string           `json:"content"`
+			ReasoningContent string           `json:"reasoning_content"`
+			Role             string           `json:"role"`
+			ToolCalls        []map[string]any `json:"tool_calls"`
+		} `json:"message"`
+	} `json:"choices"`
+	SystemFingerprint string `json:"system_fingerprint"`
+	Usage             struct {
+		CompletionTokens        int `json:"completion_tokens"`
+		PromptTokens            int `json:"prompt_tokens"`
+		TotalTokens             int `json:"total_tokens"`
+		CompletionTokensDetails struct {
+			ReasoningTokens int `json:"reasoning_tokens"`
+		} `json:"completion_tokens_details"`
+		PromptTokensDetails struct {
+			CachedTokens int `json:"cached_tokens"`
+		} `json:"prompt_tokens_details"`
+	} `json:"usage"`
+}
+
 // ChatWithMessages sends multiple messages with roles and returns response
 func (s *SiliconflowModel) ChatWithMessages(ctx context.Context, modelName string, messages []Message, apiConfig *APIConfig, chatModelConfig *ChatConfig, modelUsage *common.ModelUsage) (*ChatResponse, error) {
 	if err := s.baseModel.APIConfigCheck(apiConfig); err != nil {
@@ -114,58 +146,48 @@ func (s *SiliconflowModel) ChatWithMessages(ctx context.Context, modelName strin
 		return nil, fmt.Errorf("API request failed with status %d: %s", resp.StatusCode, string(body))
 	}
 
-	// Parse response
-	var result map[string]interface{}
-	if err := json.Unmarshal(body, &result); err != nil {
-		return nil, fmt.Errorf("failed to parse response: %w", err)
-	}
+	return parseChatCompletionResponse(body, chatModelConfig, modelUsage, func(body []byte, chatConfig *ChatConfig) (chatResponseParts, error) {
+		var result SiliconflowChatResponse
+		if err := json.Unmarshal(body, &result); err != nil {
+			return chatResponseParts{}, fmt.Errorf("failed to parse response: %w", err)
+		}
+		if len(result.Choices) == 0 {
+			return chatResponseParts{}, fmt.Errorf("no choices in response")
+		}
 
-	choices, ok := result["choices"].([]interface{})
-	if !ok || len(choices) == 0 {
-		return nil, fmt.Errorf("no choices in response")
-	}
+		choice := &result.Choices[0]
+		content := choice.Message.Content
+		if content == "" && len(choice.Message.ToolCalls) == 0 {
+			return chatResponseParts{}, fmt.Errorf("invalid content format")
+		}
 
-	firstChoice, ok := choices[0].(map[string]interface{})
-	if !ok {
-		return nil, fmt.Errorf("invalid choice format")
-	}
-
-	messageMap, ok := firstChoice["message"].(map[string]interface{})
-	if !ok {
-		return nil, fmt.Errorf("invalid message format")
-	}
-
-	toolCalls := extractToolCalls(messageMap)
-	content, hasContent := messageMap["content"].(string)
-	if !hasContent && len(toolCalls) == 0 {
-		return nil, fmt.Errorf("invalid content format")
-	}
-
-	var reasonContent string
-	if chatModelConfig != nil && chatModelConfig.Thinking != nil && *chatModelConfig.Thinking {
-		reasonContent, ok = messageMap["reasoning_content"].(string)
-		if !ok {
-			// If reasoning_content not in response, try parsing from content tags
-			reasoning, answer := GetThinkingAndAnswer(chatModelConfig.ModelClass, &content)
-			if reasoning != nil {
-				reasonContent = *reasoning
-				content = *answer
+		reasonContent := ""
+		if chatConfig != nil && chatConfig.Thinking != nil && *chatConfig.Thinking {
+			reasonContent = choice.Message.ReasoningContent
+			if reasonContent == "" {
+				reasoning, answer := GetThinkingAndAnswer(chatConfig.ModelClass, &content)
+				if reasoning != nil {
+					reasonContent = *reasoning
+					content = *answer
+				}
 			}
-		} else {
-			// if first char of reasonContent is \n remove the '\n'
 			if reasonContent != "" && reasonContent[0] == '\n' {
 				reasonContent = reasonContent[1:]
 			}
 		}
-	}
 
-	chatResponse := &ChatResponse{
-		Answer:        &content,
-		ReasonContent: &reasonContent,
-		ToolCalls:     toolCalls,
-	}
-
-	return chatResponse, nil
+		return chatResponseParts{
+			RequestID:     result.ID,
+			Content:       &content,
+			ReasonContent: &reasonContent,
+			ToolCalls:     choice.Message.ToolCalls,
+			Usage: &TokenUsage{
+				PromptTokens:     result.Usage.PromptTokens,
+				CompletionTokens: result.Usage.CompletionTokens,
+				TotalTokens:      result.Usage.TotalTokens,
+			},
+		}, nil
+	})
 }
 
 // ChatStreamlyWithSender sends messages and streams response via sender function (best performance, no channel)
@@ -186,8 +208,12 @@ func (s *SiliconflowModel) ChatStreamlyWithSender(ctx context.Context, modelName
 
 	// Build request body with streaming enabled
 	reqBody := buildRequestBody(chatModelConfig, modelName, messages, true)
+	reqBody["stream_options"] = map[string]interface{}{"include_usage": true}
 	if chatModelConfig != nil {
 		chatModelConfig.ToolCallsResult = nil
+		if chatModelConfig.Thinking != nil {
+			reqBody["enable_thinking"] = *chatModelConfig.Thinking
+		}
 	}
 	if strings.Contains(strings.ToLower(modelName), "qwen3") && (chatModelConfig == nil || chatModelConfig.Thinking == nil) {
 		reqBody["enable_thinking"] = false
@@ -226,6 +252,14 @@ func (s *SiliconflowModel) ChatStreamlyWithSender(ctx context.Context, modelName
 	done, err := ParseSSEStream[map[string]interface{}](resp.Body, func(event map[string]interface{}) error {
 		common.Info(fmt.Sprintf("%v", event))
 
+		tokenUsage, found, usageErr := decodeOpenAICompatibleStreamUsage(event)
+		if usageErr != nil {
+			return usageErr
+		}
+		if found {
+			applyStreamUsage(chatModelConfig, modelUsage, tokenUsage)
+		}
+
 		choices, ok := event["choices"].([]interface{})
 		if !ok || len(choices) == 0 {
 			return nil
@@ -281,22 +315,18 @@ func (s *SiliconflowModel) ChatStreamlyWithSender(ctx context.Context, modelName
 }
 
 type siliconflowEmbeddingResponse struct {
-	Object string                     `json:"object"`
-	Model  string                     `json:"model"`
-	Data   []siliconflowEmbeddingData `json:"data"`
-	Usage  siliconflowUsage           `json:"usage"`
-}
-
-type siliconflowEmbeddingData struct {
-	Object    string    `json:"object"`
-	Embedding []float64 `json:"embedding"`
-	Index     int       `json:"index"`
-}
-
-type siliconflowUsage struct {
-	PromptTokens     int `json:"prompt_tokens"`
-	CompletionTokens int `json:"completion_tokens"`
-	TotalTokens      int `json:"total_tokens"`
+	ID     string `json:"id"`
+	Object string `json:"object"`
+	Model  string `json:"model"`
+	Data   []struct {
+		Object    string    `json:"object"`
+		Embedding []float64 `json:"embedding"`
+		Index     int       `json:"index"`
+	} `json:"data"`
+	Usage struct {
+		PromptTokens int `json:"prompt_tokens"`
+		TotalTokens  int `json:"total_tokens"`
+	} `json:"usage"`
 }
 
 // siliconflowMaxBatchSize is the per-request input limit documented at
@@ -381,6 +411,10 @@ func (s *SiliconflowModel) Embed(ctx context.Context, modelName *string, texts [
 		embeddingData.Index = dataElem.Index
 		embeddings = append(embeddings, embeddingData)
 	}
+	recordResponseUsage(modelUsage, parsed.ID, &TokenUsage{
+		PromptTokens: parsed.Usage.PromptTokens,
+		TotalTokens:  parsed.Usage.TotalTokens,
+	}, "embedding")
 
 	return embeddings, nil
 }
@@ -572,12 +606,15 @@ func (s *SiliconflowModel) Rerank(ctx context.Context, modelName *string, query
 	if len(documents) == 0 {
 		return &RerankResponse{}, nil
 	}
+	if modelName == nil || *modelName == "" {
+		return nil, fmt.Errorf("model name is required")
+	}
 
 	apiKey := *apiConfig.ApiKey
 
-	var topN = rerankConfig.TopN
-	if rerankConfig.TopN == 0 {
-		topN = len(documents)
+	topN := len(documents)
+	if rerankConfig != nil && rerankConfig.TopN > 0 {
+		topN = rerankConfig.TopN
 	}
 
 	reqBody := SiliconflowRerankRequest{
@@ -639,6 +676,11 @@ func (s *SiliconflowModel) Rerank(ctx context.Context, modelName *string, query
 			RelevanceScore: result.RelevanceScore,
 		})
 	}
+	recordResponseUsage(modelUsage, siliconflowRerankResp.ID, &TokenUsage{
+		PromptTokens:     siliconflowRerankResp.Meta.Tokens.InputTokens,
+		CompletionTokens: siliconflowRerankResp.Meta.Tokens.OutputTokens,
+		TotalTokens:      siliconflowRerankResp.Meta.Tokens.InputTokens + siliconflowRerankResp.Meta.Tokens.OutputTokens,
+	}, "rerank")
 	return &rerankResponse, nil
 }
 

internal/entity/models/siliconflow_test.go

@@ -17,6 +17,8 @@
 package models
 
 import (
+	"encoding/json"
+	"io"
 	"net/http"
 	"net/http/httptest"
 	"testing"
@@ -47,3 +49,109 @@ func TestSiliconflowChatRejectsMissingContent(t *testing.T) {
 		t.Fatal("expected missing content error")
 	}
 }
+
+func TestSiliconflowChatWithMessagesExtractsResponseAndUsage(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		_ = json.NewEncoder(w).Encode(map[string]any{
+			"id":      "chatcmpl-siliconflow",
+			"object":  "chat.completion",
+			"created": 1,
+			"model":   "Qwen/Qwen3-8B",
+			"choices": []map[string]any{{
+				"index": 0,
+				"message": map[string]any{
+					"role":              "assistant",
+					"content":           "answer",
+					"reasoning_content": "\nreasoning",
+				},
+				"finish_reason": "stop",
+			}},
+			"usage": map[string]any{
+				"prompt_tokens":     3,
+				"completion_tokens": 5,
+				"total_tokens":      8,
+				"completion_tokens_details": map[string]any{
+					"reasoning_tokens": 2,
+				},
+				"prompt_tokens_details": map[string]any{
+					"cached_tokens": 1,
+				},
+			},
+		})
+	}))
+	defer server.Close()
+
+	apiKey := "test-key"
+	thinking := true
+	response, err := NewSiliconflowModel(
+		map[string]string{"default": server.URL},
+		URLSuffix{Chat: "chat/completions"},
+	).ChatWithMessages(
+		t.Context(),
+		"Qwen/Qwen3-8B",
+		[]Message{{Role: "user", Content: "hi"}},
+		&APIConfig{ApiKey: &apiKey},
+		&ChatConfig{Thinking: &thinking},
+		nil,
+	)
+	if err != nil {
+		t.Fatalf("ChatWithMessages: %v", err)
+	}
+	if response.Answer == nil || *response.Answer != "answer" {
+		t.Fatalf("Answer=%#v, want answer", response.Answer)
+	}
+	if response.ReasonContent == nil || *response.ReasonContent != "reasoning" {
+		t.Fatalf("ReasonContent=%#v, want reasoning", response.ReasonContent)
+	}
+	if response.Usage == nil || response.Usage.PromptTokens != 3 || response.Usage.CompletionTokens != 5 || response.Usage.TotalTokens != 8 {
+		t.Fatalf("Usage=%#v, want prompt=3 completion=5 total=8", response.Usage)
+	}
+}
+
+func TestSiliconflowChatStreamlyWithSenderCollectsUsage(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		var requestBody map[string]any
+		if err := json.NewDecoder(r.Body).Decode(&requestBody); err != nil {
+			t.Errorf("decode request: %v", err)
+			return
+		}
+		if requestBody["enable_thinking"] != true {
+			t.Errorf("enable_thinking=%#v, want true", requestBody["enable_thinking"])
+		}
+		w.Header().Set("Content-Type", "text/event-stream")
+		_, _ = io.WriteString(w, "data: {\"choices\":[{\"delta\":{\"content\":\"answer\"}}]}\n\n")
+		_, _ = io.WriteString(w, "data: {\"choices\":[],\"usage\":{\"prompt_tokens\":3,\"completion_tokens\":5,\"total_tokens\":8}}\n\n")
+		_, _ = io.WriteString(w, "data: [DONE]\n\n")
+	}))
+	defer server.Close()
+
+	apiKey := "test-key"
+	thinking := true
+	config := &ChatConfig{Thinking: &thinking}
+	err := NewSiliconflowModel(
+		map[string]string{"default": server.URL},
+		URLSuffix{Chat: "chat/completions"},
+	).ChatStreamlyWithSender(
+		t.Context(),
+		"Qwen/Qwen3-8B",
+		[]Message{{Role: "user", Content: "hi"}},
+		&APIConfig{ApiKey: &apiKey},
+		config,
+		nil,
+		func(content *string, reasoning *string) error {
+			if reasoning != nil {
+				t.Errorf("reasoning=%q", *reasoning)
+			}
+			if content != nil && *content != "answer" && *content != "[DONE]" {
+				t.Errorf("content=%q", *content)
+			}
+			return nil
+		},
+	)
+	if err != nil {
+		t.Fatalf("ChatStreamlyWithSender: %v", err)
+	}
+	if config.UsageResult == nil || config.UsageResult.PromptTokens != 3 || config.UsageResult.CompletionTokens != 5 || config.UsageResult.TotalTokens != 8 {
+		t.Fatalf("UsageResult=%#v, want prompt=3 completion=5 total=8", config.UsageResult)
+	}
+}

internal/entity/models/volcengine.go

@@ -52,6 +52,37 @@ func (v *VolcEngine) Name() string {
 	return "volcengine"
 }
 
+// VolcEngineChatResponse mirrors Ark's OpenAI-compatible chat response.
+type VolcEngineChatResponse struct {
+	ID      string `json:"id"`
+	Object  string `json:"object"`
+	Created int64  `json:"created"`
+	Model   string `json:"model"`
+	Choices []struct {
+		FinishReason string `json:"finish_reason"`
+		Index        int    `json:"index"`
+		Logprobs     any    `json:"logprobs"`
+		Message      struct {
+			Content          string           `json:"content"`
+			ReasoningContent string           `json:"reasoning_content"`
+			Role             string           `json:"role"`
+			ToolCalls        []map[string]any `json:"tool_calls"`
+		} `json:"message"`
+	} `json:"choices"`
+	SystemFingerprint string `json:"system_fingerprint"`
+	Usage             struct {
+		CompletionTokens        int `json:"completion_tokens"`
+		PromptTokens            int `json:"prompt_tokens"`
+		TotalTokens             int `json:"total_tokens"`
+		CompletionTokensDetails struct {
+			ReasoningTokens int `json:"reasoning_tokens"`
+		} `json:"completion_tokens_details"`
+		PromptTokensDetails struct {
+			CachedTokens int `json:"cached_tokens"`
+		} `json:"prompt_tokens_details"`
+	} `json:"usage"`
+}
+
 // getAPIKey extracts the actual API key from VolcEngine's stored format.
 // VolcEngine stores the api_key as a JSON string like:
 //
@@ -165,61 +196,44 @@ func (v *VolcEngine) ChatWithMessages(ctx context.Context, modelName string, mes
 		return nil, fmt.Errorf("API request failed with status %d: %s", resp.StatusCode, string(body))
 	}
 
-	// Parse response
-	var result map[string]interface{}
-	if err = json.Unmarshal(body, &result); err != nil {
-		return nil, fmt.Errorf("failed to unmarshal response: %w", err)
-	}
+	return parseChatCompletionResponse(body, chatModelConfig, modelUsage, func(body []byte, chatConfig *ChatConfig) (chatResponseParts, error) {
+		var result VolcEngineChatResponse
+		if err := json.Unmarshal(body, &result); err != nil {
+			return chatResponseParts{}, fmt.Errorf("failed to unmarshal response: %w", err)
+		}
+		if len(result.Choices) == 0 {
+			return chatResponseParts{}, fmt.Errorf("no choices in response")
+		}
 
-	choices, ok := result["choices"].([]interface{})
-	if !ok || len(choices) == 0 {
-		return nil, fmt.Errorf("no choices in response")
-	}
-
-	firstChoice, ok := choices[0].(map[string]interface{})
-	if !ok {
-		return nil, fmt.Errorf("invalid choice format")
-	}
-
-	messageMap, ok := firstChoice["message"].(map[string]interface{})
-	if !ok {
-		return nil, fmt.Errorf("invalid message format")
-	}
-
-	content, _ := messageMap["content"].(string)
-
-	var reasonContent string
-	if chatModelConfig != nil && chatModelConfig.Thinking != nil && *chatModelConfig.Thinking {
-		reasonContent, ok = messageMap["reasoning_content"].(string)
-		if !ok {
-			return nil, fmt.Errorf("invalid reasonContent format")
+		choice := &result.Choices[0]
+		content := choice.Message.Content
+		if content == "" && len(choice.Message.ToolCalls) == 0 {
+			return chatResponseParts{}, fmt.Errorf("invalid content format")
+		}
+		reasonContent := choice.Message.ReasoningContent
+		if chatConfig != nil && chatConfig.Thinking != nil && *chatConfig.Thinking && reasonContent == "" {
+			reasoning, answer := GetThinkingAndAnswer(chatConfig.ModelClass, &content)
+			if reasoning != nil {
+				reasonContent = *reasoning
+				content = *answer
+			}
 		}
 		if reasonContent != "" && reasonContent[0] == '\n' {
 			reasonContent = reasonContent[1:]
 		}
-	}
 
-	var toolCalls []map[string]any
-	if tcs, ok := messageMap["tool_calls"].([]any); ok {
-		for _, tc := range tcs {
-			if toolCall, ok := tc.(map[string]any); ok {
-				toolCalls = append(toolCalls, toolCall)
-			}
-		}
-	}
-
-	chatResponse := &ChatResponse{
-		Answer:        &content,
-		ReasonContent: &reasonContent,
-		ToolCalls:     toolCalls,
-	}
-	if pt, ct, tt := extractUsageFromMap(result); tt > 0 {
-		chatResponse.Usage = &TokenUsage{
-			PromptTokens: pt, CompletionTokens: ct, TotalTokens: tt,
-		}
-	}
-
-	return chatResponse, nil
+		return chatResponseParts{
+			RequestID:     result.ID,
+			Content:       &content,
+			ReasonContent: &reasonContent,
+			ToolCalls:     choice.Message.ToolCalls,
+			Usage: &TokenUsage{
+				PromptTokens:     result.Usage.PromptTokens,
+				CompletionTokens: result.Usage.CompletionTokens,
+				TotalTokens:      result.Usage.TotalTokens,
+			},
+		}, nil
+	})
 }
 
 // ChatStreamlyWithSender sends messages and streams response via sender function (best performance, no channel)
@@ -240,6 +254,7 @@ func (v *VolcEngine) ChatStreamlyWithSender(ctx context.Context, modelName strin
 
 	// Build request body with streaming enabled
 	reqBody := buildRequestBody(modelConfig, modelName, messages, true)
+	reqBody["stream_options"] = map[string]interface{}{"include_usage": true}
 
 	if modelConfig != nil {
 		// TODO VolcEngine has `auto` mode
@@ -275,6 +290,10 @@ func (v *VolcEngine) ChatStreamlyWithSender(ctx context.Context, modelName strin
 					thinkingFlag = "enabled"
 					reqBody["reasoning_effort"] = "xhigh"
 					break
+				case "max":
+					thinkingFlag = "enabled"
+					reqBody["reasoning_effort"] = "max"
+					break
 				default:
 					return fmt.Errorf("invalid effort level")
 				}
@@ -318,9 +337,18 @@ func (v *VolcEngine) ChatStreamlyWithSender(ctx context.Context, modelName strin
 	}
 
 	accumulatedToolCalls := make(map[int]map[string]any)
-	if _, err := ParseSSEStream[map[string]interface{}](resp.Body, func(event map[string]interface{}) error {
+	sawTerminal := false
+	done, err := ParseSSEStream[map[string]interface{}](resp.Body, func(event map[string]interface{}) error {
 		common.Info(fmt.Sprintf("%v", event))
 
+		tokenUsage, found, usageErr := decodeOpenAICompatibleStreamUsage(event)
+		if usageErr != nil {
+			return usageErr
+		}
+		if found {
+			applyStreamUsage(modelConfig, modelUsage, tokenUsage)
+		}
+
 		choices, ok := event["choices"].([]interface{})
 		if !ok || len(choices) == 0 {
 			return nil
@@ -330,6 +358,9 @@ func (v *VolcEngine) ChatStreamlyWithSender(ctx context.Context, modelName strin
 		if !ok {
 			return nil
 		}
+		if finishReason, ok := firstChoice["finish_reason"].(string); ok && finishReason != "" {
+			sawTerminal = true
+		}
 
 		delta, ok := firstChoice["delta"].(map[string]interface{})
 		if !ok {
@@ -401,9 +432,13 @@ func (v *VolcEngine) ChatStreamlyWithSender(ctx context.Context, modelName strin
 		}
 
 		return nil
-	}); err != nil {
+	})
+	if err != nil {
 		return fmt.Errorf("failed to scan response body: %w", err)
 	}
+	if !done && !sawTerminal {
+		return fmt.Errorf("volcengine: stream ended before [DONE] or finish_reason")
+	}
 
 	if len(accumulatedToolCalls) > 0 && modelConfig != nil {
 		indices := make([]int, 0, len(accumulatedToolCalls))
@@ -461,6 +496,9 @@ func (v *VolcEngine) Embed(ctx context.Context, modelName *string, texts []strin
 	if len(texts) == 0 {
 		return []EmbeddingData{}, nil
 	}
+	if modelName == nil || *modelName == "" {
+		return nil, fmt.Errorf("model name is required")
+	}
 
 	resolvedBaseURL, err := v.baseModel.GetBaseURL(apiConfig)
 	if err != nil {
@@ -469,6 +507,8 @@ func (v *VolcEngine) Embed(ctx context.Context, modelName *string, texts []strin
 	url := fmt.Sprintf("%s/%s", resolvedBaseURL, v.baseModel.URLSuffix.Embedding)
 
 	var embeddings []EmbeddingData
+	var totalUsage TokenUsage
+	var requestID string
 
 	for i, text := range texts {
 
@@ -493,7 +533,6 @@ func (v *VolcEngine) Embed(ctx context.Context, modelName *string, texts []strin
 				err,
 			)
 		}
-
 		// Run each per-text request in its own scope so the context's
 		// deadline is cancelled at the end of every iteration instead of
 		// piling up deferred cancels until the whole batch finishes.
@@ -534,12 +573,16 @@ func (v *VolcEngine) Embed(ctx context.Context, modelName *string, texts []strin
 		if err != nil {
 			return nil, err
 		}
+		requestID = parsed.ID
+		totalUsage.PromptTokens += parsed.Usage.PromptTokens
+		totalUsage.TotalTokens += parsed.Usage.TotalTokens
 
 		var embeddingData EmbeddingData
 		embeddingData.Index = i
 		embeddingData.Embedding = parsed.Data.Embedding
 		embeddings = append(embeddings, embeddingData)
 	}
+	recordResponseUsage(modelUsage, requestID, &totalUsage, "embedding")
 
 	return embeddings, nil
 }

internal/entity/models/volcengine_test.go

@@ -2,6 +2,7 @@ package models
 
 import (
 	"encoding/json"
+	"io"
 	"net/http"
 	"net/http/httptest"
 	"os"
@@ -111,3 +112,67 @@ func TestVolcEngineListModelsRequiresModelsSuffix(t *testing.T) {
 		t.Fatalf("expected missing models suffix error, got %v", err)
 	}
 }
+
+func TestVolcEngineChatStreamSupportsMaxEffortAndUsage(t *testing.T) {
+	ctx := t.Context()
+	srv := newVolcEngineServer(t, func(t *testing.T, r *http.Request, w http.ResponseWriter) {
+		var body map[string]interface{}
+		if err := json.NewDecoder(r.Body).Decode(&body); err != nil {
+			t.Fatalf("decode request: %v", err)
+		}
+		if body["reasoning_effort"] != "max" {
+			t.Errorf("reasoning_effort=%v, want max", body["reasoning_effort"])
+		}
+		streamOptions, ok := body["stream_options"].(map[string]interface{})
+		if !ok || streamOptions["include_usage"] != true {
+			t.Errorf("stream_options=%#v, want include_usage=true", body["stream_options"])
+		}
+		w.Header().Set("Content-Type", "text/event-stream")
+		_, _ = io.WriteString(w, "data: {\"choices\":[{\"delta\":{\"content\":\"answer\"},\"finish_reason\":\"stop\"}]}\n\n")
+		_, _ = io.WriteString(w, "data: {\"choices\":[],\"usage\":{\"prompt_tokens\":3,\"completion_tokens\":5,\"total_tokens\":8}}\n\n")
+		_, _ = io.WriteString(w, "data: [DONE]\n\n")
+	})
+	defer srv.Close()
+
+	apiKey := "test-key"
+	thinking := true
+	effort := "max"
+	config := &ChatConfig{Thinking: &thinking, Effort: &effort}
+	if err := newVolcEngineForTest(srv.URL).ChatStreamlyWithSender(
+		ctx,
+		"doubao-seed-2-0-pro-260215",
+		[]Message{{Role: "user", Content: "hello"}},
+		&APIConfig{ApiKey: &apiKey},
+		config,
+		nil,
+		func(*string, *string) error { return nil },
+	); err != nil {
+		t.Fatalf("ChatStreamlyWithSender: %v", err)
+	}
+	if config.UsageResult == nil || config.UsageResult.TotalTokens != 8 {
+		t.Fatalf("UsageResult=%#v, want total tokens 8", config.UsageResult)
+	}
+}
+
+func TestVolcEngineChatStreamRejectsTruncatedResponse(t *testing.T) {
+	ctx := t.Context()
+	srv := newVolcEngineServer(t, func(t *testing.T, _ *http.Request, w http.ResponseWriter) {
+		w.Header().Set("Content-Type", "text/event-stream")
+		_, _ = io.WriteString(w, "data: {\"choices\":[{\"delta\":{\"content\":\"partial\"}}]}\n\n")
+	})
+	defer srv.Close()
+
+	apiKey := "test-key"
+	err := newVolcEngineForTest(srv.URL).ChatStreamlyWithSender(
+		ctx,
+		"doubao-seed-2-0-pro-260215",
+		[]Message{{Role: "user", Content: "hello"}},
+		&APIConfig{ApiKey: &apiKey},
+		&ChatConfig{},
+		nil,
+		func(*string, *string) error { return nil },
+	)
+	if err == nil || !strings.Contains(err.Error(), "stream ended before [DONE] or finish_reason") {
+		t.Fatalf("error=%v, want truncated stream error", err)
+	}
+}