RAGFlow is an agent-native database: a Retrieval-Augmented Generation (RAG) engine that grew into context infrastructure.
Go agent-harnessagent-native-databaseagentic-aiagentic-retrievalagentic-searchaiai-agentscontext-enginecontext-engineeringcontext-managementharness-engineeringknowledge-compilationragretrieval-augmented-generationretrieval-harnesssearch-harness
refa: migrate model wrappers to model solver (#20100)
buua436317db0c939de
diff --git a/cmd/ragflow_server.go b/cmd/ragflow_server.goindex 7f08a78d3d..bd889a213a 100644--- a/cmd/ragflow_server.go+++ b/cmd/ragflow_server.go@@ -1130,7 +1130,7 @@ func startServer(ctx context.Context, serverName string, arguments *serverArgs)retrievalService := nlp.NewRetrievalService(docEngine, documentDAO)difyRetrievalHandler := handler.NewDifyRetrievalHandler(datasetsService,- modelProviderService,+ modelSolver,metadataService,retrievalService,documentDAO,diff --git a/internal/handler/dify_retrieval_handler.go b/internal/handler/dify_retrieval_handler.goindex 73dc769a0b..a3d0bf8f77 100644--- a/internal/handler/dify_retrieval_handler.go+++ b/internal/handler/dify_retrieval_handler.go@@ -47,10 +47,10 @@ type KBServiceIface interface {Accessible(ctx context.Context, kbID, userID string) bool}-// ModelServiceIface abstracts ModelProviderService for the Dify handler.-type ModelServiceIface interface {- GetEmbeddingModel(ctx context.Context, tenantID, embdID string) (*modelModule.EmbeddingModel, error)- GetChatModel(ctx context.Context, tenantID, compositeModelName string) (*modelModule.ChatModel, error)+// ModelResolver abstracts model configuration resolution for the Dify handler.+type ModelResolver interface {+ ResolveModelConfig(ctx context.Context, tenantID string, modelType entity.ModelType, modelRef string) (*service.ModelTarget, error)+ ResolveDefaultModelConfig(ctx context.Context, tenantID string, modelType entity.ModelType) (*service.ModelTarget, error)}// MetadataServiceIface abstracts MetadataService for the Dify handler.@@ -131,12 +131,12 @@ type difyRecord struct {// DifyRetrievalHandler handles Dify-compatible retrieval requests.type DifyRetrievalHandler struct {- kbSvc KBServiceIface- modelSvc ModelServiceIface- metadataSvc MetadataServiceIface- retrievalSvc RetrievalServiceIface- docDAO DocumentDAOIface- docEngine engine.DocEngine+ kbSvc KBServiceIface+ modelResolver ModelResolver+ metadataSvc MetadataServiceIface+ retrievalSvc RetrievalServiceIface+ docDAO DocumentDAOIface+ docEngine engine.DocEngine}// NewDifyRetrievalHandler creates a new DifyRetrievalHandler.@@ -144,19 +144,19 @@ type DifyRetrievalHandler struct {// a pipeline that depends on per-request model configuration.func NewDifyRetrievalHandler(kbSvc KBServiceIface,- modelSvc ModelServiceIface,+ modelResolver ModelResolver,metadataSvc MetadataServiceIface,retrievalSvc RetrievalServiceIface,docDAO DocumentDAOIface,docEngine engine.DocEngine,) *DifyRetrievalHandler {return &DifyRetrievalHandler{- kbSvc: kbSvc,- modelSvc: modelSvc,- metadataSvc: metadataSvc,- retrievalSvc: retrievalSvc,- docDAO: docDAO,- docEngine: docEngine,+ kbSvc: kbSvc,+ modelResolver: modelResolver,+ metadataSvc: metadataSvc,+ retrievalSvc: retrievalSvc,+ docDAO: docDAO,+ docEngine: docEngine,}}@@ -235,11 +235,12 @@ func (h *DifyRetrievalHandler) Retrieval(c *gin.Context) {}// Get embedding model- embModel, err := h.modelSvc.GetEmbeddingModel(ctx, kb.TenantID, kb.EmbdID)+ target, err := h.modelResolver.ResolveModelConfig(ctx, kb.TenantID, entity.ModelTypeEmbedding, kb.EmbdID)if err != nil {common.ResponseWithHttpCodeData(c, http.StatusInternalServerError, common.CodeServerError, nil, fmt.Sprintf("failed to get embedding model: %v", err))return}+ embModel := modelModule.NewEmbeddingModel(target.Driver, &target.ModelName, target.APIConfig, target.MaxTokens)// Metadata filtermetas, metaErr := h.metadataSvc.GetFlattedMetaByKBs(ctx, []string{req.KnowledgeID})@@ -291,10 +292,11 @@ func (h *DifyRetrievalHandler) Retrieval(c *gin.Context) {// KG retrieval (optional)if req.UseKG {- chatModel, kgErr := h.modelSvc.GetChatModel(ctx, kb.TenantID, "")+ target, kgErr := h.modelResolver.ResolveDefaultModelConfig(ctx, kb.TenantID, entity.ModelTypeChat)if kgErr != nil {common.Warn("KG retrieval: failed to get chat model", zap.String("kbID", req.KnowledgeID), zap.Error(kgErr))- } else if chatModel != nil {+ } else if target != nil {+ chatModel := modelModule.NewChatModel(target.Driver, &target.ModelName, target.APIConfig)kgPipeline := graph.NewPipeline(h.docEngine,[]string{req.KnowledgeID},diff --git a/internal/handler/dify_retrieval_handler_test.go b/internal/handler/dify_retrieval_handler_test.goindex 781ca3f3f1..ace1ec4878 100644--- a/internal/handler/dify_retrieval_handler_test.go+++ b/internal/handler/dify_retrieval_handler_test.go@@ -31,7 +31,6 @@ import ("ragflow/internal/engine""ragflow/internal/engine/types""ragflow/internal/entity"- modelModule "ragflow/internal/entity/models""ragflow/internal/service""ragflow/internal/service/nlp"@@ -62,24 +61,23 @@ func (m *mockKBService) Accessible(ctx context.Context, kbID, userID string) booreturn true}-type mockModelService struct {- ModelServiceIface- getEmbeddingFn func(ctx context.Context, tenantID, embdID string) (*modelModule.EmbeddingModel, error)- getChatModelFn func(ctx context.Context, tenantID, llmID string) (*modelModule.ChatModel, error)+type mockModelResolver struct {+ resolveModelConfigFn func(ctx context.Context, tenantID string, modelType entity.ModelType, modelRef string) (*service.ModelTarget, error)+ resolveDefaultModelConfigFn func(ctx context.Context, tenantID string, modelType entity.ModelType) (*service.ModelTarget, error)}-func (m *mockModelService) GetEmbeddingModel(ctx context.Context, tenantID, embdID string) (*modelModule.EmbeddingModel, error) {- if m.getEmbeddingFn != nil {- return m.getEmbeddingFn(ctx, tenantID, embdID)+func (m *mockModelResolver) ResolveModelConfig(ctx context.Context, tenantID string, modelType entity.ModelType, modelRef string) (*service.ModelTarget, error) {+ if m.resolveModelConfigFn != nil {+ return m.resolveModelConfigFn(ctx, tenantID, modelType, modelRef)}- return &modelModule.EmbeddingModel{}, nil+ return &service.ModelTarget{ModelName: "test-model"}, nil}-func (m *mockModelService) GetChatModel(ctx context.Context, tenantID, llmID string) (*modelModule.ChatModel, error) {- if m.getChatModelFn != nil {- return m.getChatModelFn(ctx, tenantID, llmID)+func (m *mockModelResolver) ResolveDefaultModelConfig(ctx context.Context, tenantID string, modelType entity.ModelType) (*service.ModelTarget, error) {+ if m.resolveDefaultModelConfigFn != nil {+ return m.resolveDefaultModelConfigFn(ctx, tenantID, modelType)}- return &modelModule.ChatModel{}, nil+ return &service.ModelTarget{ModelName: "test-model"}, nil}type mockMetadataService struct {@@ -163,12 +161,12 @@ func (m *mockDocEngine) GetChunk(ctx context.Context, _, _ string, _ []string) (func setupDifyTest(userID string) (*DifyRetrievalHandler, *gin.Engine) {h := &DifyRetrievalHandler{- kbSvc: &mockKBService{},- modelSvc: &mockModelService{},- metadataSvc: &mockMetadataService{},- retrievalSvc: &mockRetrievalService{},- docDAO: &mockDocDAO{},- docEngine: &mockDocEngine{},+ kbSvc: &mockKBService{},+ modelResolver: &mockModelResolver{},+ metadataSvc: &mockMetadataService{},+ retrievalSvc: &mockRetrievalService{},+ docDAO: &mockDocDAO{},+ docEngine: &mockDocEngine{},}gin.SetMode(gin.TestMode)@@ -184,12 +182,12 @@ func setupDifyTest(userID string) (*DifyRetrievalHandler, *gin.Engine) {func setupDifyTestNoAuth() (*DifyRetrievalHandler, *gin.Engine) {h := &DifyRetrievalHandler{- kbSvc: &mockKBService{},- modelSvc: &mockModelService{},- metadataSvc: &mockMetadataService{},- retrievalSvc: &mockRetrievalService{},- docDAO: &mockDocDAO{},- docEngine: &mockDocEngine{},+ kbSvc: &mockKBService{},+ modelResolver: &mockModelResolver{},+ metadataSvc: &mockMetadataService{},+ retrievalSvc: &mockRetrievalService{},+ docDAO: &mockDocDAO{},+ docEngine: &mockDocEngine{},}gin.SetMode(gin.TestMode)r := gin.New()diff --git a/internal/ingestion/component/tokenizer.go b/internal/ingestion/component/tokenizer.goindex 02f9a8dd0e..da393ae60f 100644--- a/internal/ingestion/component/tokenizer.go+++ b/internal/ingestion/component/tokenizer.go@@ -38,7 +38,7 @@// - EMBEDDING MODEL RESOLUTION: mirrored. Python uses// `LLMBundle(tenant_id, embd_id).encode([...])` from// `rag/flow/tokenizer/tokenizer.py:54-66`; the Go port goes-// through `service.ModelProviderService.GetEmbeddingModel`+// through the model solver injected by the ingestion task package// (callers inject the resolver, see `DefaultEmbedderResolver`).// The component does NOT directly construct a model driver —// the resolution path depends on tenant/DAO context that livesdiff --git a/internal/ingestion/task/embedder.go b/internal/ingestion/task/embedder.goindex 56c7f6af41..0effbb5276 100644--- a/internal/ingestion/task/embedder.go+++ b/internal/ingestion/task/embedder.go@@ -35,6 +35,7 @@ import ("ragflow/internal/common""ragflow/internal/dao"+ "ragflow/internal/entity""ragflow/internal/entity/models"componentpkg "ragflow/internal/ingestion/component""ragflow/internal/service"@@ -686,6 +687,7 @@ func newEmbedderResolver(// so the concrete resolver is injected here - the task package is the// composition root for ingestion runs.func init() {+ modelSolver := service.NewModelSolver()componentpkg.DefaultEmbedderResolver = newEmbedderResolver(func(ctx context.Context, kbID string) (string, error) {kb, err := dao.NewKnowledgebaseDAO().GetByID(ctx, dao.DB, kbID)@@ -697,6 +699,12 @@ func init() {}return kb.EmbdID, nil},- service.NewModelProviderService().GetEmbeddingModel,+ func(ctx context.Context, tenantID, embdID string) (*models.EmbeddingModel, error) {+ target, err := modelSolver.ResolveModelConfig(ctx, tenantID, entity.ModelTypeEmbedding, embdID)+ if err != nil {+ return nil, err+ }+ return models.NewEmbeddingModel(target.Driver, &target.ModelName, target.APIConfig, target.MaxTokens), nil+ },)}diff --git a/internal/ingestion/task/knowledge_compiler_wiring.go b/internal/ingestion/task/knowledge_compiler_wiring.goindex ca2b8ba66e..eceb7f28b5 100644--- a/internal/ingestion/task/knowledge_compiler_wiring.go+++ b/internal/ingestion/task/knowledge_compiler_wiring.go@@ -714,9 +714,8 @@ const kcChatAttemptTimeout = 20 * time.Minute// kcChatRetryDelay is the initial exponential-backoff delay between retries.const kcChatRetryDelay = 2 * time.Second-// kcEmbedder adapts service.ModelProviderService.GetEmbeddingModel to the-// knowledge_compiler Embedder seam. Vectors are returned as []float32 to match-// the component's product schema.+// kcEmbedder adapts model resolution to the knowledge_compiler Embedder seam.+// Vectors are returned as []float32 to match the component's product schema.type kcEmbedder struct {svc *service.ModelProviderServicesolver *service.ModelSolver@@ -800,14 +799,14 @@ func (e *kcEmbedder) Encode(ctx context.Context, texts []string) ([][]float32, e// loudly instead of silently producing empty vectors.func (e *kcEmbedder) resolveModel(ctx context.Context) (*models.EmbeddingModel, error) {if embdID := strings.TrimSpace(e.embdID); embdID != "" {- mdl, err := e.svc.GetEmbeddingModel(ctx, e.tenantID, embdID)+ target, err := e.solver.ResolveModelConfig(ctx, e.tenantID, entity.ModelTypeEmbedding, embdID)if err != nil {return nil, fmt.Errorf("knowledge_compiler: resolve embedding model: %w", err)}- if mdl == nil || mdl.ModelDriver == nil {+ if target == nil || target.Driver == nil || target.ModelName == "" {return nil, fmt.Errorf("knowledge_compiler: embedding model %q is unavailable", embdID)}- return mdl, nil+ return models.NewEmbeddingModel(target.Driver, &target.ModelName, target.APIConfig, target.MaxTokens), nil}target, err := e.solver.ResolveDefaultModelConfig(ctx, e.tenantID, entity.ModelTypeEmbedding)if err != nil {@@ -816,7 +815,7 @@ func (e *kcEmbedder) resolveModel(ctx context.Context) (*models.EmbeddingModel,if target == nil || target.Driver == nil || target.ModelName == "" {return nil, fmt.Errorf("knowledge_compiler: embedding_model is required (tenant default embedding model unavailable)")}- return &models.EmbeddingModel{ModelDriver: target.Driver, ModelName: &target.ModelName, APIConfig: target.APIConfig}, nil+ return models.NewEmbeddingModel(target.Driver, &target.ModelName, target.APIConfig, target.MaxTokens), nil}func (e *kcEmbedder) Dimensions() int { return int(e.dim.Load()) }diff --git a/internal/service/chunk/chunk.go b/internal/service/chunk/chunk.goindex 609e4b4e00..efb59bac77 100644--- a/internal/service/chunk/chunk.go+++ b/internal/service/chunk/chunk.go@@ -1690,7 +1690,11 @@ func (s *ChunkService) getEmbeddingModel(ctx context.Context, tenantID, embdID sif s.getEmbeddingModelFunc != nil {return s.getEmbeddingModelFunc(tenantID, embdID)}- return service.NewModelProviderService().GetEmbeddingModel(ctx, tenantID, embdID)+ target, err := service.NewModelSolver().ResolveModelConfig(ctx, tenantID, entity.ModelTypeEmbedding, embdID)+ if err != nil {+ return nil, err+ }+ return models.NewEmbeddingModel(target.Driver, &target.ModelName, target.APIConfig, target.MaxTokens), nil}func (s *ChunkService) incrementChunkStats(docID, kbID string, tokenNum, chunkNum int64, duration float64) error {diff --git a/internal/service/model_chat.go b/internal/service/model_chat.goindex d3651f3e6b..b3b197a826 100644--- a/internal/service/model_chat.go+++ b/internal/service/model_chat.go@@ -24,6 +24,7 @@ import ("go.uber.org/zap""ragflow/internal/common"+ "ragflow/internal/entity"modelModule "ragflow/internal/entity/models")@@ -35,18 +36,19 @@ const streamDoneSentinel = "[DONE]"var errStreamDone = errors.New("chat stream done")func (m *ModelProviderService) Chat(ctx context.Context, tenantID, modelID string, messages []modelModule.Message, config *modelModule.ChatConfig) (*modelModule.ChatResponse, error) {- chatModel, err := m.GetChatModel(ctx, tenantID, modelID)+ target, err := m.modelSolver().ResolveModelConfig(ctx, tenantID, entity.ModelTypeChat, modelID)if err != nil {return nil, err}- return chatModel.ModelDriver.ChatWithMessages(ctx, *chatModel.ModelName, messages, chatModel.APIConfig, config, nil)+ return target.Driver.ChatWithMessages(ctx, target.ModelName, messages, target.APIConfig, config, nil)}func (m *ModelProviderService) ChatStream(ctx context.Context, tenantID, modelID string, messages []modelModule.Message, config *modelModule.ChatConfig) (<-chan string, <-chan error, error) {- chatModel, err := m.GetChatModel(ctx, tenantID, modelID)+ target, err := m.modelSolver().ResolveModelConfig(ctx, tenantID, entity.ModelTypeChat, modelID)if err != nil {return nil, nil, err}+ chatModel := modelModule.NewChatModel(target.Driver, &target.ModelName, target.APIConfig)ch, errCh := chatStreamWithContext(ctx, chatModel, messages, config)return ch, errCh, nil}diff --git a/internal/service/model_service.go b/internal/service/model_service.goindex 5e2a065081..516c29b341 100644--- a/internal/service/model_service.go+++ b/internal/service/model_service.go@@ -3056,33 +3056,6 @@ func (m *ModelProviderService) ParseFile(ctx context.Context, providerName, instreturn response, common.CodeSuccess, nil}-// GetEmbeddingModel returns an EmbeddingModel wrapper for the given tenant-func (m *ModelProviderService) GetEmbeddingModel(ctx context.Context, tenantID, compositeModelName string) (*modelModule.EmbeddingModel, error) {- target, err := m.modelSolver().ResolveModelConfig(ctx, tenantID, entity.ModelTypeEmbedding, compositeModelName)- if err != nil {- return nil, err- }- return modelModule.NewEmbeddingModel(target.Driver, &target.ModelName, target.APIConfig, target.MaxTokens), nil-}--// GetChatModel returns a ChatModel wrapper for the given tenant-func (m *ModelProviderService) GetChatModel(ctx context.Context, tenantID, compositeModelName string) (*modelModule.ChatModel, error) {- target, err := m.modelSolver().ResolveModelConfig(ctx, tenantID, entity.ModelTypeChat, compositeModelName)- if err != nil {- return nil, err- }- return modelModule.NewChatModel(target.Driver, &target.ModelName, target.APIConfig), nil-}--// GetRerankModel returns a RerankModel wrapper for the given tenant-func (m *ModelProviderService) GetRerankModel(ctx context.Context, tenantID, compositeModelName string) (*modelModule.RerankModel, error) {- target, err := m.modelSolver().ResolveModelConfig(ctx, tenantID, entity.ModelTypeRerank, compositeModelName)- if err != nil {- return nil, err- }- return modelModule.NewRerankModel(target.Driver, &target.ModelName, target.APIConfig, target.MaxTokens), nil-}-type AddModelRequest struct {ProviderName string `json:"provider_name"`InstanceName string `json:"instance_name"`diff --git a/internal/service/nav_embedder.go b/internal/service/nav_embedder.goindex 125b704ddd..ee4594ff99 100644--- a/internal/service/nav_embedder.go+++ b/internal/service/nav_embedder.go@@ -68,11 +68,11 @@ func (e *NavEmbedder) encode(ctx context.Context, tenantID string, texts []strin}model = modelModule.NewEmbeddingModel(target.Driver, &target.ModelName, target.APIConfig, target.MaxTokens)} else {- var err error- model, err = e.modelSvc.GetEmbeddingModel(ctx, tenantID, name)+ target, err := e.modelSvc.modelSolver().ResolveModelConfig(ctx, tenantID, entity.ModelTypeEmbedding, name)if err != nil {return nil, fmt.Errorf("datasetnav: resolve embedding model for tenant %s: %w", tenantID, err)}+ model = modelModule.NewEmbeddingModel(target.Driver, &target.ModelName, target.APIConfig, target.MaxTokens)}nonEmpty := make([]string, 0, len(texts))for _, t := range texts {diff --git a/internal/service/skill_indexer.go b/internal/service/skill_indexer.goindex 2d20172732..b48a1a1b51 100644--- a/internal/service/skill_indexer.go+++ b/internal/service/skill_indexer.go@@ -923,10 +923,11 @@ func (s *SkillIndexerService) generateEmbedding(ctx context.Context, text, embdIreturn nil, fmt.Errorf("embedding model ID not configured")}- embeddingModel, err := s.modelProvider.GetEmbeddingModel(ctx, tenantID, embdID)+ target, err := s.modelProvider.modelSolver().ResolveModelConfig(ctx, tenantID, entity.ModelTypeEmbedding, embdID)if err != nil {return nil, fmt.Errorf("failed to get embedding model: %w", err)}+ embeddingModel := models.NewEmbeddingModel(target.Driver, &target.ModelName, target.APIConfig, target.MaxTokens)// Truncate text to prevent exceeding model's max input lengthmaxLen := embeddingModel.MaxTokens@@ -961,11 +962,12 @@ func (s *SkillIndexerService) generateEmbeddings(ctx context.Context, texts []st}common.Info(fmt.Sprintf("Getting embedding model for %s", embdID))- embeddingModel, err := s.modelProvider.GetEmbeddingModel(ctx, tenantID, embdID)+ target, err := s.modelProvider.modelSolver().ResolveModelConfig(ctx, tenantID, entity.ModelTypeEmbedding, embdID)if err != nil {common.Error(fmt.Sprintf("Failed to get embedding model: %v", err), err)return nil, fmt.Errorf("failed to get embedding model: %w", err)}+ embeddingModel := models.NewEmbeddingModel(target.Driver, &target.ModelName, target.APIConfig, target.MaxTokens)// Truncate texts to prevent exceeding model's max input lengthmaxLen := embeddingModel.MaxTokens@@ -1019,10 +1021,11 @@ func (s *SkillIndexerService) getEmbeddingDimension(ctx context.Context, tenantIreturn 0, fmt.Errorf("embedding model ID not configured")}- embeddingModel, err := s.modelProvider.GetEmbeddingModel(ctx, tenantID, embdID)+ target, err := s.modelProvider.modelSolver().ResolveModelConfig(ctx, tenantID, entity.ModelTypeEmbedding, embdID)if err != nil {return 0, fmt.Errorf("failed to get embedding model: %w", err)}+ embeddingModel := models.NewEmbeddingModel(target.Driver, &target.ModelName, target.APIConfig, target.MaxTokens)// Use simple test text like Python does: embedding_model.encode(["ok"])testText := "ok"diff --git a/internal/service/skill_search.go b/internal/service/skill_search.goindex b380342f1f..593c748efa 100644--- a/internal/service/skill_search.go+++ b/internal/service/skill_search.go@@ -667,10 +667,11 @@ func (s *SkillSearchService) getEmbedding(ctx context.Context, text, embdID, tenreturn nil, fmt.Errorf("embedding model ID not configured")}- embeddingModel, err := s.modelProvider.GetEmbeddingModel(ctx, tenantID, embdID)+ target, err := s.modelProvider.modelSolver().ResolveModelConfig(ctx, tenantID, entity.ModelTypeEmbedding, embdID)if err != nil {return nil, fmt.Errorf("failed to get embedding model: %w", err)}+ embeddingModel := models.NewEmbeddingModel(target.Driver, &target.ModelName, target.APIConfig, target.MaxTokens)// Truncate text to prevent exceeding model's max input lengthmaxLen := embeddingModel.MaxTokens