Skip to content

RAGFlow is an agent-native database: a Retrieval-Augmented Generation (RAG) engine that grew into context infrastructure.

fix(models): infer model types on catalog miss so Qwen VLMs show LLM+VLM tags (#18571)

euvre7d8143850353
diff --git a/internal/entity/models/base_model.go b/internal/entity/models/base_model.go
index 8e29359328..5018e19ce2 100644
--- a/internal/entity/models/base_model.go
+++ b/internal/entity/models/base_model.go
@@ -383,6 +383,11 @@ func ParseSSEStreamTolerant[T any](r io.Reader, onEvent func(event T) error) (do
// ParseListModel Parse model list. Empty/whitespace IDs are skipped so
// upstream typos do not surface as blank entries in the UI.
+//
+// Entries the catalog cannot type fall back to name-based inference
+// (InferModelTypes), mirroring Python's
+// OpenAIAPICompatible._format_model_list (rag/llm/model_meta.py) so remote
+// entries never surface type-less.
func ParseListModel(modelList ModelList) []ListModelResponse {
var models []ListModelResponse
pm := GetProviderManager()
@@ -405,7 +410,18 @@ func ParseListModel(modelList ModelList) []ListModelResponse {
modelResponse.MaxOutput = modelEntity.MaxOutput
modelResponse.ModelTypes = modelEntity.ModelTypes
modelResponse.Thinking = modelEntity.Thinking
- modelResponse.Dimensions = modelEntity.Dimensions
+ }
+
+ // The provider-list merge treats remote entries as authoritative
+ // (internal/handler/providers.go) and the instance save path
+ // persists whatever types this list carries, so a catalog miss
+ // must not leave ModelTypes empty — the UI renders type-less
+ // models with an LLM-only badge. Infer types from the model name
+ // (vision models like qwen-vl-plus keep their VLM tag even before
+ // the catalog knows them); InferModelTypes always returns at
+ // least ["chat"].
+ if len(modelResponse.ModelTypes) == 0 {
+ modelResponse.ModelTypes = InferModelTypes(modelName)
}
models = append(models, modelResponse)
diff --git a/internal/entity/models/base_model_test.go b/internal/entity/models/base_model_test.go
new file mode 100644
index 0000000000..9312060006
--- /dev/null
+++ b/internal/entity/models/base_model_test.go
@@ -0,0 +1,83 @@
+//
+// 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 (
+ "slices"
+ "testing"
+)
+
+// ParseListModel feeds the provider-model list endpoint whose merge treats
+// remote entries as authoritative: every returned entry must carry non-empty
+// model types. Catalog hits keep the catalog types; catalog misses fall back
+// to name-based inference (defaulting to chat), mirroring Python's
+// OpenAIAPICompatible._format_model_list.
+func TestParseListModelAssignsModelTypes(t *testing.T) {
+ dir, restore := setupProviderTestDir(t, "aliyun.json")
+ defer restore()
+
+ if err := InitProviderManager(dir); err != nil {
+ t.Fatalf("InitProviderManager: %v", err)
+ }
+
+ got := ParseListModel(ModelList{Models: []ModelListItem{
+ {ID: "qwen3-vl-plus"}, // catalog hit: chat+vision+ocr
+ {ID: "qwen-vl-plus"}, // catalog hit: chat+vision
+ {ID: "qwen3-vl-flash-2099-01-01"}, // miss: future dated variant -> inferred
+ {ID: "gpt-4o-2099-01-01"}, // miss -> inferred
+ {ID: "text-embedding-v4"}, // miss -> inferred
+ {ID: "qwen3-reranker-2099"}, // miss -> inferred
+ {ID: "some-unknown-model"}, // miss -> default chat
+ {ID: " qwen-vl-max-2099-01-01 "}, // trimmed, miss -> inferred
+ {ID: " "}, // skipped
+ }})
+
+ want := map[string][]string{
+ "qwen3-vl-plus": {"chat", "vision", "ocr"}, // catalog wins over inference
+ "qwen-vl-plus": {"chat", "vision"},
+ "qwen3-vl-flash-2099-01-01": {"chat", "vision"},
+ "gpt-4o-2099-01-01": {"chat", "vision"},
+ "text-embedding-v4": {"embedding"},
+ "qwen3-reranker-2099": {"rerank"},
+ "some-unknown-model": {"chat"},
+ "qwen-vl-max-2099-01-01": {"chat", "vision"},
+ }
+
+ if len(got) != len(want) {
+ t.Fatalf("ParseListModel returned %d entries, want %d (blank IDs skipped)", len(got), len(want))
+ }
+
+ for _, m := range got {
+ wantTypes, ok := want[m.Name]
+ if !ok {
+ t.Fatalf("unexpected entry %q in result", m.Name)
+ }
+ if !slices.Equal(m.ModelTypes, wantTypes) {
+ t.Errorf("model %q: ModelTypes = %v, want %v", m.Name, m.ModelTypes, wantTypes)
+ }
+ }
+
+ // Catalog metadata still flows through on a hit (qwen3-vl-plus declares
+ // max_output 8192 in conf/all_models.json).
+ for _, m := range got {
+ if m.Name == "qwen3-vl-plus" {
+ if m.MaxOutput == nil || *m.MaxOutput != 8192 {
+ t.Errorf("qwen3-vl-plus: MaxOutput = %v, want 8192", m.MaxOutput)
+ }
+ }
+ }
+}
diff --git a/internal/entity/models/gitee_test.go b/internal/entity/models/gitee_test.go
index 0c19d93f33..f894a2e18b 100644
--- a/internal/entity/models/gitee_test.go
+++ b/internal/entity/models/gitee_test.go
@@ -184,8 +184,10 @@ func TestGiteeListModelsMapsAllDeepSeekAliasesToModelMetadata(t *testing.T) {
if unknown.MaxOutput != nil {
t.Fatalf("unknown.MaxOutput=%v, want nil", *unknown.MaxOutput)
}
- if len(unknown.ModelTypes) != 0 {
- t.Fatalf("unknown.ModelTypes=%v, want empty", unknown.ModelTypes)
+ // Catalog misses fall back to name-based inference, which defaults to
+ // chat — remote entries must never surface type-less.
+ if len(unknown.ModelTypes) != 1 || unknown.ModelTypes[0] != "chat" {
+ t.Fatalf("unknown.ModelTypes=%v, want [chat]", unknown.ModelTypes)
}
}