Skip to content

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

refactor(parser): drop write-only setup fields from picture/audio parsers (#19971)

jay77721f4cf4ba9f167
diff --git a/internal/ingestion/component/parser.go b/internal/ingestion/component/parser.go
index 78ae0e24d6..353ba66b6d 100644
--- a/internal/ingestion/component/parser.go
+++ b/internal/ingestion/component/parser.go
@@ -67,10 +67,10 @@
// - The Python _param.check() business validation
// (parse_method whitelist, conditional lang checks) is mirrored
// by (*ParserComponent).Check() below, which NewParserComponent
-// runs at construction time. The Python flow check() also
-// validates audio/video vlm.llm_id, but Go media_dispatch uses
-// tenant default models (resolveTenantModelByType) rather than
-// setup["vlm"]["llm_id"], so that check is intentionally omitted.
+// runs at construction time. Neither backend validates
+// audio/video vlm.llm_id: Python's check() has no such branch,
+// and the audio model is resolved at dispatch time with a
+// tenant-default fallback.
//
// - NO PERSISTENCE: structured parser items live only in the per-run
// output map.
@@ -210,11 +210,10 @@ func cloneParserSetupValue(value any) any {
// (Python raises ValueError on the first failure).
//
// NOT covered here (intentional):
-// - audio/video vlm.llm_id: Go media_dispatch uses tenant default
-// models (resolveTenantModelByType), not setup["vlm"]["llm_id"].
-// The Python flow check() for vlm.llm_id does not apply — Go
-// never reads that field, and validating it would block every
-// valid audio/video pipeline (see ingestion_pipeline_audio.json).
+// - audio/video vlm.llm_id: Python's check() does not validate it
+// either, and audio dispatch resolves a missing/empty model to
+// the tenant default, so validating it here would only block
+// otherwise valid pipelines (see ingestion_pipeline_audio.json).
func (c *ParserComponent) Check() error {
// PDF family (parser.py:252-261).
if pdf, ok := c.setups["pdf"]; ok {
diff --git a/internal/ingestion/component/parser_check_test.go b/internal/ingestion/component/parser_check_test.go
index 7213e50f90..63d526848f 100644
--- a/internal/ingestion/component/parser_check_test.go
+++ b/internal/ingestion/component/parser_check_test.go
@@ -26,9 +26,9 @@ import (
// TestParserComponent_Check covers the construction-time business
// validation for parser methods that require language configuration.
//
-// Go does NOT validate audio/video vlm.llm_id because media_dispatch
-// resolves tenant default models via resolveTenantModelByType, not
-// setup["vlm"]["llm_id"]. See plan: quantum-forging-curie-sZ_7zRZb.
+// audio/video vlm.llm_id is not validated: Python's check() has no
+// such branch, and audio resolves its model at dispatch time with a
+// tenant-default fallback.
func TestParserComponent_Check(t *testing.T) {
cases := []struct {
name string
@@ -99,17 +99,17 @@ func TestParserComponent_Check(t *testing.T) {
setups: map[string]schema.ParserSetup{"image": {"lang": "English"}},
},
- // --- audio/video: vlm.llm_id NOT validated in Go ---
+ // --- audio/video: vlm.llm_id not validated (matches Python check()) ---
{
- name: "audio: no vlm field → pass (Go uses tenant default ASR)",
+ name: "audio: no vlm field → pass (model falls back to tenant default)",
setups: map[string]schema.ParserSetup{"audio": {"output_format": "text"}},
},
{
- name: "video: no vlm field → pass (Go uses tenant default VISION)",
+ name: "video: no vlm field → pass",
setups: map[string]schema.ParserSetup{"video": {"output_format": "text"}},
},
{
- name: "audio: vlm.llm_id empty → pass (Go ignores vlm.llm_id)",
+ name: "audio: vlm.llm_id empty → pass (empty falls back to tenant default)",
setups: map[string]schema.ParserSetup{"audio": {"vlm": map[string]any{"llm_id": ""}}},
},
diff --git a/internal/parser/parser/audio_parser.go b/internal/parser/parser/audio_parser.go
index fab1c63cf3..8bdf5e0188 100644
--- a/internal/parser/parser/audio_parser.go
+++ b/internal/parser/parser/audio_parser.go
@@ -38,11 +38,8 @@ var audioExtensions = map[string]bool{
"vqf": true, "oggvorbis": true, "ape": true,
}
-// AudioParser handles audio files for transcription. The struct mirrors
-// the configuration from setups["audio"]:output_format and
-// setups["audio"].vlm.llm_id.
+// AudioParser handles audio files for transcription.
type AudioParser struct {
- VLMModelID string // vlm.llm_id — identifies the speech-to-text model
OutputFormat string
}
@@ -52,16 +49,11 @@ func NewAudioParser() *AudioParser {
}
// ConfigureFromSetup reads audio-specific configuration from the
-// parser setup map. It extracts vlm.llm_id and output_format.
+// parser setup map.
func (p *AudioParser) ConfigureFromSetup(setup map[string]any) {
if p == nil || setup == nil {
return
}
- if vlm, ok := setup["vlm"].(map[string]any); ok {
- if llmID, ok := vlm["llm_id"].(string); ok && llmID != "" {
- p.VLMModelID = llmID
- }
- }
if v, ok := setup["output_format"].(string); ok && v != "" {
p.OutputFormat = v
}
@@ -82,7 +74,6 @@ func (p *AudioParser) ParseWithResult(ctx context.Context, filename string, data
}
}
- // OutputFormat and VLMModelID are consumed by maybeDispatchAudio.
outFmt := p.OutputFormat
if outFmt == "" {
outFmt = "text"
diff --git a/internal/parser/parser/audio_parser_test.go b/internal/parser/parser/audio_parser_test.go
index 19f9dbdba4..12e03393fe 100644
--- a/internal/parser/parser/audio_parser_test.go
+++ b/internal/parser/parser/audio_parser_test.go
@@ -26,9 +26,6 @@ func TestAudioParser_ValidExtension(t *testing.T) {
p := NewAudioParser()
p.ConfigureFromSetup(map[string]any{
"output_format": "text",
- "vlm": map[string]any{
- "llm_id": "whisper-1",
- },
})
result := p.ParseWithResult(ctx, "test.mp3", []byte{1, 2, 3, 4})
@@ -66,18 +63,6 @@ func TestAudioParser_InvalidExtension(t *testing.T) {
}
}
-func TestAudioParser_ConfigureVLM(t *testing.T) {
- p := NewAudioParser()
- p.ConfigureFromSetup(map[string]any{
- "vlm": map[string]any{
- "llm_id": "openai-whisper",
- },
- })
- if p.VLMModelID != "openai-whisper" {
- t.Errorf("VLMModelID = %q, want openai-whisper", p.VLMModelID)
- }
-}
-
func TestAudioParser_ConfigureNilSafe(t *testing.T) {
ctx := t.Context()
p := NewAudioParser()
diff --git a/internal/parser/parser/picture_parser.go b/internal/parser/parser/picture_parser.go
index 558ea70e6c..9148080038 100644
--- a/internal/parser/parser/picture_parser.go
+++ b/internal/parser/parser/picture_parser.go
@@ -51,17 +51,8 @@ var imageExtensions = map[string]bool{
}
// PictureParser handles image files for OCR and VLM description.
-// Mirrors the configuration from setups["picture"]:
-// - VLMModelID ← setup parse_method (non-"ocr"), vlm.llm_id, or top-level llm_id (in order)
-// - output_format
-// - image_context_size
-// - layout_recognize → ("@PaddleOCR:model_name" or empty)
type PictureParser struct {
- VLMModelID string // per-call IMAGE2TEXT model reference (parse_method / vlm.llm_id / llm_id)
- OutputFormat string
- ImageContextSize int // default 0
- LayoutRecognize string // layout_recognize (e.g. "@PaddleOCR")
- VideoPrompt string // for video-in-image detection (Python fallback)
+ OutputFormat string
}
// NewPictureParser constructs a PictureParser.
@@ -70,37 +61,14 @@ func NewPictureParser() *PictureParser {
}
// ConfigureFromSetup reads picture-specific configuration from the
-// parser setup map. Extracts the per-call IMAGE2TEXT model reference
-// (parse_method non-"ocr" > vlm.llm_id > llm_id), output_format,
-// image_context_size, layout_recognize, and video_prompt.
+// parser setup map.
func (p *PictureParser) ConfigureFromSetup(setup map[string]any) {
if p == nil || setup == nil {
return
}
- if pm, ok := setup["parse_method"].(string); ok && pm != "" && !strings.EqualFold(pm, "ocr") {
- p.VLMModelID = pm
- } else if vlm, ok := setup["vlm"].(map[string]any); ok {
- if llmID, ok := vlm["llm_id"].(string); ok && llmID != "" {
- p.VLMModelID = llmID
- }
- }
- if p.VLMModelID == "" {
- if llmID, ok := setup["llm_id"].(string); ok && llmID != "" {
- p.VLMModelID = llmID
- }
- }
if v, ok := setup["output_format"].(string); ok && v != "" {
p.OutputFormat = v
}
- if v, ok := setup["image_context_size"].(float64); ok {
- p.ImageContextSize = int(v)
- }
- if v, ok := setup["layout_recognize"].(string); ok && v != "" {
- p.LayoutRecognize = v
- }
- if v, ok := setup["video_prompt"].(string); ok && v != "" {
- p.VideoPrompt = v
- }
}
// ParseWithResult implements ParseResultProducer. It validates the
@@ -127,8 +95,6 @@ func (p *PictureParser) ParseWithResult(ctx context.Context, filename string, da
}
}
- // OutputFormat, VLMModelID, ImageContextSize, and LayoutRecognize
- // are consumed by maybeDispatchImage at the component layer.
// Parser output is normalized to JSON at the component boundary, so an
// absent backend format defaults to JSON here as well.
outFmt := p.OutputFormat
diff --git a/internal/parser/parser/picture_parser_test.go b/internal/parser/parser/picture_parser_test.go
index b3e3ccdc39..c53360984d 100644
--- a/internal/parser/parser/picture_parser_test.go
+++ b/internal/parser/parser/picture_parser_test.go
@@ -29,40 +29,16 @@ func TestNewPictureParser(t *testing.T) {
if p.OutputFormat != "" {
t.Errorf("OutputFormat = %q, want empty", p.OutputFormat)
}
- if p.VLMModelID != "" {
- t.Errorf("VLMModelID = %q, want empty", p.VLMModelID)
- }
- if p.ImageContextSize != 0 {
- t.Errorf("ImageContextSize = %d, want 0", p.ImageContextSize)
- }
}
func TestPictureParser_ConfigureFromSetup(t *testing.T) {
p := NewPictureParser()
p.ConfigureFromSetup(map[string]any{
- "vlm": map[string]any{
- "llm_id": "gpt-4-vision",
- },
- "output_format": "text",
- "image_context_size": float64(3),
- "layout_recognize": "@PaddleOCR",
- "video_prompt": "summarize",
+ "output_format": "text",
})
- if p.VLMModelID != "gpt-4-vision" {
- t.Errorf("VLMModelID = %q, want gpt-4-vision", p.VLMModelID)
- }
if p.OutputFormat != "text" {
t.Errorf("OutputFormat = %q, want text", p.OutputFormat)
}
- if p.ImageContextSize != 3 {
- t.Errorf("ImageContextSize = %d, want 3", p.ImageContextSize)
- }
- if p.LayoutRecognize != "@PaddleOCR" {
- t.Errorf("LayoutRecognize = %q, want @PaddleOCR", p.LayoutRecognize)
- }
- if p.VideoPrompt != "summarize" {
- t.Errorf("VideoPrompt = %q, want summarize", p.VideoPrompt)
- }
}
func TestPictureParser_ParseWithResult_NilSetup(t *testing.T) {