Skip to content

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

Fix Agent cited not show during chat (#19986)

Wang Qia46e64f53b27
diff --git a/internal/agent/component/agent.go b/internal/agent/component/agent.go
index aeb0e5677d..ac94a62fce 100644
--- a/internal/agent/component/agent.go
+++ b/internal/agent/component/agent.go
@@ -28,6 +28,7 @@ import (
"github.com/cloudwego/eino/schema"
"gorm.io/gorm"
+ "ragflow/internal/agent/chat"
"ragflow/internal/agent/component/prompts"
"ragflow/internal/agent/runtime"
agenttool "ragflow/internal/agent/tool"
@@ -211,9 +212,9 @@ func runEinoReActAgent(ctx context.Context, p AgentParam) (*schema.Message, erro
// graph releases its output stream, so agent.Stream does not return
// until the model finishes. GetMessageStreams blocks on the future's
// started signal (closed by the graph onStart callback), so starting
- // the collector first lets thinking deltas stream out in real time
+ // the collector first lets non-citation deltas stream out in real time
// while the checker runs, instead of buffering the entire round.
- emitDone := emitAgentModelStreams(ctx, future)
+ emitDone := emitAgentModelStreams(ctx, future, p.Cite)
stream, err := agent.Stream(ctx, input, opt)
if err != nil {
// Drain the collector so its goroutine exits before we return.
@@ -345,7 +346,7 @@ func buildAgentInputMessages(ctx context.Context, p AgentParam) []*schema.Messag
return input
}
-func emitAgentModelStreams(ctx context.Context, future react.MessageFuture) <-chan error {
+func emitAgentModelStreams(ctx context.Context, future react.MessageFuture, cite bool) <-chan error {
done := make(chan error, 1)
go func() {
var firstErr error
@@ -386,6 +387,11 @@ func emitAgentModelStreams(ctx context.Context, future react.MessageFuture) <-ch
if msg.Content == "" && msg.ReasoningContent == "" {
continue
}
+ // Retrieval tools can add chunks during this call. Hold the first
+ // answer until we know whether citation grounding will run.
+ if cite {
+ continue
+ }
if runtime.AgentMessageEventsEmitted(ctx) && !runtime.HasDeferredAgentMessageSink(ctx) {
continue
}
@@ -446,9 +452,9 @@ func addToolCallMemory(ctx context.Context, db *gorm.DB, p AgentParam, msg *sche
// the model to insert [ID:N] tags into the assistant's final
// content.
//
-// Returns the grounded content on success, the original content
-// unchanged when no chunks are available or the call fails. Mirrors
-// Python's `cite_letter` / `generate_with_citation` flow.
+// Streams grounded deltas when the invoker supports streaming. Returns the
+// original content when no chunks are available or the call fails before
+// emitting anything.
func applyCitationGrounding(ctx context.Context, db *gorm.DB, p AgentParam, content string, chunks []prompts.CitationSource) (string, error) {
if !p.Cite {
return content, nil
@@ -461,7 +467,7 @@ func applyCitationGrounding(ctx context.Context, db *gorm.DB, p AgentParam, cont
}
systemPrompt, _ := prompts.CitationPlusPrompt(chunks)
inv := getDefaultChatInvoker()
- resp, err := inv.Invoke(ctx, db, ChatInvokeRequest{
+ req := ChatInvokeRequest{
Driver: p.Driver,
ModelName: p.ModelID,
APIKey: p.APIKey,
@@ -471,18 +477,46 @@ func applyCitationGrounding(ctx context.Context, db *gorm.DB, p AgentParam, cont
{Role: schema.User, Content: content},
},
TopP: p.TopP,
- })
+ }
+ var resp *ChatInvokeResponse
+ var err error
+ if streamer, ok := inv.(chat.StreamingInvoker); ok {
+ var leadingWhitespace strings.Builder
+ emitting := false
+ resp, err = streamer.Stream(ctx, db, req, func(delta string, isThink bool) error {
+ if ctx.Err() != nil {
+ return ctx.Err()
+ }
+ if isThink || delta == "" {
+ return nil
+ }
+ if !emitting {
+ leadingWhitespace.WriteString(delta)
+ if strings.TrimSpace(leadingWhitespace.String()) == "" {
+ return nil
+ }
+ delta = leadingWhitespace.String()
+ emitting = true
+ }
+ runtime.EmitAgentMessage(ctx, delta, "")
+ return nil
+ })
+ } else {
+ resp, err = inv.Invoke(ctx, db, req)
+ }
if err != nil {
// Grounding is best-effort. Return the original content
// so the message still flows; the caller can decide
// whether to surface the error.
return content, err
}
- grounded := strings.TrimSpace(resp.Content)
- if grounded == "" {
+ if resp == nil {
+ return content, fmt.Errorf("citation grounding returned no response")
+ }
+ if strings.TrimSpace(resp.Content) == "" {
return content, nil
}
- return grounded, nil
+ return resp.Content, nil
}
// chunksFromState extracts the recorded retrieval chunks from
@@ -963,9 +997,8 @@ func (c *AgentComponent) invokeNow(ctx context.Context, db *gorm.DB, inputs map[
// the canvas state has recorded retrieval chunks (populated
// by the Retrieval tool during the ReAct loop), make a second
// LLM call to insert [ID:N] tags into the final content. The
- // grounding call is best-effort — on failure the original
- // content is kept and the error is surfaced under
- // outputs["grounding_error"].
+ // grounding call is best-effort until it emits a visible delta. An
+ // interrupted citation stream must not be followed by the first answer.
// Diagnostic sentinel (temporary — see plan): log the post-
// agentRunner state right before the `msg.Content` deref so a
// subsequent panic shows whether the agent returned (nil, nil).
@@ -995,6 +1028,9 @@ func (c *AgentComponent) invokeNow(ctx context.Context, db *gorm.DB, inputs map[
content = grounded
groundingStatus = "applied"
} else if gErr != nil {
+ if ctx.Err() != nil || runtime.AgentMessageEventsEmitted(ctx) || runtime.DeferredAgentMessageEventsEmitted(ctx) {
+ return nil, fmt.Errorf("component: Agent citation grounding stream: %w", gErr)
+ }
groundingStatus = "error: " + gErr.Error()
}
}
diff --git a/internal/agent/component/agent_test.go b/internal/agent/component/agent_test.go
index e092d32146..ab2fc4031b 100644
--- a/internal/agent/component/agent_test.go
+++ b/internal/agent/component/agent_test.go
@@ -306,7 +306,7 @@ func TestReactCheckerStreamsThinkingBeforeAgentReturns(t *testing.T) {
})
opt, future := react.WithMessageFuture()
- emitDone := emitAgentModelStreams(ctx, future)
+ emitDone := emitAgentModelStreams(ctx, future, false)
streamCh := make(chan *schema.StreamReader[*schema.Message], 1)
errCh := make(chan error, 1)
diff --git a/internal/agent/component/grounding_test.go b/internal/agent/component/grounding_test.go
index 69194cf667..d9073d9139 100644
--- a/internal/agent/component/grounding_test.go
+++ b/internal/agent/component/grounding_test.go
@@ -46,6 +46,26 @@ type groundingTestInvoker struct {
calls int
}
+type groundingStreamingInvoker struct {
+ groundingTestInvoker
+ streamCalls int
+}
+
+func (g *groundingStreamingInvoker) Stream(ctx context.Context, db *gorm.DB, req ChatInvokeRequest, onDelta func(string, bool) error) (*ChatInvokeResponse, error) {
+ g.streamCalls++
+ resp, err := g.Invoke(ctx, db, req)
+ if err != nil {
+ return nil, err
+ }
+ mid := len(resp.Content) / 2
+ for _, delta := range []string{resp.Content[:mid], resp.Content[mid:]} {
+ if err := onDelta(delta, false); err != nil {
+ return nil, err
+ }
+ }
+ return resp, nil
+}
+
func (g *groundingTestInvoker) Invoke(_ context.Context, _ *gorm.DB, req ChatInvokeRequest) (*ChatInvokeResponse, error) {
g.mu.Lock()
defer g.mu.Unlock()
@@ -64,7 +84,7 @@ func (g *groundingTestInvoker) Invoke(_ context.Context, _ *gorm.DB, req ChatInv
// TestGrounding_Applied: Cite=true + state has chunks → second
// LLM call is made and the grounded content replaces the original.
func TestGrounding_Applied(t *testing.T) {
- inv := &groundingTestInvoker{responses: []string{"grounded answer [ID:0]"}}
+ inv := &groundingStreamingInvoker{groundingTestInvoker: groundingTestInvoker{responses: []string{"grounded answer [ID:0]"}}}
prev := getDefaultChatInvoker()
SetDefaultChatInvoker(inv)
defer SetDefaultChatInvoker(prev)
@@ -78,6 +98,12 @@ func TestGrounding_Applied(t *testing.T) {
{"id": "0", "content": "the source content"},
})
ctx := runtime.WithState(t.Context(), state)
+ var streamed string
+ var deltas int
+ ctx = runtime.WithAgentMessageEmitter(ctx, func(content, _ string) {
+ streamed += content
+ deltas++
+ })
c := NewAgentComponent(AgentParam{ModelID: "stub", Cite: true})
out, err := c.Invoke(ctx, nil, map[string]any{"user_prompt": "q"})
@@ -93,6 +119,9 @@ func TestGrounding_Applied(t *testing.T) {
if inv.calls != 1 {
t.Errorf("expected 1 chat call, got %d", inv.calls)
}
+ if inv.streamCalls != 1 || deltas != 2 || streamed != "grounded answer [ID:0]" {
+ t.Errorf("stream calls=%d, deltas=%d, streamed=%q, want 1 / 2 / grounded answer [ID:0]", inv.streamCalls, deltas, streamed)
+ }
// System message should contain the citation prompt + sources block.
if got := inv.lastReq.Messages[0].Role; got != schema.System {
t.Errorf("first message role=%v, want System", got)