Skip to content

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

fix: honor Go chat citation settings and preserve final streamed citations (#19797)

adws52cf6fe586ea
diff --git a/internal/service/chat_pipeline.go b/internal/service/chat_pipeline.go
index e161f06b02..f4a8c6f686 100644
--- a/internal/service/chat_pipeline.go
+++ b/internal/service/chat_pipeline.go
@@ -409,6 +409,15 @@ func (s *ChatPipelineService) AsyncChat(
// === Phase 6: SQL Retrieval ===
// Retrieve field_map for SQL retrieval (preferred over vector search)
promptConfig := chat.PromptConfig
+ // Either the chat setting or the request can disable citations. Resolve
+ // this once before any retrieval path can return early.
+ quote := true
+ if v, ok := kwargs["quote"].(bool); ok {
+ quote = v
+ }
+ if promptConfigQuote, ok := promptConfig["quote"].(bool); ok {
+ quote = quote && promptConfigQuote
+ }
fieldMap, fmErr := s.kbDAO.GetFieldMap(ctx, dao.DB, kbIDStrings(kbs))
if fmErr != nil {
common.Warn("get_field_map failed; proceeding without field_map", zap.Error(fmErr))
@@ -421,11 +430,6 @@ func (s *ChatPipelineService) AsyncChat(
if len(fieldMap) > 0 && chatModel != nil && len(kbs) > 0 {
common.Info("Phase 6: Use SQL to retrieval")
common.Debug("field_map retrieved", zap.Any("field_map", fieldMap))
- quote := true
- if v, ok := promptConfig["quote"].(bool); ok {
- quote = v
- }
-
ans, sqlErr := s.useSQL(
ctx, chat, kbs, questions[len(questions)-1], chatModel, fieldMap, quote,
)
@@ -693,16 +697,10 @@ func (s *ChatPipelineService) AsyncChat(
// When false, the entire block is skipped.
if hasKnowledgeParam {
if useReasoning && chatModel != nil && len(kbs) > 0 {
- // Reasoning chat (level 1..4): drive the agentic-RAG harness at
- // the corresponding mode, mirroring Python's dialog_service →
- // RAGTools → harness/*.
- //
- // The harness 'rag' tool composes the final cited answer itself
- // (terminal tool, mirroring Python). When it returns a non-empty
- // Answer we emit that answer directly (decorated per Python's
- // decorate_answer: references resolved, prompt empty) and stop —
- // there is no second-generation pass in Phase 10/11. Otherwise
- // we keep its evidence as kbinfos and fall through.
+ // The harness collects evidence and composes the answer for
+ // reasoning levels 1..4. Apply citation visibility to its stream
+ // and final answer. If no answer is returned, keep the evidence
+ // and continue through the regular generation path.
thinkingMode := harnessModeForLevel(reasoningLevel)
question := strings.Join(questions, " ")
// Stream the answer as the harness composes it (Python
@@ -735,16 +733,17 @@ func (s *ChatPipelineService) AsyncChat(
// Engine progress (B) is pushed from concurrent research-slot
// goroutines while the final compose runs on the main goroutine, so
// serialize the think-framing state machine and the out<-send.
+ var answerFilter, reasoningFilter citationStreamFilter
var sinkMu sync.Mutex
+ send := func(ev AsyncChatResult) {
+ select {
+ case out <- ev:
+ case <-ctx.Done():
+ }
+ }
sink := func(delta string, isThink bool) {
sinkMu.Lock()
defer sinkMu.Unlock()
- send := func(ev AsyncChatResult) {
- select {
- case out <- ev:
- case <-ctx.Done():
- }
- }
if isThink {
if answerStreamed {
// Never reopen a thought block after the answer: drop the
@@ -760,6 +759,9 @@ func (s *ChatPipelineService) AsyncChat(
StartToThink: true,
})
}
+ if !quote {
+ delta = reasoningFilter.write(delta)
+ }
if delta != "" {
send(AsyncChatResult{
Reasoning: delta,
@@ -771,6 +773,9 @@ func (s *ChatPipelineService) AsyncChat(
return
}
if harnessThinking {
+ if text := reasoningFilter.flush(); text != "" {
+ send(AsyncChatResult{Reasoning: text, Reference: map[string]interface{}{}, CreatedAt: float64(time.Now().Unix())})
+ }
harnessThinking = false
send(AsyncChatResult{
Reference: map[string]interface{}{},
@@ -781,6 +786,11 @@ func (s *ChatPipelineService) AsyncChat(
}
if delta != "" {
answerStreamed = true
+ }
+ if !quote {
+ delta = answerFilter.write(delta)
+ }
+ if delta != "" {
send(AsyncChatResult{
Answer: delta,
Reference: map[string]interface{}{},
@@ -836,6 +846,11 @@ func (s *ChatPipelineService) AsyncChat(
// EndToThink after the Phase 10/11 answer or the Final below.
// No-op when no block is open.
sink("", false)
+ sinkMu.Lock()
+ if text := answerFilter.flush(); text != "" {
+ send(AsyncChatResult{Answer: text, Reference: map[string]interface{}{}, CreatedAt: float64(time.Now().Unix())})
+ }
+ sinkMu.Unlock()
if hErr != nil {
common.Warn("harness retrieval failed", zap.Error(hErr))
} else {
@@ -843,7 +858,7 @@ func (s *ChatPipelineService) AsyncChat(
if harnessAnswer != "" {
common.Info("harness produced final cited answer; short-circuiting",
zap.Int("answer_chars", len(harnessAnswer)))
- final := s.decorateHarnessAnswer(harnessAnswer, kbinfos, slotCites, citeChunkIDs)
+ final := s.decorateHarnessAnswer(harnessAnswer, kbinfos, slotCites, citeChunkIDs, quote)
final.Final = true
out <- final
return
@@ -1031,13 +1046,17 @@ func (s *ChatPipelineService) AsyncChat(
// otherwise see an empty reply.
if emptyResponseApplies(len(knowledges), attachments, hasImageAttachments) {
if emptyResp, ok := promptConfig["empty_response"].(string); ok && emptyResp != "" {
+ finalReference := kbinfos
+ if !quote {
+ finalReference = map[string]interface{}{}
+ }
out <- AsyncChatResult{
Answer: emptyResp,
Reference: map[string]interface{}{},
}
out <- AsyncChatResult{
Answer: emptyResp,
- Reference: kbinfos,
+ Reference: finalReference,
AudioBinary: s.synthesizeTTS(ctx, ttsModel, emptyResp),
Prompt: fmt.Sprintf("\n\n### Query:\n%s", strings.Join(questions, " ")),
Final: true,
@@ -1075,13 +1094,6 @@ func (s *ChatPipelineService) AsyncChat(
// Build citation prompt if quoting is enabled.
prompt4citation := ""
- quote := true
- if v, ok := kwargs["quote"].(bool); ok {
- quote = v
- }
- if promptConfigQuote, ok := promptConfig["quote"].(bool); ok {
- quote = quote && promptConfigQuote
- }
if len(knowledges) > 0 && quote {
// Python's citation_prompt() (generator.py:226) renders
// citation_prompt.md — the full rules with examples, which the agent
@@ -1253,6 +1265,34 @@ func (s *ChatPipelineService) AsyncChat(
// Streaming path: accumulate answer, emit deltas.
var fullAnswer string
thinkState := &ThinkStreamState{}
+ var answerFilter, reasoningFilter citationStreamFilter
+ emit := func(result AsyncChatResult) {
+ if !result.Final && result.Answer != "" {
+ result.AudioBinary = s.synthesizeTTS(ctx, ttsModel, result.Answer)
+ }
+ out <- result
+ }
+ send := func(result AsyncChatResult) {
+ if !quote {
+ // Flush before a boundary so buffered text stays in its section.
+ if result.StartToThink || result.EndToThink || result.Final {
+ if text := reasoningFilter.flush(); text != "" {
+ emit(AsyncChatResult{Reasoning: text, Reference: map[string]interface{}{}, CreatedAt: float64(time.Now().Unix())})
+ }
+ if text := answerFilter.flush(); text != "" {
+ emit(AsyncChatResult{Answer: text, Reference: map[string]interface{}{}, CreatedAt: float64(time.Now().Unix())})
+ }
+ }
+ if !result.Final && (result.Answer != "" || result.Reasoning != "") {
+ result.Answer = answerFilter.write(result.Answer)
+ result.Reasoning = reasoningFilter.write(result.Reasoning)
+ if result.Answer == "" && result.Reasoning == "" {
+ return
+ }
+ }
+ }
+ emit(result)
+ }
// Tool routing: use tool-loop method when tools are bound.
var driverErr error
@@ -1271,48 +1311,47 @@ func (s *ChatPipelineService) AsyncChat(
if text == "<think>" {
inThink = true
- out <- AsyncChatResult{
+ send(AsyncChatResult{
Answer: "",
Reference: map[string]interface{}{},
AudioBinary: nil,
CreatedAt: float64(time.Now().Unix()),
Final: false,
StartToThink: true,
- }
+ })
return nil
}
if text == "</think>" {
inThink = false
- out <- AsyncChatResult{
+ send(AsyncChatResult{
Answer: "",
Reference: map[string]interface{}{},
AudioBinary: nil,
CreatedAt: float64(time.Now().Unix()),
Final: false,
EndToThink: true,
- }
+ })
return nil
}
if inThink {
// Reasoning text — route to Reasoning field so
// the SSE handler maps it to
// `delta.reasoning_content`.
- out <- AsyncChatResult{
+ send(AsyncChatResult{
Reasoning: text,
Reference: map[string]interface{}{},
AudioBinary: nil,
CreatedAt: float64(time.Now().Unix()),
Final: false,
- }
+ })
} else {
// Regular answer content
- out <- AsyncChatResult{
- Answer: text,
- Reference: map[string]interface{}{},
- AudioBinary: s.synthesizeTTS(ctx, ttsModel, text),
- CreatedAt: float64(time.Now().Unix()),
- Final: false,
- }
+ send(AsyncChatResult{
+ Answer: text,
+ Reference: map[string]interface{}{},
+ CreatedAt: float64(time.Now().Unix()),
+ Final: false,
+ })
}
return nil
})
@@ -1322,26 +1361,25 @@ func (s *ChatPipelineService) AsyncChat(
func(answer *string, reason *string) error {
if reason != nil && *reason != "" {
if thinkState.EnterReasoning() {
- out <- AsyncChatResult{
+ send(AsyncChatResult{
Answer: "",
Reference: map[string]interface{}{},
AudioBinary: nil,
CreatedAt: float64(time.Now().Unix()),
Final: false,
StartToThink: true,
- }
+ })
}
deltas := NextThinkDelta(thinkState, *reason, 16)
for _, d := range deltas {
if d.Kind == ThinkDeltaText && d.Value != "" {
fullAnswer += d.Value
- out <- AsyncChatResult{
- Answer: d.Value,
- Reference: map[string]interface{}{},
- AudioBinary: s.synthesizeTTS(ctx, ttsModel, d.Value),
- CreatedAt: float64(time.Now().Unix()),
- Final: false,
- }
+ send(AsyncChatResult{
+ Answer: d.Value,
+ Reference: map[string]interface{}{},
+ CreatedAt: float64(time.Now().Unix()),
+ Final: false,
+ })
}
}
}
@@ -1350,35 +1388,33 @@ func (s *ChatPipelineService) AsyncChat(
for _, d := range FlushRemaining(thinkState) {
if d.Kind == ThinkDeltaText && d.Value != "" {
fullAnswer += d.Value
- out <- AsyncChatResult{
- Answer: d.Value,
- Reference: map[string]interface{}{},
- AudioBinary: s.synthesizeTTS(ctx, ttsModel, d.Value),
- CreatedAt: float64(time.Now().Unix()),
- Final: false,
- }
+ send(AsyncChatResult{
+ Answer: d.Value,
+ Reference: map[string]interface{}{},
+ CreatedAt: float64(time.Now().Unix()),
+ Final: false,
+ })
}
}
- out <- AsyncChatResult{
+ send(AsyncChatResult{
Answer: "",
Reference: map[string]interface{}{},
AudioBinary: nil,
CreatedAt: float64(time.Now().Unix()),
Final: false,
EndToThink: true,
- }
+ })
}
fullAnswer += *answer
deltas := BufferAnswerDelta(thinkState, *answer, 16)
for _, d := range deltas {
if d.Kind == ThinkDeltaText && d.Value != "" {
- out <- AsyncChatResult{
- Answer: d.Value,
- Reference: map[string]interface{}{},
- AudioBinary: s.synthesizeTTS(ctx, ttsModel, d.Value),
- CreatedAt: float64(time.Now().Unix()),
- Final: false,
- }
+ send(AsyncChatResult{
+ Answer: d.Value,
+ Reference: map[string]interface{}{},
+ CreatedAt: float64(time.Now().Unix()),
+ Final: false,
+ })
}
}
}
@@ -1387,10 +1423,10 @@ func (s *ChatPipelineService) AsyncChat(
)
}
if driverErr != nil {
- out <- AsyncChatResult{
+ send(AsyncChatResult{
Answer: fmt.Sprintf("**ERROR**: %s", driverErr.Error()),
Final: true,
- }
+ })
return
}
@@ -1401,36 +1437,35 @@ func (s *ChatPipelineService) AsyncChat(
for _, d := range FlushRemaining(thinkState) {
if d.Kind == ThinkDeltaMarker && d.Value == "</think>" {
hadThinkClose = true
- out <- AsyncChatResult{
+ send(AsyncChatResult{
Answer: "",
Reference: map[string]interface{}{},
AudioBinary: nil,
CreatedAt: float64(time.Now().Unix()),
Final: false,
EndToThink: true,
- }
+ })
} else if d.Kind == ThinkDeltaText && d.Value != "" {
- out <- AsyncChatResult{
- Answer: d.Value,
- Reference: map[string]interface{}{},
- AudioBinary: s.synthesizeTTS(ctx, ttsModel, d.Value),
- CreatedAt: float64(time.Now().Unix()),
- Final: false,
- }
+ send(AsyncChatResult{
+ Answer: d.Value,
+ Reference: map[string]interface{}{},
+ CreatedAt: float64(time.Now().Unix()),
+ Final: false,
+ })
}
}
// Close reasoning if the stream ended while still in reasoning mode
// (e.g. model returned only reasoning chunks with no content delta).
// Skip when FlushRemaining already emitted a </think> marker.
if !hadThinkClose && thinkState.ExitReasoning() {
- out <- AsyncChatResult{
+ send(AsyncChatResult{
Answer: "",
Reference: map[string]interface{}{},
AudioBinary: nil,
CreatedAt: float64(time.Now().Unix()),
Final: false,
EndToThink: true,
- }
+ })
}
// Decorate and yield the final answer.
@@ -1443,7 +1478,7 @@ func (s *ChatPipelineService) AsyncChat(
final.Final = true
final.AudioBinary = nil
timer.Exit(common.PhaseGenerateAnswer)
- out <- final
+ send(final)
} else {
// Non-streaming: get the answer synchronously.
var answer string
@@ -3072,6 +3107,10 @@ func (s *ChatPipelineService) decorateAnswer(
ans = strings.TrimSpace(parts[1])
}
}
+ if !quote {
+ think = stripCitations(think)
+ ans = stripCitations(ans)
+ }
var citationIdx map[int]struct{}
var refs map[string]interface{}
@@ -3178,10 +3217,10 @@ func (s *ChatPipelineService) decorateAnswer(
}
}
- // Build refs only when the answer contains a citation that resolves to a
- // chunk. Retrieved evidence without a citation must not be exposed as a
- // document reference.
- if hasKnowledges && len(citationIdx) > 0 {
+ // Include sources only when citations are enabled and at least one citation
+ // actually resolves to a chunk, stripping chunk vectors. Retrieved evidence
+ // without a citation must not be exposed as a document reference.
+ if hasKnowledges && quote && len(citationIdx) > 0 {
refs = make(map[string]interface{})
for k, v := range kbinfos {
refs[k] = v
@@ -3200,6 +3239,8 @@ func (s *ChatPipelineService) decorateAnswer(
}
refs["chunks"] = chunksFormat(newChunks)
}
+ } else if !quote {
+ refs = map[string]interface{}{}
}
// Check for invalid API key errors (outside knowledges guard).
@@ -3276,13 +3317,13 @@ func (s *ChatPipelineService) decorateAnswer(
}
}
-// decorateHarnessAnswer mirrors Python's rag_agent decorate_answer used on the
-// reasoning>=1 path (dialog_service.py:2092-2137). The harness 'rag' tool
-// already composed the final answer WITH its own [ID:N] citations, so — unlike
-// decorateAnswer for the native async_chat path — we never run insert_citations
-// here. We only resolve the existing markers, repair bad formats, filter
-// doc_aggs to the cited docs, and build the reference from the harness citation
-// pool (chunks stripped of their vectors). Prompt stays empty, matching Python.
+// decorateHarnessAnswer formats the reasoning chat's final answer. The harness
+// 'rag' tool already composed the answer WITH its own [ID:N] citations, so —
+// unlike decorateAnswer for the native async_chat path — we never run
+// insert_citations here. We only resolve the existing markers, repair bad
+// formats, filter doc_aggs to the cited docs, and build the reference from the
+// harness citation pool (chunks stripped of their vectors). Prompt stays empty,
+// matching Python.
//
// The answer's [ID:n] markers index the evidence list the compose rendered for
// the model — harness.Kbinfos.CiteChunkIDs, passed in as citeChunkIDs. That list
@@ -3297,11 +3338,9 @@ func (s *ChatPipelineService) decorateAnswer(
// returned whenever the pool is non-empty, so an answer the model composed without
// markers still carries its sources.
//
-// Precondition for marker n to open passage n: the rendered blocks resolve to pool
-// chunks in order. A block whose id the pool does not know keeps its slot — an empty
-// entry the reference carries — so the later blocks keep their numbers; its own
-// citations are dropped, since the user could not open them.
-func (s *ChatPipelineService) decorateHarnessAnswer(answer string, kbinfos map[string]interface{}, slotCitations map[string][]string, citeChunkIDs []string) AsyncChatResult {
+// When quote is false the citations are removed entirely and the reference is
+// cleared (see stripCitations).
+func (s *ChatPipelineService) decorateHarnessAnswer(answer string, kbinfos map[string]interface{}, slotCitations map[string][]string, citeChunkIDs []string, quote bool) AsyncChatResult {
think := ""
ans := answer
if strings.Contains(answer, "</think>") {
@@ -3311,6 +3350,20 @@ func (s *ChatPipelineService) decorateHarnessAnswer(answer string, kbinfos map[s
}
}
+ // Invalid-key hint (dialog_service.py:2134-2135).
+ if strings.Contains(strings.ToLower(ans), "invalid key") ||
+ strings.Contains(strings.ToLower(ans), "invalid api") {
+ ans += " Please set LLM API-Key in 'User Setting -> Model providers -> API-Key'"
+ }
+
+ if !quote {
+ return AsyncChatResult{
+ Answer: stripCitations(think + ans),
+ Reference: map[string]interface{}{},
+ CreatedAt: float64(time.Now().Unix()),
+ }
+ }
+
chunksRaw, _ := kbinfos["chunks"].([]map[string]interface{})
// The markers exactly as the model wrote them (before any repair), for the
// citation observability log below.
@@ -3411,12 +3464,6 @@ func (s *ChatPipelineService) decorateHarnessAnswer(answer string, kbinfos map[s
refs = ref
}
- // Invalid-key hint (dialog_service.py:2134-2135).
- if strings.Contains(strings.ToLower(ans), "invalid key") ||
- strings.Contains(strings.ToLower(ans), "invalid api") {
- ans += " Please set LLM API-Key in 'User Setting -> Model providers -> API-Key'"
- }
-
return AsyncChatResult{
Answer: think + ans,
Reference: refs,
@@ -3770,23 +3817,6 @@ Please correct the error and write SQL again using the exact field names above,
//
// - err: non-nil when something went wrong; caller should log and fall
// through.
-//
-// StructuredQuery is the exported narrow entrypoint for the agentic-search
-// structured_query tool: translate a natural-language question to SQL over the
-// given tabular KBs and return the answer + reference chunks. It forwards to the
-// internal useSQL with quote=false (the agent tool returns the answer directly,
-// not a cited natural-language response).
-func (s *ChatPipelineService) StructuredQuery(
- ctx context.Context,
- chat *entity.Chat,
- kbs []*entity.Knowledgebase,
- question string,
- chatModel *modelModule.ChatModel,
- fieldMap map[string]interface{},
-) (ans map[string]interface{}, err error) {
- return s.useSQL(ctx, chat, kbs, question, chatModel, fieldMap, false)
-}
-
func (s *ChatPipelineService) useSQL(
ctx context.Context,
chat *entity.Chat,
@@ -3888,7 +3918,7 @@ func (s *ChatPipelineService) useSQL(
// repair doesn't yield source columns, fall through to the
// best-effort answer (matches Python's `returning best-effort
// answer` log at line 1221).
- if !isAggregateSQL(sqlText) && !hasSourceColumns(rows) {
+ if quote && !isAggregateSQL(sqlText) && !hasSourceColumns(rows) {
common.Debug("SQL retrieval: result missing source columns; attempting repair",
zap.String("sql", sqlText))
expectedCol := expectedDocNameColumn(engineName)
@@ -3924,7 +3954,7 @@ func (s *ChatPipelineService) useSQL(
// and best-effort empty refs.
answerStr, ref := s.buildSQLReference(
ctx, docEngine, tableName, sqlText, rows,
- sysPrompt, engineName, kbs, fieldMap,
+ sysPrompt, engineName, kbs, fieldMap, quote,
)
return map[string]interface{}{
"answer": answerStr,
@@ -4638,8 +4668,12 @@ func (s *ChatPipelineService) buildSQLReference(
sysPrompt, engineName string,
kbs []*entity.Knowledgebase,
fieldMap map[string]interface{},
+ quote bool,
) (string, map[string]interface{}) {
if len(rows) == 0 {
+ if !quote {
+ return "No results.", map[string]interface{}{}
+ }
return "No results.", map[string]interface{}{
"chunks": []map[string]interface{}{},
"doc_aggs": []interface{}{},
@@ -4650,6 +4684,9 @@ func (s *ChatPipelineService) buildSQLReference(
// Scalar shortcut — matches the previous renderSQLAnswer behavior.
if len(rows) == 1 && len(rows[0]) == 1 {
for _, v := range rows[0] {
+ if !quote {
+ return cleanCellValue(v), map[string]interface{}{}
+ }
return cleanCellValue(v), map[string]interface{}{
"chunks": []map[string]interface{}{},
"doc_aggs": []interface{}{},
@@ -4662,6 +4699,7 @@ func (s *ChatPipelineService) buildSQLReference(
docIDIdx, docNameIdx, kbIDIdx, columns := extractSourceColumnIndexes(rows)
expectedCol := expectedDocNameColumn(engineName)
hasSrc := len(docIDIdx) > 0 && len(docNameIdx) > 0
+ showSource := quote && hasSrc
// Build the set of "display column" indices (everything except
// doc_id, docnm*, kb_id*). Python uses set subtraction at
@@ -4690,13 +4728,13 @@ func (s *ChatPipelineService) buildSQLReference(
header.WriteString(mapColumnName(columns[i], fieldMap))
header.WriteString("|")
}
- if hasSrc {
+ if showSource {
header.WriteString("Source|")
}
// --- Separator (Python L1285) ---
sep := strings.Repeat("|------", len(displayCols)) + "|"
- if hasSrc {
+ if showSource {
sep += "------|"
}
@@ -4709,7 +4747,7 @@ func (s *ChatPipelineService) buildSQLReference(
cells.WriteString(cleanCellValue(r[columns[i]]))
cells.WriteString("|")
}
- if hasSrc {
+ if showSource {
cells.WriteString(fmt.Sprintf(" ##%d$$|", rowIdx))
}
// Skip rows that are entirely empty/whitespace (Python's
@@ -4722,6 +4760,9 @@ func (s *ChatPipelineService) buildSQLReference(
rowsJoined := stripISOTimestamps(strings.Join(bodyRows, "\n"))
answer := strings.Join([]string{header.String(), sep, rowsJoined}, "\n")
+ if !quote {
+ return answer, map[string]interface{}{}
+ }
// --- Reference: chunks + doc_aggs ---
ref := map[string]interface{}{
diff --git a/internal/service/chat_pipeline_test.go b/internal/service/chat_pipeline_test.go
index b009e30e7b..c64e5ae6b4 100644
--- a/internal/service/chat_pipeline_test.go
+++ b/internal/service/chat_pipeline_test.go
@@ -235,8 +235,8 @@ func TestDecorateAnswer_InvalidKeySuffix(t *testing.T) {
}
// TestDecorateAnswer_LeavesCanonicalMarkers covers P0: the decorator
-// passes canonical [ID:N] markers through unchanged when there are
-// no chunks to cite (so insertCitations is skipped).
+// passes canonical [ID:N] markers through unchanged when quoting is enabled
+// and there are no chunks to cite (so insertCitations is skipped).
func TestDecorateAnswer_LeavesCanonicalMarkers(t *testing.T) {
s := &ChatPipelineService{}
timer, _ := newTimerAndPrompt()
@@ -248,7 +248,7 @@ func TestDecorateAnswer_LeavesCanonicalMarkers(t *testing.T) {
[]string{"q"},
0,
timer,
- nil, 0.0, false,
+ nil, 0.0, true,
nil,
"",
nil,
@@ -262,14 +262,14 @@ func TestDecorateAnswer_LeavesCanonicalMarkers(t *testing.T) {
}
// TestDecorateAnswer_RepairNotRunWhenNoQuote covers P0.10: when
-// quote=false, the citation-repair branch is gated off and the
-// answer is preserved verbatim.
+// quote=false, citation markers are removed without rewriting ordinary
+// bracketed numbers.
func TestDecorateAnswer_RepairNotRunWhenNoQuote(t *testing.T) {
s := &ChatPipelineService{}
timer, _ := newTimerAndPrompt()
result := s.decorateAnswer(
t.Context(),
- "see (ID: 12) for details",
+ "<think>Checking[ID:0]##0$$</think>see (ID: 12)[ID:0]##1$$ in [2024] for details",
map[string]interface{}{"chunks": []interface{}{}, "doc_aggs": []interface{}{}},
"system prompt",
[]string{"q"},
@@ -283,8 +283,8 @@ func TestDecorateAnswer_RepairNotRunWhenNoQuote(t *testing.T) {
nil,
false,
)
- if result.Answer != "see (ID: 12) for details" {
- t.Errorf("quote=false must not repair, got %q", result.Answer)
+ if result.Answer != "<think>Checking</think>see in [2024] for details" {
+ t.Errorf("quote=false citation cleanup = %q", result.Answer)
}
}
@@ -649,7 +649,7 @@ func TestBuildSQLReference_Scalar(t *testing.T) {
ans, ref := s.buildSQLReference(
t.Context(), nil, "", "",
[]map[string]interface{}{{"count": 42.0}},
- "", "", nil, nil,
+ "", "", nil, nil, true,
)
if ans != "42" {
t.Errorf("buildSQLReference scalar answer = %q, want %q", ans, "42")
@@ -676,7 +676,7 @@ func TestBuildSQLReference_MultiRowTable(t *testing.T) {
ans, ref := s.buildSQLReference(
t.Context(), nil, "", "select id, name from t",
rows,
- "sys", "elasticsearch", nil, nil,
+ "sys", "elasticsearch", nil, nil, true,
)
// No source columns → empty chunks/doc_aggs.
if chunks, _ := ref["chunks"].([]map[string]interface{}); len(chunks) != 0 {
@@ -792,6 +792,23 @@ func TestDecorateAnswer_VectorStrippedFromReference(t *testing.T) {
}
}
+func TestDecorateAnswer_NoReferencesWhenQuoteDisabled(t *testing.T) {
+ s := &ChatPipelineService{}
+ timer, _ := newTimerAndPrompt()
+ kb := map[string]interface{}{
+ "chunks": []map[string]interface{}{{"chunk_id": "c1", "content_with_weight": "Evidence", "doc_id": "d1"}},
+ "doc_aggs": []interface{}{map[string]interface{}{"doc_id": "d1", "doc_name": "Source"}},
+ }
+ result := s.decorateAnswer(t.Context(), "Answer", kb, "", nil, 0, timer,
+ nil, 0, false, nil, "", nil, "", nil, true)
+ if result.Reference == nil || len(result.Reference) != 0 {
+ t.Fatalf("disabled citations must explicitly clear references: %#v", result.Reference)
+ }
+ if result.Answer != "Answer" {
+ t.Fatalf("answer changed: %q", result.Answer)
+ }
+}
+
// --- normalizeInternetFlag / shouldUseWebSearch parity with Python ---
// TestNormalizeInternetFlag_PythonParity pins the three-state return of
@@ -1405,7 +1422,7 @@ func TestBuildSQLReference_EmptyRows(t *testing.T) {
s := &ChatPipelineService{}
ans, ref := s.buildSQLReference(
t.Context(), nil, "", "", nil,
- "", "", nil, nil,
+ "", "", nil, nil, true,
)
if ans != "No results." {
t.Errorf("ans = %q, want %q", ans, "No results.")
@@ -1427,7 +1444,7 @@ func TestBuildSQLReference_NonAggregateWithSourceColumns(t *testing.T) {
s := &ChatPipelineService{}
ans, ref := s.buildSQLReference(
t.Context(), nil, "t", "select doc_id, docnm_kwd, title from t",
- rows, "", "elasticsearch", kbs, nil,
+ rows, "", "elasticsearch", kbs, nil, true,
)
if !strings.Contains(ans, "Source|") {
t.Errorf("expected Source column in answer, got:\n%s", ans)
@@ -1449,6 +1466,17 @@ func TestBuildSQLReference_NonAggregateWithSourceColumns(t *testing.T) {
t.Errorf("chunks[%d].dataset_id = %v, want kb_a", i, cm["dataset_id"])
}
}
+
+ ans, ref = s.buildSQLReference(
+ t.Context(), nil, "t", "select doc_id, docnm_kwd, title from t",
+ rows, "", "elasticsearch", kbs, nil, false,
+ )
+ if strings.Contains(ans, "Source|") || strings.Contains(ans, "##0$$") {
+ t.Errorf("quote=false leaked SQL citations in answer:\n%s", ans)
+ }
+ if len(ref) != 0 {
+ t.Errorf("quote=false reference = %#v, want empty", ref)
+ }
}
// TestBuildSQLReference_AggregateMissingSourceColumnsSecondaryFetch
@@ -1475,7 +1503,7 @@ func TestBuildSQLReference_AggregateMissingSourceColumnsSecondaryFetch(t *testin
ans, ref := s.buildSQLReference(
t.Context(), sqlEngine, "t",
"select count(*) from t where x = 1",
- rows, "", "elasticsearch", kbs, nil,
+ rows, "", "elasticsearch", kbs, nil, true,
)
// Multi-cell aggregate → renders as a table, not a scalar.
if !strings.Contains(ans, "|42|") {
@@ -1498,7 +1526,7 @@ func TestBuildSQLReference_NonAggregateMissingSourceEmptyRefs(t *testing.T) {
s := &ChatPipelineService{}
ans, ref := s.buildSQLReference(
t.Context(), nil, "t", "select title from t",
- rows, "", "elasticsearch", nil, nil,
+ rows, "", "elasticsearch", nil, nil, true,
)
if !strings.Contains(ans, "T1") || !strings.Contains(ans, "T2") {
t.Errorf("expected table data in answer, got:\n%s", ans)
@@ -1526,7 +1554,7 @@ func TestBuildSQLReference_DisplayNameTranslation(t *testing.T) {
s := &ChatPipelineService{}
ans, _ := s.buildSQLReference(
t.Context(), nil, "t", "select doc_id, docnm_kwd, title from t",
- rows, "", "elasticsearch", nil, fieldMap,
+ rows, "", "elasticsearch", nil, fieldMap, true,
)
if !strings.Contains(ans, "|My Title|") {
t.Errorf("expected translated column name, got:\n%s", ans)
@@ -1545,7 +1573,7 @@ func TestBuildSQLReference_ISOTimestampStripped(t *testing.T) {
s := &ChatPipelineService{}
ans, _ := s.buildSQLReference(
t.Context(), nil, "t", "select doc_id, docnm_kwd, created_at from t",
- rows, "", "elasticsearch", nil, nil,
+ rows, "", "elasticsearch", nil, nil, true,
)
if strings.Contains(ans, "T13:24:55") {
t.Errorf("expected ISO timestamp stripped, got:\n%s", ans)
@@ -1864,7 +1892,7 @@ func TestDecorateHarnessAnswerReferenceUsesClientChunkShape(t *testing.T) {
}
s := &ChatPipelineService{}
- res := s.decorateHarnessAnswer("The answer [ID:0]", kbinfos, nil, nil)
+ res := s.decorateHarnessAnswer("The answer [ID:0]", kbinfos, nil, nil, true)
if res.Reference == nil {
t.Fatal("a cited answer must carry a reference")
}
@@ -1960,6 +1988,7 @@ func TestDecorateHarnessAnswerRewritesSlotCitations(t *testing.T) {
kbinfos,
map[string][]string{"0": {"c9"}},
nil,
+ true,
)
if strings.Contains(res.Answer, "[ID:Slot") {
t.Fatalf("final answer still carries the internal slot citation: %q", res.Answer)
@@ -1981,6 +2010,114 @@ func TestDecorateHarnessAnswerRewritesSlotCitations(t *testing.T) {
}
}
+func TestAsyncChatHarnessQuote(t *testing.T) {
+ db := setupChatPipelineVisionTestDB(t)
+ seedChatPipelineVisionTenant(t, db, "glm-4-flash@ZHIPU-AI")
+ if err := db.AutoMigrate(&entity.Knowledgebase{}); err != nil {
+ t.Fatal(err)
+ }
+ if err := db.Create(&entity.Knowledgebase{ID: "kb-1", TenantID: "tenant-1", Name: "Knowledge"}).Error; err != nil {
+ t.Fatal(err)
+ }
+ previousHarness := harnessRetriever
+ t.Cleanup(func() { harnessRetriever = previousHarness })
+ harnessRetriever = func(_ context.Context, req HarnessRequest) (HarnessResult, error) {
+ req.AnswerSink("Research[ID:", true)
+ req.AnswerSink("0] done", true)
+ for _, delta := range []string{"Answer[", "ID:0", "] end"} {
+ req.AnswerSink(delta, false)
+ }
+ return HarnessResult{
+ Answer: "Answer[ID:0] end",
+ Chunks: []map[string]interface{}{{"chunk_id": "c1", "content_with_weight": "Evidence", "doc_id": "d1"}},
+ DocAggs: []map[string]interface{}{{"doc_id": "d1", "doc_name": "Source"}},
+ }, nil
+ }
+ for _, tt := range []struct {
+ name string
+ config, request any
+ wantQuote bool
+ }{
+ {"default enabled", nil, nil, true},
+ {"chat disabled", false, true, false},
+ {"request disabled", true, false, false},
+ } {
+ t.Run(tt.name, func(t *testing.T) {
+ chat := dialForTest("glm-4-flash@ZHIPU-AI")
+ chat.KBIDs = entity.JSONSlice{"kb-1"}
+ chat.PromptConfig["system"] = "Answer using {knowledge}"
+ chat.PromptConfig["quote"] = tt.config
+ results, err := NewChatPipelineService().AsyncChat(t.Context(), "user-1", chat,
+ []map[string]interface{}{{"role": "user", "content": "Question"}}, true,
+ map[string]interface{}{"reasoning": 1, "quote": tt.request})
+ if err != nil {
+ t.Fatal(err)
+ }
+ var streamed, reasoning strings.Builder
+ var final AsyncChatResult
+ want := "Answer end"
+ if tt.wantQuote {
+ want = "Answer[ID:0] end"
+ }
+ for result := range results {
+ reasoning.WriteString(result.Reasoning)
+ if result.Final {
+ final = result
+ } else {
+ streamed.WriteString(result.Answer)
+ if !strings.HasPrefix(want, streamed.String()) {
+ t.Fatalf("stream leaked citation text: %q", streamed.String())
+ }
+ }
+ }
+ if !final.Final || final.Answer != want || streamed.String() != want {
+ t.Fatalf("stream=%q final=%+v, want %q", streamed.String(), final, want)
+ }
+ if (len(final.Reference) > 0) != tt.wantQuote {
+ t.Fatalf("references=%#v, quote=%v", final.Reference, tt.wantQuote)
+ }
+ if !tt.wantQuote && (final.Reference == nil || strings.Contains(reasoning.String(), "[ID:")) {
+ t.Fatal("disabled citations must be removed from reasoning and explicitly clear references")
+ }
+ })
+ }
+}
+
+func TestAsyncChatEmptyResponseQuoteDisabled(t *testing.T) {
+ db := setupChatPipelineVisionTestDB(t)
+ seedChatPipelineVisionTenant(t, db, "glm-4-flash@ZHIPU-AI")
+ if err := db.AutoMigrate(&entity.Knowledgebase{}); err != nil {
+ t.Fatal(err)
+ }
+ if err := db.Create(&entity.Knowledgebase{ID: "kb-1", TenantID: "tenant-1", Name: "Knowledge"}).Error; err != nil {
+ t.Fatal(err)
+ }
+
+ chat := dialForTest("glm-4-flash@ZHIPU-AI")
+ chat.KBIDs = entity.JSONSlice{"kb-1"}
+ chat.PromptConfig["system"] = "Answer using {knowledge}"
+ chat.PromptConfig["empty_response"] = "Nothing found."
+ results, err := NewChatPipelineService().AsyncChat(t.Context(), "user-1", chat,
+ []map[string]interface{}{{"role": "user", "content": "Question"}}, true,
+ map[string]interface{}{"quote": false})
+ if err != nil {
+ t.Fatal(err)
+ }
+
+ var final AsyncChatResult
+ for result := range results {
+ if result.Final {
+ final = result
+ }
+ }
+ if !final.Final || final.Answer != "Nothing found." {
+ t.Fatalf("final=%+v, want empty-response result", final)
+ }
+ if final.Reference == nil || len(final.Reference) != 0 {
+ t.Fatalf("quote=false reference = %#v, want explicit empty object", final.Reference)
+ }
+}
+
// TestThinkEventWireKeys pins the JSON keys of a fully populated step: they ARE
// the client contract (the SSE `think_event` field), so renaming or dropping a tag
// breaks every client while neither the compiler nor any Go caller notices.
@@ -2061,7 +2198,7 @@ func TestDecorateHarnessAnswerUncitedOmitsReference(t *testing.T) {
}
s := &ChatPipelineService{}
- res := s.decorateHarnessAnswer("小狼的颜色是灰色的。", kbinfos, nil, nil)
+ res := s.decorateHarnessAnswer("小狼的颜色是灰色的。", kbinfos, nil, nil, true)
if res.Reference != nil {
t.Fatalf("uncited answer must not carry a reference, got %#v", res.Reference)
}
@@ -2085,7 +2222,7 @@ func TestDecorateHarnessAnswerCitesThinkBlock(t *testing.T) {
}
s := &ChatPipelineService{}
- res := s.decorateHarnessAnswer("<think>Supported by [ID:0].</think>小狼的颜色是灰色的。", kbinfos, nil, []string{"c1"})
+ res := s.decorateHarnessAnswer(" thinkingSupported by [ID:0]. response小狼的颜色是灰色的。", kbinfos, nil, []string{"c1"}, true)
if res.Reference == nil {
t.Fatal("a think-block citation must carry a reference")
}
@@ -2115,7 +2252,7 @@ func TestDecorateHarnessAnswerResolvesRenderedPosition(t *testing.T) {
}
s := &ChatPipelineService{}
- res := s.decorateHarnessAnswer("小狼的颜色是灰色的 [ID:0]。", kbinfos, nil, []string{"c1"})
+ res := s.decorateHarnessAnswer("小狼的颜色是灰色的 [ID:0]。", kbinfos, nil, []string{"c1"}, true)
if !strings.Contains(res.Answer, "[ID:0]") {
t.Fatalf("the marker is the client's index and must survive, got %q", res.Answer)
}
@@ -2129,7 +2266,7 @@ func TestDecorateHarnessAnswerResolvesRenderedPosition(t *testing.T) {
// The old numbering (a 1-based position) names no entry in a one-passage
// reference, so it must be dropped rather than shipped as a dead marker.
- res = s.decorateHarnessAnswer("小狼的颜色是灰色的 [ID:1]。", kbinfos, nil, []string{"c1"})
+ res = s.decorateHarnessAnswer("小狼的颜色是灰色的 [ID:1]。", kbinfos, nil, []string{"c1"}, true)
if strings.Contains(res.Answer, "[ID:1]") {
t.Fatalf("an unresolvable marker must be dropped, got %q", res.Answer)
}
@@ -2163,6 +2300,7 @@ func TestDecorateHarnessAnswerOrdersReferenceByRenderedOrder(t *testing.T) {
kbinfos,
nil,
[]string{"c3", "c1"},
+ true,
)
chunks, _ := res.Reference["chunks"].([]map[string]interface{})
if len(chunks) != 3 {
@@ -2203,6 +2341,7 @@ func TestDecorateHarnessAnswerDropsOnlyCanonicalMarkers(t *testing.T) {
kbinfos,
nil,
[]string{"c1"},
+ true,
)
if !strings.Contains(res.Answer, "[2024]") {
t.Errorf("prose brackets must survive, got %q", res.Answer)
@@ -2239,6 +2378,7 @@ func TestDecorateHarnessAnswerKeepsPositionsAfterUnknownBlock(t *testing.T) {
kbinfos,
nil,
[]string{"c1", "gone", "c3"},
+ true,
)
if !strings.Contains(res.Answer, "[ID:2]") {
t.Fatalf("a marker past the unknown block must survive, got %q", res.Answer)
diff --git a/internal/service/citation.go b/internal/service/citation.go
index b4a145bc34..0c4a26429b 100644
--- a/internal/service/citation.go
+++ b/internal/service/citation.go
@@ -477,6 +477,42 @@ func ExpandRangeCitations(answer string, poolSize int) string {
return b.String()
}
+// Keep an unfinished citation until its closing delimiter arrives. Retaining
+// the trailing word also preserves the word boundaries used by refN markers.
+var citationStreamTailPattern = regexp.MustCompile(`(?i)[\[(【][\s*idslot::0-9\x{0660}-\x{0669}\x{06F0}-\x{06F9}–—~~-]*$|#{1,2}(?:[0-9]+\$?)?$|\bref\s*[0-9\x{0660}-\x{0669}\x{06F0}-\x{06F9}]*$|\w+$`)
+
+func stripCitations(text string) string {
+ text = cleanCitationMarkers(text)
+ text = canonicalIDMarkerPattern.ReplaceAllString(text, "")
+ text = slotCitationPattern.ReplaceAllString(text, "")
+ text = rangeCitationPattern.ReplaceAllString(text, "")
+ for _, pattern := range badCitationPatterns {
+ text = pattern.ReplaceAllString(text, "")
+ }
+ return text
+}
+
+type citationStreamFilter struct {
+ pending string
+}
+
+func (f *citationStreamFilter) write(delta string) string {
+ f.pending += delta
+ end := len(f.pending)
+ if tail := citationStreamTailPattern.FindStringIndex(f.pending); tail != nil {
+ end = tail[0]
+ }
+ text := stripCitations(f.pending[:end])
+ f.pending = f.pending[end:]
+ return text
+}
+
+func (f *citationStreamFilter) flush() string {
+ text := stripCitations(f.pending)
+ f.pending = ""
+ return text
+}
+
// ResolveCitationMarkers drops the answer's canonical [ID:n] markers that name
// nothing and returns the pool positions of the markers that do resolve.
//
@@ -529,7 +565,7 @@ func ResolveCitationMarkers(answer string, citeIdx []int) (string, []int) {
// "ID:" prefix — the form the citation rules prescribe, and the only one that can
// be removed from the answer without risking ordinary text like "[2024]".
var canonicalIDMarkerPattern = regexp.MustCompile(
- `(?i)^\[\s*ID\s*[:: ]*\s*[0-9\x{0660}-\x{0669}\x{06F0}-\x{06F9}]+\s*\]$`)
+ `(?i)\[\s*ID\s*[:: ]*\s*[0-9\x{0660}-\x{0669}\x{06F0}-\x{06F9}]+\s*\]`)
// rawCitationMarkers returns the citation numbers the model wrote, in order and
// including duplicates — the repetition is what the citation observability log in
diff --git a/internal/service/citation_test.go b/internal/service/citation_test.go
index 8cfc0808a7..30e4a87945 100644
--- a/internal/service/citation_test.go
+++ b/internal/service/citation_test.go
@@ -495,7 +495,7 @@ func TestDecorateHarnessAnswerExpandsRangeCitations(t *testing.T) {
},
}
s := &ChatPipelineService{}
- res := s.decorateHarnessAnswer("The range claim holds [ID:1-3].", kbinfos, nil, nil)
+ res := s.decorateHarnessAnswer("The range claim holds [ID:1-3].", kbinfos, nil, nil, true)
if strings.Contains(res.Answer, "1-3") {
t.Fatalf("final answer still carries the range citation: %q", res.Answer)
@@ -513,3 +513,19 @@ func TestDecorateHarnessAnswerExpandsRangeCitations(t *testing.T) {
t.Fatalf("reference doc_aggs = %#v, want the single cited doc", res.Reference["doc_aggs"])
}
}
+
+func TestCitationStreamFilter(t *testing.T) {
+ var filter citationStreamFilter
+ want := "Answer\n## Heading\n[2024][guide](https://example.com) end"
+ got := ""
+ for _, delta := range []string{"Answer[", "ID:0][ID:1-", "3][ID:Slot 0](ID: 2)#", "#0", "$", "$\n#", "# Heading\n[2024][guide](https://example.com) end"} {
+ got += filter.write(delta)
+ if !strings.HasPrefix(want, got) {
+ t.Fatalf("stream leaked citation text: %q", got)
+ }
+ }
+ got += filter.flush()
+ if got != want {
+ t.Fatalf("filtered answer=%q, want %q", got, want)
+ }
+}
diff --git a/web/src/components/markdown-content/index.tsx b/web/src/components/markdown-content/index.tsx
index b9c80102c7..507fa31040 100644
--- a/web/src/components/markdown-content/index.tsx
+++ b/web/src/components/markdown-content/index.tsx
@@ -41,7 +41,6 @@ import 'katex/dist/katex.min.css'; // `rehype-katex` does not import the CSS for
import { useFetchDocumentThumbnailsByIds } from '@/hooks/use-document-request';
import { useLoadingPause } from '@/hooks/use-loading-pause';
import {
- currentReg,
escapeUnmatchedAngleBrackets,
parseCitationIndex,
preprocessLaTeX,
@@ -66,6 +65,7 @@ import { sanitizeHtmlWithImagesAsText } from '@/utils/dom-util';
import { SafeImg } from '@/components/safe-img';
const getChunkIndex = (match: string) => parseCitationIndex(match);
+const ReferenceMarkerReg = /(\[(?:ID:)?[0-9\u0660-\u0669\u06F0-\u06F9]+\])/g;
// Wraps every text node so citation markers can be replaced by React elements.
// Defined at module scope: react-markdown rebuilds its whole processor whenever
@@ -326,26 +326,40 @@ const MarkdownContent = ({
const renderReference = useCallback(
(text: string) => {
- const replacedText = reactStringReplace(text, currentReg, (match, i) => {
- const chunkIndex = getChunkIndex(match) as number;
+ const replacedText = reactStringReplace(
+ text,
+ ReferenceMarkerReg,
+ (match, i) => {
+ const chunkIndex = getChunkIndex(match);
+ if (typeof chunkIndex !== 'number') {
+ return match;
+ }
+ const hasReference = !!reference?.chunks?.[chunkIndex];
+ // Explicit markers can render while their sources are still arriving.
+ if (!hasReference && !(loading && match.startsWith('[ID:'))) {
+ return match;
+ }
- return (
- <HoverCard key={i}>
- <HoverCardTrigger>
- <bdi className="text-text-secondary bg-bg-card rounded-2xl px-1 mx-1 text-nowrap inline-block">
- [{chunkIndex + 1}]
- </bdi>
- </HoverCardTrigger>
- <HoverCardContent className="max-w-3xl">
- {getPopoverContent(chunkIndex)}
- </HoverCardContent>
- </HoverCard>
- );
- });
+ return (
+ <HoverCard key={i}>
+ <HoverCardTrigger>
+ <bdi className="text-text-secondary bg-bg-card rounded-2xl px-1 mx-1 text-nowrap inline-block">
+ [{chunkIndex + 1}]
+ </bdi>
+ </HoverCardTrigger>
+ {hasReference && (
+ <HoverCardContent className="max-w-3xl">
+ {getPopoverContent(chunkIndex)}
+ </HoverCardContent>
+ )}
+ </HoverCard>
+ );
+ },
+ );
return replacedText;
},
- [getPopoverContent],
+ [getPopoverContent, loading, reference?.chunks],
);
const dir = getDirAttribute(content.replace(citationMarkerReg, ''));
diff --git a/web/src/hooks/logic-hooks.ts b/web/src/hooks/logic-hooks.ts
index 7c9da6d2c5..2e775d97b2 100644
--- a/web/src/hooks/logic-hooks.ts
+++ b/web/src/hooks/logic-hooks.ts
@@ -32,7 +32,7 @@ import { IKnowledgeFile } from '@/interfaces/database/dataset';
import { changeLanguageAsync } from '@/locales/config';
import api from '@/utils/api';
import { getAuthorization } from '@/utils/authorization-util';
-import { buildMessageUuid } from '@/utils/chat';
+import { buildMessageUuid, mergeAnswerChunk } from '@/utils/chat';
import {
consumeListDeletionMarker,
discardListDeletionMarker,
@@ -365,31 +365,9 @@ export const useSendMessageWithSse = () => {
const d = val?.data;
if (typeof d !== 'boolean') {
setAnswer((prev) => {
- const prevAnswer = prev.answer || '';
- // Skip final-chunk answer only when prior stream chunks exist (avoids duplicate).
- // Empty-response and other single-shot answers arrive with final=true only.
- // const currentAnswer = d.final ? '' : d.answer || '';
- const currentAnswer =
- d.final && prevAnswer ? '' : d.answer || '';
-
- let newAnswer: string;
- if (prevAnswer && currentAnswer.startsWith(prevAnswer)) {
- newAnswer = currentAnswer;
- } else {
- newAnswer = prevAnswer + currentAnswer;
- }
-
- if (d.start_to_think === true) {
- newAnswer = newAnswer + '<think>';
- }
-
- if (d.end_to_think === true) {
- newAnswer = newAnswer + '</think>';
- }
-
return {
...d,
- answer: newAnswer,
+ answer: mergeAnswerChunk(prev.answer || '', d),
conversationId: body?.session_id ?? body?.conversation_id,
chatBoxId: body.chatBoxId,
};
diff --git a/web/src/pages/next-chats/chat-stream/run-stream.ts b/web/src/pages/next-chats/chat-stream/run-stream.ts
index 25c2227c8d..628e38e5ec 100644
--- a/web/src/pages/next-chats/chat-stream/run-stream.ts
+++ b/web/src/pages/next-chats/chat-stream/run-stream.ts
@@ -10,9 +10,9 @@ import {
readJsonSafely,
requestChatCompletionStream,
} from '@/services/chat-completion-stream';
+import { mergeAnswerChunk } from '@/utils/chat';
import { AnswerFlushIntervalMs } from './constants';
import { useChatStreamStore } from './store';
-import { mergeAnswerChunk } from './utils';
export type RunChatCompletionStreamParams = {
conversationId: string;
diff --git a/web/src/pages/next-chats/chat-stream/utils.ts b/web/src/pages/next-chats/chat-stream/utils.ts
index f01e3fcc7c..ec6f3447e8 100644
--- a/web/src/pages/next-chats/chat-stream/utils.ts
+++ b/web/src/pages/next-chats/chat-stream/utils.ts
@@ -2,68 +2,8 @@ import { MessageType } from '@/constants/chat';
import { IAnswer, IMessage } from '@/interfaces/database/chat';
import { buildMessageUuid } from '@/utils/chat';
import { omit } from 'lodash';
-import { CompletionChunk } from '@/services/chat-completion-stream';
import { PrologueMessageIdPrefix } from './constants';
-/**
- * Accumulates a streamed answer chunk onto the previously accumulated text.
- * Ported verbatim from the legacy `setAnswer` reducer in
- * `useSendMessageWithSse` (src/hooks/logic-hooks.ts) to preserve behaviour:
- * the final chunk is skipped only when earlier chunks exist (single-shot
- * answers arrive with `final: true` only), a chunk that already contains the
- * previous text replaces it instead of being appended, and think markers are
- * injected inline.
- */
-export function mergeAnswerChunk(
- previousAnswer: string,
- chunk: CompletionChunk,
-): string {
- // A final chunk carries the COMPLETE answer (citations inserted, references
- // attached server-side). It replaces the accumulated ANSWER — but NOT the
- // thinking accumulated so far, which is the only record of the ReAct
- // trajectory the user watched unfold (the final chunk ships the deliverable
- // alone). A naive final is self-contained (`think + decorated answer`), so
- // it replaces everything wholesale instead of duplicating the think span.
- if (chunk.final === true && chunk.answer) {
- if (chunk.answer.includes('<think>')) {
- return chunk.answer;
- }
- const closeTag = '</think>';
- const closeIdx = previousAnswer.lastIndexOf(closeTag);
- if (closeIdx !== -1) {
- return (
- previousAnswer.slice(0, closeIdx + closeTag.length) +
- '\n\n' +
- chunk.answer
- );
- }
- return chunk.answer;
- }
-
- const currentAnswer = chunk.final && previousAnswer ? '' : chunk.answer || '';
-
- let nextAnswer: string;
- if (previousAnswer && currentAnswer.startsWith(previousAnswer)) {
- nextAnswer = currentAnswer;
- } else {
- nextAnswer = previousAnswer + currentAnswer;
- }
-
- if (chunk.start_to_think === true) {
- nextAnswer = nextAnswer + '<think>';
- }
-
- if (chunk.end_to_think === true) {
- // The closing marker must be followed by a blank line: the answer's first
- // line lands right after it, and a heading (`## Candidate Matrix`) glued
- // onto '</think>' stops being a heading — markdown only parses `#` at
- // line begin.
- nextAnswer = nextAnswer + '</think>\n\n';
- }
-
- return nextAnswer;
-}
-
/** Builds the assistant message that replaces the trailing placeholder. */
export function buildAssistantMessageFromAnswer(answer: IAnswer): IMessage {
return {
diff --git a/web/src/utils/__tests__/chat.test.ts b/web/src/utils/__tests__/chat.test.ts
index 085147e033..ff4aa6aacb 100644
--- a/web/src/utils/__tests__/chat.test.ts
+++ b/web/src/utils/__tests__/chat.test.ts
@@ -1,4 +1,56 @@
-import { preprocessLaTeX, replaceThinkToSection } from '../chat';
+import {
+ mergeAnswerChunk,
+ preprocessLaTeX,
+ replaceThinkToSection,
+} from '../chat';
+
+describe('mergeAnswerChunk', () => {
+ it.each([
+ ['First fact. Second fact.', 'First fact.[ID:0] Second fact.[ID:1]'],
+ ['A fact.[ID:0-1]', 'A fact.[ID:0][ID:1]'],
+ ])('replaces %s with the final cited answer', (previous, answer) => {
+ expect(mergeAnswerChunk(previous, { answer, final: true })).toBe(answer);
+ });
+
+ it('keeps streamed thinking when replacing the visible answer', () => {
+ const streamed = [
+ { start_to_think: true },
+ { answer: 'Checking sources.' },
+ { end_to_think: true },
+ { answer: '## First fact. ' },
+ { answer: 'Second fact.' },
+ ].reduce(mergeAnswerChunk, '');
+ expect(streamed).toBe(
+ '<think>Checking sources.</think>\n\n## First fact. Second fact.',
+ );
+ const answer = mergeAnswerChunk(streamed, {
+ answer: '## First fact.[ID:0] Second fact.[ID:1]',
+ final: true,
+ });
+
+ expect(answer).toBe(
+ '<think>Checking sources.</think>\n\n## First fact.[ID:0] Second fact.[ID:1]',
+ );
+ expect(mergeAnswerChunk(answer, { answer, final: true })).toBe(answer);
+ });
+
+ it.each(['', undefined])(
+ 'keeps the answer for an empty final chunk (%s)',
+ (answer) => {
+ const previous = '<think>Checking sources.</think>A fact.[ID:0]';
+ expect(mergeAnswerChunk(previous, { answer, final: true })).toBe(
+ previous,
+ );
+ },
+ );
+
+ it('accepts a single-shot final answer without duplicating a repeated one', () => {
+ const chunk = { answer: 'A fact.[ID:0]', final: true };
+ const answer = mergeAnswerChunk('', chunk);
+ expect(answer).toBe(chunk.answer);
+ expect(mergeAnswerChunk(answer, chunk)).toBe(answer);
+ });
+});
describe('preprocessLaTeX', () => {
it('converts block \\[ \\] to $$ $$', () => {
diff --git a/web/src/utils/chat.ts b/web/src/utils/chat.ts
index 7041624fb1..382b6ab24e 100644
--- a/web/src/utils/chat.ts
+++ b/web/src/utils/chat.ts
@@ -19,6 +19,7 @@ import {
EmptyConversationId,
} from '@/constants/chat';
import { IMessage, Message } from '@/interfaces/database/chat';
+import type { CompletionChunk } from '@/services/chat-completion-stream';
import { omit } from 'lodash';
import { v4 as uuid } from 'uuid';
import {
@@ -58,6 +59,43 @@ export const buildMessageUuidWithRole = (
return `${message.role}_${message.id}`;
};
+/** Merges streamed deltas, replacing the visible text with a final answer. */
+export function mergeAnswerChunk(
+ previousAnswer: string,
+ chunk: CompletionChunk,
+): string {
+ const currentAnswer = chunk.answer || '';
+
+ let nextAnswer: string;
+ if (chunk.final && currentAnswer) {
+ // Final answers can add or repair citations anywhere in the text.
+ // Keep streamed thinking when the final payload only contains the answer.
+ const thinkEnd = previousAnswer.lastIndexOf('</think>');
+ const thinking =
+ thinkEnd >= 0 &&
+ !currentAnswer.includes('<think>') &&
+ !currentAnswer.includes('</think>')
+ ? previousAnswer.slice(0, thinkEnd + '</think>'.length) + '\n\n'
+ : '';
+ nextAnswer = thinking + currentAnswer;
+ } else if (previousAnswer && currentAnswer.startsWith(previousAnswer)) {
+ nextAnswer = currentAnswer;
+ } else {
+ nextAnswer = previousAnswer + currentAnswer;
+ }
+
+ if (chunk.start_to_think === true) {
+ nextAnswer = nextAnswer + '<think>';
+ }
+
+ if (chunk.end_to_think === true) {
+ // Keep the answer's first Markdown heading on its own line.
+ nextAnswer = nextAnswer + '</think>\n\n';
+ }
+
+ return nextAnswer;
+}
+
// Preprocess LaTeX equations to be rendered by KaTeX
// ref: https://github.com/remarkjs/react-markdown/issues/785
//