fix(suggestion): avoid echoing the current query (#2582)

* fix(suggestion): avoid echoing the current query

* test(suggestion): use generic follow-up examples
main
HJX|翻斗花园二蛋 2026-09-21 13:06:01 +08:00 committed by GitHub
parent a3c7be10ca
commit 2f2bcc7aa7
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 119 additions and 1 deletions

View File

@ -379,6 +379,11 @@ func (s *messageSuggestionService) generateWithModel(
return nil, types.TokenUsage{}, err
}
items, err := parseGeneratedSuggestions(response.Content, config.Categories, count)
if err == nil {
// The system prompt already asks the model not to repeat prior user
// questions, but that is guidance the model does not always honour.
items = filterSuggestionItemsAgainstQuery(items, generationContext.CurrentQuery)
}
return items, response.Usage, err
}
@ -469,7 +474,18 @@ func (s *messageSuggestionService) generateFromKnowledge(
items := make(types.SuggestionItems, 0, len(candidates))
for _, candidate := range candidates {
text := strings.TrimSpace(candidate.Question)
if text == "" {
// The knowledge path never reaches the model, so the "do not repeat
// prior user questions" system-prompt rule cannot apply here. Worse,
// rankKnowledgeSuggestions ranks candidates against a context that
// starts with CurrentQuery and IsContentContained adds +1, so a
// candidate identical to the question just asked is actively promoted
// to the top of the knowledge pool.
// Typical trigger: the user asks "介绍一下X" while X has an
// entity/summary wiki page, which wikiSuggestionFromPage turns into
// the very same "介绍一下X".
// Skip via continue rather than filtering afterwards so that later
// candidates can backfill and the configured count is preserved.
if text == "" || suggestionMatchesQuery(text, generationContext.CurrentQuery) {
continue
}
item := types.SuggestionItem{
@ -916,6 +932,36 @@ func mergeHybridSuggestionItems(model, knowledge types.SuggestionItems, limit in
return result
}
// filterSuggestionItemsAgainstQuery drops suggestions that are, after
// normalization, exactly the question the user just asked.
// Only exact matches are removed, deliberately: a similarity threshold would
// also kill legitimate follow-ups about the same entity, which is precisely
// what the feature exists for.
func filterSuggestionItemsAgainstQuery(
items types.SuggestionItems,
currentQuery string,
) types.SuggestionItems {
if normalizeSuggestionText(currentQuery) == "" {
return items
}
filtered := make(types.SuggestionItems, 0, len(items))
for _, item := range items {
if suggestionMatchesQuery(item.Text, currentQuery) {
continue
}
filtered = append(filtered, item)
}
return filtered
}
// suggestionMatchesQuery reports whether a candidate is the current question.
// It reuses normalizeSuggestionText so punctuation, whitespace and case
// variants ("Tell me about Foo?" vs "tell me about foo") are caught too.
func suggestionMatchesQuery(value, currentQuery string) bool {
queryKey := normalizeSuggestionText(currentQuery)
return queryKey != "" && normalizeSuggestionText(value) == queryKey
}
func normalizeSuggestionText(value string) string {
return strings.Map(func(r rune) rune {
if unicode.IsSpace(r) || strings.ContainsRune("??!!,,.。::;;\"'", r) {

View File

@ -33,6 +33,78 @@ func TestParseGeneratedSuggestionsFiltersAndDeduplicates(t *testing.T) {
}
}
func TestFilterSuggestionItemsAgainstQueryDropsNormalizedEchoes(t *testing.T) {
const currentQuery = "介绍一下手冲咖啡"
items := types.SuggestionItems{
{ID: "model-echo", Text: "介绍一下 手冲咖啡?", Source: "model"},
{ID: "wiki-echo", Text: "介绍一下手冲咖啡。", Source: "wiki"},
{ID: "keep", Text: "手冲咖啡适合用什么水温?", Source: "model"},
}
got := filterSuggestionItemsAgainstQuery(items, currentQuery)
if len(got) != 1 || got[0].ID != "keep" {
t.Fatalf("filterSuggestionItemsAgainstQuery() = %#v", got)
}
}
func TestFilterSuggestionItemsAgainstQueryKeepsAllWhenQueryEmpty(t *testing.T) {
items := types.SuggestionItems{
{ID: "1", Text: "A?", Source: "model"},
{ID: "2", Text: "B?", Source: "wiki"},
}
for _, query := range []string{"", " ", "??"} {
got := filterSuggestionItemsAgainstQuery(items, query)
if len(got) != 2 {
t.Fatalf("query %q: len = %d, want 2", query, len(got))
}
}
}
func TestSuggestionMatchesQueryIgnoresPunctuationAndCase(t *testing.T) {
cases := []struct {
value, query string
want bool
}{
{"Tell me about Foo?", "tell me about foo", true},
{"介绍一下X", "介绍一下X?", true},
{"介绍一下 X", "介绍一下X", true},
{"什么是X?", "介绍一下X", false},
{"介绍一下XY", "介绍一下X", false},
{"介绍一下X", "", false},
}
for _, c := range cases {
if got := suggestionMatchesQuery(c.value, c.query); got != c.want {
t.Fatalf("suggestionMatchesQuery(%q, %q) = %v, want %v", c.value, c.query, got, c.want)
}
}
}
// Removing the echo must not collapse the hybrid layout: the knowledge slot
// should be filled by the next knowledge candidate, not stolen by the model.
func TestMergeHybridKeepsLayoutAfterEchoRemoved(t *testing.T) {
const currentQuery = "介绍一下手冲咖啡"
model := types.SuggestionItems{
{ID: "m1", Text: "手冲咖啡适合用什么水温?", Source: "model"},
{ID: "m2", Text: "如何选择咖啡豆?", Source: "model"},
}
knowledge := filterSuggestionItemsAgainstQuery(types.SuggestionItems{
{ID: "k-echo", Text: currentQuery, Source: "wiki"},
{ID: "k2", Text: "咖啡豆应该如何保存?", Source: "wiki"},
}, currentQuery)
got := mergeHybridSuggestionItems(model, knowledge, 3)
if len(got) != 3 {
t.Fatalf("len = %d, want 3: %#v", len(got), got)
}
if got[0].ID != "m1" || got[1].ID != "m2" || got[2].ID != "k2" {
t.Fatalf("mergeHybridSuggestionItems() = %#v", got)
}
for _, item := range got {
if item.Text == currentQuery {
t.Fatalf("follow-up still echoes the current question: %#v", got)
}
}
}
func TestMergeSuggestionItemsPreservesPriorityAndLimit(t *testing.T) {
primary := types.SuggestionItems{{ID: "1", Text: "A?", Source: "model"}}
fallback := types.SuggestionItems{