refactor: prioritize cognitive interpretation over deterministic rules
CI / test (push) Failing after 14s

This commit is contained in:
m
2026-08-18 17:46:14 +03:00
parent 2e0c0236aa
commit d50d709d69
4 changed files with 198 additions and 46 deletions
+116 -30
View File
@@ -175,7 +175,8 @@ func semanticScores(tokens []string, hasQuestionMark bool, explicitCorrection bo
interaction.SpeechActConversational: 0,
}
if hasQuestionMark {
scores[interaction.SpeechActQuestion] += 3
scores[interaction.SpeechActQuestion] += 4
scores[interaction.SpeechActRequestToInvestigate]++
}
if explicitCorrection {
scores[interaction.SpeechActCorrection] += 6
@@ -192,14 +193,17 @@ func semanticScores(tokens []string, hasQuestionMark bool, explicitCorrection bo
scores[interaction.SpeechActConfirmation] += 2
}
questionWords := map[string]struct{}{"кто": {}, "что": {}, "как": {}, "почему": {}, "зачем": {}, "где": {}, "когда": {}, "сколько": {}, "какой": {}, "какая": {}, "какие": {}, "чей": {}}
conversationWords := []string{"привет", "здравствуйте", "спасибо", "ку", "здравствуй", "добрый", "день", "добрый день"}
investigateStems := []string{"разбер", "проверь", "проанализ", "исслед", "выясн", "посмотри", "узнай", "расскаж", "объясн", "поясн"}
directiveStems := []string{"сдел", "запуст", "созд", "настро", "добав", "измени", "помен", "поменя", "измен", "использ", "удал", "напом", "подключ", "исправ", "обрати", "поправ", "обнов", "ответь", "посмотри"}
conversationWords := []string{"привет", "здравствуйте", "спасибо", "ку", "здравствуй", "добрый"}
investigateStems := []string{"разбер", "проверь", "провери", "проанализ", "исслед", "выясн", "узнай"}
directiveStems := []string{"сдел", "запуст", "созд", "настро", "добав", "измени", "помен", "поменя", "измен", "использ", "удал", "напом", "подключ", "исправ", "обрати", "поправ", "обнов", "ответь", "посмотри", "проверь"}
preferenceStems := []string{"предпоч", "хочу", "лучше", "не используй"}
hypothesisStems := []string{"думаю", "кажется", "гипотез", "похоже"}
decisionStems := []string{"решил", "решение", "выбираю"}
arithmeticMarker := false
for index, token := range tokens {
if firstPersonPastReport(tokens, index) && containsStem(token, directiveStems...) {
continue
}
if _, ok := questionWords[token]; ok && !(token == "что" && index+1 < len(tokens) && tokens[index+1] == "нибудь") {
scores[interaction.SpeechActQuestion] += 4
}
@@ -226,37 +230,24 @@ func semanticScores(tokens []string, hasQuestionMark bool, explicitCorrection bo
arithmeticMarker = true
}
}
if len(tokens) >= 2 && len(tokens[0]) >= 4 && (strings.HasSuffix(tokens[0], "й") || strings.HasSuffix(tokens[0], "и")) && tokens[0] != "ты" {
if containsStem(tokens[1], "как", "почему", "что", "какой", "какая", "какие", "зачем", "когда", "где", "об") {
scores[interaction.SpeechActQuestion] += 3
scores[interaction.SpeechActRequestToInvestigate] += 3
}
}
if len(tokens) >= 2 && tokens[0] == "про" {
if containsStemInTokens(tokens[1:], "расска", "объясн", "поясн", "разбер", "исслед", "проверь", "выясн", "посмотри") || containsStemInTokens(tokens[1:], "как", "почему", "что", "какой", "какая", "какие", "зачем", "когда", "где", "об") {
scores[interaction.SpeechActQuestion] += 4
scores[interaction.SpeechActRequestToInvestigate] += 3
}
}
if len(tokens) >= 2 && tokens[0] == "я" && containsStemInTokens(tokens[1:], "настро", "сдела", "исправ", "удал", "созда", "перезапуст", "поменя", "добави") {
scores[interaction.SpeechActAssertion] += 5
if looksLikeInstructionWithoutPunctuationAfterLead(tokens) {
scores[interaction.SpeechActQuestion] += 4
scores[interaction.SpeechActRequestToInvestigate] += 2
}
if arithmeticMarker {
scores[interaction.SpeechActQuestion] += 2
}
if len(tokens) > 0 && strings.HasSuffix(tokens[0], "ешь") {
scores[interaction.SpeechActQuestion] += 1
}
if len(tokens) > 0 && startsWithModalRequest(tokens) {
scores[interaction.SpeechActQuestion] += 2
scores[interaction.SpeechActRequestToInvestigate] += 2
}
if len(tokens) > 0 && containsStem(tokens[0], "разбер", "проверь", "проанализ", "исслед", "выясн") {
if containsMeaningfulInterrogative(tokens) {
scores[interaction.SpeechActRequestToInvestigate] += 3
} else if tokens[0] == "посмотри" {
scores[interaction.SpeechActRequestToInvestigate] += 1
}
if looksLikeInformationRequest(tokens) {
scores[interaction.SpeechActQuestion] += 4
scores[interaction.SpeechActRequestToInvestigate] += 2
}
if !hasQuestionMark && containsInterrogative(tokens) {
scores[interaction.SpeechActQuestion] += 4
scores[interaction.SpeechActRequestToInvestigate] += 2
}
if containsModelReference(tokens) {
scores[interaction.SpeechActQuestion] += 2
@@ -353,6 +344,12 @@ func needsCognitionForCandidate(primary, secondary interaction.SpeechAct, tokens
if primary == interaction.SpeechActConversational && secondary == interaction.SpeechActDirective {
return true
}
if primary == interaction.SpeechActQuestion && containsMeaningfulInterrogative(tokens) {
return true
}
if isAssistantStatusProbe(primary, secondary, tokens) {
return true
}
if primary == interaction.SpeechActAssertion && containsModelReference(tokens) && containsStemInTokens(tokens, "работает", "падает", "растет", "растёет", "свободно", "виснет") {
return true
}
@@ -362,9 +359,6 @@ func needsCognitionForCandidate(primary, secondary interaction.SpeechAct, tokens
if secondary == interaction.SpeechActQuestion || secondary == interaction.SpeechActRequestToInvestigate {
return true
}
if isAssistantStatusProbe(primary, secondary, tokens) {
return true
}
if primary == interaction.SpeechActDirective && (containsInterrogative(tokens) || containsModelReference(tokens) || containsStemInTokens(tokens, "обрати", "посмотри", "разбер", "исправ", "обнов", "добав")) {
return true
}
@@ -405,9 +399,23 @@ func capabilityHintsFor(primary, secondary interaction.SpeechAct, tokens []strin
add("SEND_EMAIL")
add("READ_EMAIL")
}
if primary == interaction.SpeechActDirective && containsStemInTokens(tokens, "добав", "реализ", "поддержк") {
add("MODIFY_CODE")
}
if containsToken(tokens, "напомни") || containsToken(tokens, "напомин") {
add("SCHEDULE_REMINDER")
}
if primary == interaction.SpeechActRequestToInvestigate || (primary == interaction.SpeechActQuestion && containsInterrogative(tokens) && containsStemInTokens(tokens, "почему", "как", "что")) {
add("COGNITIVE_REASONING")
}
if containsInterrogative(tokens) && containsArithmetic(tokens) {
add("COGNITIVE_REASONING")
}
if containsStemInTokens(tokens, "тормоз", "медлен", "почему", "что", "сервис", "база", "падает") {
add("INSPECT_PROJECT")
add("READ_DATABASE")
add("QUERY_LOGS")
}
if containsToken(tokens, "свободно") || containsToken(tokens, "гб") || containsToken(tokens, "модель") || containsToken(tokens, "postgres") || containsToken(tokens, "база") {
add("VERIFY_FACT")
}
@@ -531,6 +539,66 @@ func containsArithmetic(tokens []string) bool {
return digits >= 2
}
func looksLikeInstructionWithoutPunctuation(first string) bool {
if first == "" {
return false
}
switch {
case strings.HasSuffix(first, "й"), strings.HasSuffix(first, "и"), strings.HasSuffix(first, "ите"):
return utf8.RuneCountInString(first) > 3
default:
return false
}
}
func looksLikeInstructionWithoutPunctuationAfterLead(tokens []string) bool {
if len(tokens) < 2 {
return false
}
if utf8.RuneCountInString(tokens[0]) > 3 {
return false
}
switch tokens[0] {
case "про", "по", "на", "в", "за":
default:
return false
}
for _, token := range tokens[1:] {
if utf8.RuneCountInString(token) <= 3 {
continue
}
switch {
case strings.HasSuffix(token, "ешь"), strings.HasSuffix(token, "и"), strings.HasSuffix(token, "й"), strings.HasSuffix(token, "те"), strings.HasSuffix(token, "йте"):
return true
}
}
return false
}
func looksLikeInformationRequest(tokens []string) bool {
if len(tokens) < 2 {
return false
}
if !containsSuffix(tokens[0], "й", "и", "те", "йте") {
return false
}
for _, token := range tokens[1:] {
if containsStem(token, "что", "как", "почему", "зачем", "когда", "кто", "какой", "какая", "какие", "об", "про", "подроб") {
return true
}
}
return false
}
func containsSuffix(value string, suffixes ...string) bool {
for _, suffix := range suffixes {
if strings.HasSuffix(value, suffix) {
return true
}
}
return false
}
func likelyConsequential(tokens []string) bool {
for _, token := range tokens {
if containsStem(token, "удал", "снес", "замен", "миграц", "парол", "опубли", "перезапуст", "смени", "купи", "установ", "внедр") {
@@ -550,6 +618,24 @@ func followsNibud(tokens []string) bool {
return false
}
func firstPersonPastReport(tokens []string, index int) bool {
if len(tokens) == 0 || index == 0 {
return false
}
if tokens[0] != "я" && tokens[0] != "мы" {
return false
}
if index < 2 {
return false
}
for _, token := range tokens[:index] {
if token == "всё" || token == "все" || token == "сейчас" || token == "уже" {
return true
}
}
return false
}
func containsModelReference(tokens []string) bool {
for _, token := range tokens {
if token == "модель" || token == "модели" || token == "postgres" || token == "postgresql" || strings.Contains(token, "gemma") || strings.Contains(token, "гемма") || strings.Contains(token, "qwen") {
@@ -129,6 +129,8 @@ func TestInterpretUtteranceNaturalRussianHelpflowUsesContextAwareCognitionInput(
{message: "Расскажи как это работает", wantSpeech: "", wantCognition: true, allowNoMatch: true},
{message: "можешь объяснить почему база растёт", wantSpeech: "", wantCognition: true, allowNoMatch: true},
{message: "ПОЧЕМУ postgres использует так много памяти", wantSpeech: "", wantCognition: true, allowNoMatch: true},
{message: "ПОЧЕМУ postgres использует столько памяти", wantSpeech: "", wantCognition: true, allowNoMatch: true},
{message: "разберись почему сервис упал", wantSpeech: "", wantCognition: true, allowNoMatch: true},
{message: "разберись почему сервис упал", wantSpeech: "", wantCognition: true, allowNoMatch: true},
{message: "посмотри что произошло", wantSpeech: "", wantCognition: true, allowNoMatch: true},
{message: "исправь ошибку если найдёшь", wantSpeech: "", wantCognition: true, allowNoMatch: true},