package asr

import (
	"strings"
	"unicode"
)

// looksLikePromptLeak reports whether text appears to be the model echoing
// the --prompt biasing content back verbatim instead of genuinely
// transcribing the segment's audio — a known failure mode of
// prompt-conditioned generation, confirmed via real testing against
// qwen_asr (see transcriptor-ui/CLAUDE.md, "Glossary prompt-leakage bug"):
// a whole segment came back as literally the glossary terms, twice, on two
// separate 5-minute test runs, even after switching to the tool's
// recommended "Preserve spelling: ..." prompt framing (which reduces but
// does not eliminate the risk).
//
// Heuristic: if most of text's words also appear among prompt's words, and
// there are enough of them to rule out a legitimate short mention of one
// or two glossary terms in real speech, treat it as a leak. Both observed
// real leaks were near-total reproductions of the prompt's term list, so a
// strict threshold (80% word overlap, at least 4 words) catches them
// without plausibly false-flagging a normal sentence that happens to
// mention "Claude" or "Mistral" once — normal speech has enough other
// words to keep the overlap ratio well below that.
func looksLikePromptLeak(text, prompt string) bool {
	if prompt == "" || text == "" {
		return false
	}
	textWords := normalizeWords(text)
	if len(textWords) < 4 {
		return false
	}
	promptWords := wordSet(normalizeWords(prompt))

	matched := 0
	for _, w := range textWords {
		if promptWords[w] {
			matched++
		}
	}
	return float64(matched)/float64(len(textWords)) >= 0.8
}

// normalizeWords lowercases and splits on anything that isn't a letter or
// digit — unicode.IsLetter (not an ASCII-only check) so accented
// characters (French: é, è, ï, ç, ...) stay part of their word instead of
// being split into fragments.
func normalizeWords(s string) []string {
	return strings.FieldsFunc(strings.ToLower(s), func(r rune) bool {
		return !unicode.IsLetter(r) && !unicode.IsDigit(r)
	})
}

func wordSet(words []string) map[string]bool {
	set := make(map[string]bool, len(words))
	for _, w := range words {
		set[w] = true
	}
	return set
}