Files
audita/internal/modules/homophones/prompt.go

86 lines
4.6 KiB
Go

package homophones
import (
"encoding/json"
"fmt"
"gitea.maximumdirect.net/eric/audita/internal/core/schema"
"gitea.maximumdirect.net/eric/audita/internal/framework/contracts"
"gitea.maximumdirect.net/eric/audita/internal/framework/promptcontext"
)
type promptSegment struct {
ID int `json:"id"`
Speaker string `json:"speaker"`
Start float64 `json:"start"`
End float64 `json:"end"`
Text string `json:"text"`
Categories []string `json:"categories,omitempty"`
}
type promptTranscriptSection struct {
SectionIndex int `json:"section_index"`
Segments []promptSegment `json:"segments"`
}
// BuildProposalMessages constrains corrections to conservative homophone and
// mistranscription updates.
func BuildProposalMessages(transcript *schema.Transcript, glossary *schema.Glossary, sectionIndex int, transcriptDescription string) ([]contracts.LLMMessage, error) {
glossaryJSON, err := json.MarshalIndent(glossary, "", " ")
if err != nil {
return nil, fmt.Errorf("marshal glossary prompt context: %w", err)
}
sectionPayload := promptTranscriptSection{
SectionIndex: sectionIndex,
Segments: make([]promptSegment, 0),
}
if transcript != nil {
for _, s := range transcript.Segments {
sectionPayload.Segments = append(sectionPayload.Segments, promptSegment{
ID: s.ID,
Speaker: s.Speaker,
Start: s.Start,
End: s.End,
Text: s.Text,
Categories: append([]string(nil), s.Categories...),
})
}
}
sectionJSON, err := json.MarshalIndent(sectionPayload, "", " ")
if err != nil {
return nil, fmt.Errorf("marshal transcript prompt context: %w", err)
}
system := "You are Audita, a conservative homophone correction assistant. Identify only transcript changes that plausibly reflect homophones, phonetic similarity, or common mistranscriptions of spoken English. Do not make punctuation, capitalization, spacing, filler-word, repetition, style, or grammar edits. Do not paraphrase, summarize, or rewrite content."
user := "Review this transcript section and return only homophone or spoken-form corrections that should be applied.\n\n" +
"Rules:\n" +
"- Approve only corrections where the original text is plausibly a mistaken homophone, phonetic rendering, or mistranscription of what was likely spoken.\n" +
"- Allow examples such as changing \"dam\" to \"damn\", \"rank\" to \"Hrank\", or \"gestures\" to \"Jesters\" when local context supports the correction.\n" +
"- Reject unrelated substitutions like changing \"Lyra\" to \"Jesters\".\n" +
"- Reject antonyms or reversals such as changing \"visible\" to \"invisible\".\n" +
"- Do not add or remove punctuation, alter capitalization only, normalize spacing, remove filler words, collapse repetitions, or make general readability edits.\n" +
"- Treat glossary names and aliases as protected spellings and context.\n" +
"- You may correct toward glossary names, aliases, or their plural forms when the correction is acoustically plausible and supported by local context.\n" +
"- Do not replace, Anglicize, normalize, lowercase, or otherwise alter protected glossary names or aliases that already appear correctly in the transcript.\n" +
"- Preserve canonical glossary capitalization for protected names and aliases, even if they look unusual.\n" +
"- If a segment includes categories, treat them as additional transcript context.\n" +
"- Use the exact id from the input segment.\n" +
"- For returned corrections, original_text must be only the exact text span that needs replacement, not the full segment text unless the whole segment is the replacement span.\n" +
"- Choose an original_text span that appears exactly once in the current segment text.\n" +
"- corrected_text must be only the replacement text for that span, not the full corrected segment text unless the whole segment is the replacement span.\n" +
"- Each returned correction must contain only id, original_text, corrected_text, and confidence.\n" +
"- Do not return corrections where original_text and corrected_text are identical.\n" +
"- Do not return speaker, start, or end fields.\n" +
"- Return only changed segments; do not return entries for unchanged segments.\n" +
"- confidence must be between 0.0 and 1.0.\n" +
"- If no corrections are needed, return an empty corrections list.\n\n" +
promptcontext.TranscriptDescriptionBlock(transcriptDescription) +
fmt.Sprintf("Protected glossary/context:\n%s\n\nTranscript section:\n%s", string(glossaryJSON), string(sectionJSON))
return []contracts.LLMMessage{
{Role: "system", Content: system},
{Role: "user", Content: user},
}, nil
}