Refactor to request matching/replacement substrings from LLMs, rather than requesting a complete replacement for the full original text

src/audita/corrections.py

@@ -1,5 +1,5 @@
 from dataclasses import asdict, dataclass
-from typing import Dict, Iterable, List, Optional, Tuple
+from typing import Iterable, List, Optional, Tuple
 
 from .errors import AuditaValidationError
 from .schemas import CorrectionCandidate, TranscriptSegment
@@ -32,33 +32,20 @@ def apply_corrections(
     if not 0.0 <= confidence_threshold <= 1.0:
         raise AuditaValidationError("Confidence threshold must be between 0.0 and 1.0.")
 
-    correction_by_id: Dict[int, CorrectionCandidate] = {}
+    revised = list(transcript)
     skipped: List[SkippedCorrection] = []
     for correction in corrections:
-        segment_id = correction.segment_id
-        if segment_id in correction_by_id:
-            skipped.append(
-                _skip(
-                    correction,
-                    "duplicate correction for segment already handled",
-                )
-            )
+        if correction.confidence < confidence_threshold:
             continue
 
-        reason, actual_text = _target_error(transcript, correction)
+        reason, actual_text = _target_error(revised, correction)
         if reason is not None:
             skipped.append(_skip(correction, reason, actual_text=actual_text))
             continue
 
-        correction_by_id[segment_id] = correction
-
-    revised: List[TranscriptSegment] = []
-    for original_id, segment in enumerate(transcript):
-        correction = correction_by_id.get(original_id)
-        if correction is not None and correction.confidence >= confidence_threshold:
-            revised.append(segment.model_copy(update={"text": correction.corrected_text}))
-        else:
-            revised.append(segment)
+        segment = revised[correction.segment_id]
+        revised_text = segment.text.replace(correction.original_text, correction.corrected_text, 1)
+        revised[correction.segment_id] = segment.model_copy(update={"text": revised_text})
 
     indexed_revised = list(enumerate(revised))
     indexed_revised.sort(key=lambda item: (item[1].start, item[1].end, item[0]))
@@ -76,8 +63,16 @@ def _target_error(
         return "segment_id does not exist in transcript", None
 
     segment = transcript[correction.segment_id]
-    if correction.original_text != segment.text:
-        return "original_text does not exactly match segment text", segment.text
+    if correction.original_text == "":
+        return "original_text is empty", segment.text
+    if correction.original_text == correction.corrected_text:
+        return "original_text and corrected_text are identical", segment.text
+
+    match_count = segment.text.count(correction.original_text)
+    if match_count == 0:
+        return "original_text does not match any substring in segment text", segment.text
+    if match_count > 1:
+        return "original_text appears multiple times in segment text", segment.text
 
     return None, None
 

src/audita/prompts.py

@@ -31,11 +31,13 @@ def build_glossary_correction_messages(section: TranscriptSection, glossary: Glo
         "- Inappropriate example: correcting \"Lyra\" to \"Jesters\" should be omitted because those words are not similar in spoken English, even if \"Jesters\" appears in the glossary.\n"
         "- Do not replace one clear glossary term, character name, location, or ordinary word with a different glossary term unless it is a plausible mishearing.\n"
         "- Assign high confidence only when the correction is supported by glossary evidence, local context, and spoken-word similarity; otherwise omit the correction.\n"
-        "- Use the exact segment_id and original_text from the input segment.\n"
+        "- Use the exact segment_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"
+        "- 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 segment_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"
-        "- corrected_text must contain the full corrected text for that segment.\n"
         "- confidence must be between 0.0 and 1.0.\n"
         "- If no corrections are needed, return an empty corrections list.\n\n"
         f"Glossary:\n{glossary_json}\n\n"