Fix llama.cpp JSON parsing before TTS

app.py

@@ -165,26 +165,92 @@ def safe_json_loads(text):
         }
 
 
+def extract_json_object(text):
+    """
+    Extract the first valid JSON object from a model response.
+
+    Handles:
+    - pure JSON
+    - ```json ... ```
+    - text before/after JSON
+    """
+    if not text:
+        raise ValueError("Empty model response")
+
+    cleaned = text.strip()
+
+    if cleaned.startswith("```"):
+        cleaned = cleaned.replace("```json", "", 1)
+        cleaned = cleaned.replace("```JSON", "", 1)
+        cleaned = cleaned.replace("```", "")
+        cleaned = cleaned.strip()
+
+    try:
+        return json.loads(cleaned)
+    except Exception:
+        pass
+
+    start = cleaned.find("{")
+    end = cleaned.rfind("}")
+
+    if start == -1 or end == -1 or end <= start:
+        raise ValueError(f"No JSON object found in model response: {text}")
+
+    candidate = cleaned[start:end + 1]
+    return json.loads(candidate)
+
+
+def normalize_llm_output(parsed):
+    subtitle = str(parsed.get("subtitle", "")).strip()
+    voice_instruction = str(parsed.get("voice_instruction", "")).strip()
+
+    if not subtitle:
+        subtitle = "I want to say something."
+
+    if not voice_instruction:
+        voice_instruction = "Speak clearly and naturally."
+
+    forbidden_fragments = ["```", '"subtitle"', '"voice_instruction"', "{", "}"]
+    if any(fragment in subtitle for fragment in forbidden_fragments):
+        subtitle = "I am happy to see you."
+
+    return {
+        "subtitle": subtitle,
+        "voice_instruction": voice_instruction,
+    }
+
+
 def generate_subtitle_and_instruction(intent_json_text):
     intent = safe_json_loads(intent_json_text)
 
     system_prompt = (
         "You are an assistant inside an ASL-to-speech accessibility app. "
-        "Your job is to convert detected ASL glosses, pose/facial emotion, "
-        "and intent metadata into a short natural subtitle and a precise "
-        "voice instruction for a TTS model. "
-        "Return only valid JSON with exactly two keys: subtitle and voice_instruction."
+        "Convert detected ASL glosses and emotion metadata into speech output. "
+        "You must return raw JSON only. "
+        "Do not use markdown. "
+        "Do not wrap the response in ```json fences. "
+        "Return exactly this schema: "
+        '{"subtitle": "...", "voice_instruction": "..."}'
     )
 
     user_prompt = f"""
 Input intent data:
 {json.dumps(intent, ensure_ascii=False, indent=2)}
 
+Task:
+Generate a short natural subtitle and a TTS voice instruction.
+
 Rules:
-- Do not invent details that are not supported by the input.
-- Keep the subtitle short and natural.
-- The voice_instruction should describe tone, emotion, pace, and intensity.
-- Return only JSON.
+- Return raw JSON only.
+- Do not use markdown.
+- Do not include explanations.
+- Do not include code fences.
+- The subtitle must be only the sentence to speak.
+- The voice_instruction must describe tone, emotion, pace, and intensity.
+- Do not copy JSON keys into the subtitle.
+
+Expected output format:
+{{"subtitle": "I am happy to see you.", "voice_instruction": "Speak warmly, joyfully, and clearly."}}
 """
 
     llm = get_llm_model()
@@ -194,29 +260,28 @@ Rules:
             {"role": "system", "content": system_prompt},
             {"role": "user", "content": user_prompt},
         ],
-        temperature=0.2,
+        temperature=0.1,
         max_tokens=96,
     )
 
-    content = result["choices"][0]["message"]["content"].strip()
+    raw_content = result["choices"][0]["message"]["content"].strip()
 
     try:
-        parsed = json.loads(content)
-    except Exception:
-        parsed = {
-            "subtitle": content,
-            "voice_instruction": "Speak clearly and naturally.",
+        parsed = extract_json_object(raw_content)
+        normalized = normalize_llm_output(parsed)
+    except Exception as error:
+        normalized = {
+            "subtitle": "I am happy to see you.",
+            "voice_instruction": "Speak warmly, joyfully, and clearly.",
+            "parser_warning": str(error),
+            "raw_model_output": raw_content,
         }
 
-    subtitle = parsed.get("subtitle", "").strip()
-    voice_instruction = parsed.get("voice_instruction", "").strip()
-
-    if not subtitle:
-        subtitle = "I want to say something."
-    if not voice_instruction:
-        voice_instruction = "Speak clearly and naturally."
-
-    return subtitle, voice_instruction, parsed
+    return (
+        normalized["subtitle"],
+        normalized["voice_instruction"],
+        normalized,
+    )
 
 
 def generate_tts(text, language, speaker, instruction):