fix(accuracy): sentiment reason, summarization bullet fmt, NER escalation schema

handlers/remote_handlers.py

@@ -20,17 +20,31 @@ def __init__(self) -> None:
         self.system_prompt = load_prompt_template("remote_general.txt")
 
     async def handle(self, prompt: str, category: str = "LOCAL_GENERAL") -> str:
-        # Set max_tokens based on category
+        # Category-specific token budgets
         max_tokens = {
             "LOCAL_GENERAL": 80,
-            "LOCAL_SENTIMENT": 20,
+            "LOCAL_SENTIMENT": 60,  # label + one-sentence reason
             "LOCAL_NER": 200,
             "API_LONG_CONTEXT": 200,
         }.get(category, 80)
 
+        # Preserve schema constraints for structured-output categories
+        if category == "LOCAL_NER":
+            # Must keep JSON list format intact — remote_general.txt is unstructured prose
+            system_prompt = load_prompt_template("ner.txt")
+        elif category == "LOCAL_SENTIMENT":
+            system_prompt = (
+                "Classify the sentiment of the text as Positive, Negative, or Neutral. "
+                "Then provide exactly one sentence explaining your reasoning. "
+                "Format: <LABEL>. <one sentence reason>  "
+                "Example: Neutral. The packaging was damaged but the product itself works perfectly."
+            )
+        else:
+            system_prompt = self.system_prompt
+
         return await self.engine.generate(
             prompt=prompt,
             category=category,
-            system_prompt=self.system_prompt,
+            system_prompt=system_prompt,
             max_tokens=max_tokens,
         )

handlers/sentiment.py

@@ -8,10 +8,10 @@
 _VALID = {"Positive", "Negative", "Neutral"}
 
 _SYSTEM_PROMPT = (
-    "You are a zero-filler text classification engine. "
-    "Analyze the input text sentiment. Output EXACTLY one word "
-    "from these options: Positive, Negative, Neutral. "
-    "No preamble. No explanations. No markdown formatting."
+    "Classify the sentiment of the text as Positive, Negative, or Neutral. "
+    "Then provide exactly one sentence explaining your reasoning. "
+    "Format your response EXACTLY as: <LABEL>. <one sentence reason> "
+    "Example: Neutral. The packaging was damaged but the product itself works perfectly."
 )
 
 
@@ -30,16 +30,19 @@ def handle(self, prompt: str) -> str:
         res = self.engine.generate(
             f"Text: {prompt}\nSentiment:",
             system_prompt=_SYSTEM_PROMPT,
-            max_tokens=2,
+            max_tokens=60,  # enough for "<Label>. <one-sentence reason>"
             temperature=0.0,
         )
-        # __ESCALATE__ is unreachable at max_tokens=2 (no length truncation),
-        # but guard defensively.
         if res == "__ESCALATE__":
             return "Neutral"
-        cleaned = res.strip().title()
+        cleaned = res.strip()
+        # Return the full label+reason string if it starts with a valid label
         for label in _VALID:
-            if label in cleaned:
-                return label
+            if cleaned.startswith(label):
+                return cleaned
+        # Fallback: scan for label anywhere in response (model omitted format)
+        for label in _VALID:
+            if label in cleaned.title():
+                return cleaned
         logger.info("SentimentHandler: non-standard output '%s'. Defaulting to Neutral.", res)
         return "Neutral"

handlers/summarization.py

@@ -11,7 +11,13 @@
     re.IGNORECASE,
 )
 
-_SYSTEM_PROMPT = "Summarize the text in under 3 concise sentences. " "Output ONLY the final summary. No intro. No structural wrappers."
+_SYSTEM_PROMPT_PROSE = "Summarize the text in under 3 concise sentences. " "Output ONLY the final summary. No intro. No structural wrappers."
+
+_SYSTEM_PROMPT_BULLETS = (
+    "Summarize the text in EXACTLY 3 bullet points. "
+    "Each bullet point must be 15 words or fewer. "
+    "Start each bullet with '• '. No intro, no headers, no extra text."
+)
 
 _WORD_COUNT_THRESHOLD = 1200
 
@@ -37,14 +43,24 @@ def handle(self, prompt: str) -> str:
         word_count = len(cleaned.split())
         if word_count >= _WORD_COUNT_THRESHOLD:
             logger.info(
-                "SummarizationHandler: word_count=%d ≥ threshold=%d → __ESCALATE__",
+                "SummarizationHandler: word_count=%d >= threshold=%d → __ESCALATE__",
                 word_count,
                 _WORD_COUNT_THRESHOLD,
             )
             return "__ESCALATE__"
+
+        # Detect bullet-format directive in the original prompt
+        p_lower = prompt.lower()
+        if any(w in p_lower for w in ["bullet", "bullet point", "• ", "points"]):
+            system_prompt = _SYSTEM_PROMPT_BULLETS
+            max_tok = 120  # 3 bullets × ~15 words × ~1.3 tokens/word
+        else:
+            system_prompt = _SYSTEM_PROMPT_PROSE
+            max_tok = 90
+
         return self.engine.generate(
             cleaned,
-            system_prompt=_SYSTEM_PROMPT,
-            max_tokens=90,
+            system_prompt=system_prompt,
+            max_tokens=max_tok,
             temperature=0.1,
         )

prompts/sentiment.txt

@@ -1,4 +1,6 @@
 You are a sentiment classification agent.
 Analyze the user's text and classify its sentiment as exactly one of: Positive, Negative, or Neutral.
-You must output ONLY one of those three words, with no explanations, no preamble, and no extra punctuation.
-If the input text is ambiguous, meaningless, or you are unsure, output: __ESCALATE__
+Then provide exactly one sentence explaining your reasoning.
+Format your response EXACTLY as: <LABEL>. <one sentence reason>
+Example: Neutral. The packaging was damaged but the product itself works perfectly.
+Do not include any preamble, extra punctuation, or additional sentences.