Update evo_inference.py

evo_inference.py

@@ -1,12 +1,16 @@
 """
-evo_inference.py — Step 8
+evo_inference.py — Step 8 (refined)
 Adds a GENERATIVE path using a small plugin (FLAN-T5 stand-in) while keeping the
 old EXTRACTIVE fallback (bullet points) if generation isn't available.
 
+What's new in this refinement:
+- Answers are explicitly labeled **[Generative]** or **[Extractive]** so you
+  can tell which path ran at a glance.
+
 How it works:
 - We try to import your real evo plugin (evo_plugin.py). If not found, we load
   evo_plugin_example.py instead. If both fail, we stay in extractive mode.
-- synthesize_with_evo(...) now accepts mode/temp/max_tokens from the UI.
+- synthesize_with_evo(...) accepts mode/temp/max_tokens from the UI.
 """
 
 from typing import List, Dict
@@ -26,14 +30,17 @@ except Exception:
 
 MAX_SNIPPET_CHARS = 400
 
+
 def _snippet(text: str) -> str:
     text = " ".join(text.split())
     return text[:MAX_SNIPPET_CHARS] + ("..." if len(text) > MAX_SNIPPET_CHARS else "")
 
+
 def _extractive_answer(user_query: str, lang: str, hits: List[Dict]) -> str:
-    """Old safe mode: show top snippets + standard steps."""
+    """Old safe mode: show top snippets + standard steps, now labeled."""
     if not hits:
-        return L(lang, "intro_err")
+        return "**[Extractive]**\n\n" + L(lang, "intro_err")
+
     bullets = [f"- {_snippet(h['text'])}" for h in hits[:4]]
     steps = {
         "en": [
@@ -55,13 +62,16 @@ def _extractive_answer(user_query: str, lang: str, hits: List[Dict]) -> str:
             "• Step 4: Gard referans/reso; swiv letan tretman.",
         ],
     }[normalize_lang(lang)]
+
     return (
+        "**[Extractive]**\n\n"
         f"**{L(lang, 'intro_ok')}**\n\n"
         f"**Q:** {user_query}\n\n"
         f"**Key information:**\n" + "\n".join(bullets) + "\n\n"
         f"**Suggested steps:**\n" + "\n".join(steps)
     )
 
+
 def _build_grounded_prompt(question: str, lang: str, hits: List[Dict]) -> str:
     """Create a compact prompt that includes the question + top retrieved snippets."""
     lang = normalize_lang(lang)
@@ -76,6 +86,7 @@ def _build_grounded_prompt(question: str, lang: str, hits: List[Dict]) -> str:
         system = ("You are the Mauritius Government Copilot. Answer clearly and step-by-step using "
                   "ONLY the provided context. Include: required documents, fees, where to apply, "
                   "processing time. State if anything is missing.")
+
     ctx = "\n".join([f"[Context #{i+1}] {_snippet(h['text'])}" for i, h in enumerate(hits[:6])]) or "[Context] (none)"
     return (
         f"{system}\n\n[Question]\n{question}\n\n{ctx}\n\n"
@@ -83,6 +94,7 @@ def _build_grounded_prompt(question: str, lang: str, hits: List[Dict]) -> str:
         f"- Do not invent links/fees\n- Answer in language code: {lang}\n[Answer]\n"
     )
 
+
 def synthesize_with_evo(
     user_query: str,
     lang: str,
@@ -91,13 +103,32 @@ def synthesize_with_evo(
     max_new_tokens: int = 192,
     temperature: float = 0.4,
 ) -> str:
-    """If mode=='generative' and a generator exists, generate; else use extractive fallback."""
+    """
+    If mode=='generative' and a generator exists, generate a grounded answer
+    and label it **[Generative]**. Otherwise, return the labeled extractive fallback.
+    """
     lang = normalize_lang(lang)
+
+    # No retrieved context? Stay safe.
+    if not hits:
+        return _extractive_answer(user_query, lang, hits)
+
     if mode != "generative" or _GENERATOR is None:
         return _extractive_answer(user_query, lang, hits)
+
     prompt = _build_grounded_prompt(user_query, lang, hits)
     try:
-        text = _GENERATOR.generate(prompt, max_new_tokens=int(max_new_tokens), temperature=float(temperature))
-        return text.strip() or _extractive_answer(user_query, lang, hits)
+        text = _GENERATOR.generate(
+            prompt,
+            max_new_tokens=int(max_new_tokens),
+            temperature=float(temperature),
+        ).strip()
+
+        if not text:
+            return _extractive_answer(user_query, lang, hits)
+
+        return "**[Generative]**\n\n" + text
+
     except Exception:
+        # Any runtime issue falls back to safe mode
         return _extractive_answer(user_query, lang, hits)