Update app.py

app.py

@@ -78,6 +78,89 @@ MODEL_ID = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"
 tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
 pipe = pipeline("text-generation", model=MODEL_ID, tokenizer=tokenizer)
 
+def _fmt_num(x, sig=4):
+    """Safe, short formatting for numbers (returns string)."""
+    try:
+        if x is None:
+            return "N/A"
+        if isinstance(x, (int,)):
+            return str(x)
+        if isinstance(x, float):
+            # use general format with sig significant digits
+            return f"{x:.{sig}g}"
+        return str(x)
+    except Exception:
+        return str(x)
+
+def _llm_generate(prompt: str, max_tokens: int = 300) -> str:
+    """
+    Run the local pipeline, then strip any echoed prompt and common instruction text.
+    If the model echoes instructions, do one gentle retry with a simplified prompt.
+    """
+    # Primary generation: deterministic (no sampling) is usually safer for engineering text.
+    try:
+        out = pipe(
+            prompt,
+            max_new_tokens=max_tokens,
+            do_sample=False,
+            temperature=0.0,
+            return_full_text=True,
+        )
+    except Exception:
+        # fallback: try return_full_text=False if first attempt fails for this model
+        out = pipe(
+            prompt,
+            max_new_tokens=max_tokens,
+            do_sample=False,
+            temperature=0.0,
+            return_full_text=False,
+        )
+
+    # get text (handle both pipeline variants)
+    text = ""
+    if isinstance(out, list) and len(out) > 0:
+        text = out[0].get("generated_text", "") or out[0].get("text", "") or ""
+    text = text or ""
+    # If the model returned the prompt + output, strip the prompt if present
+    if text.startswith(prompt):
+        text = text[len(prompt):]
+    text = text.strip()
+
+    # If output looks like it merely repeated instructions, try a simpler short-prompt retry
+    low_quality_indicators = [
+        "Use bullet points", "Be sure to include", "Do not", "Do NOT",
+        "Now produce", "System:", "User:", "Instruction:"
+    ]
+    if (not text) or any(ind in text for ind in low_quality_indicators) or len(text) < 10:
+        # simple short retry prompt asking for only the final answer
+        simple_prompt = prompt + "\n\nNow produce ONLY the requested explanation below (no re-statement of the prompt or instructions):\n"
+        try:
+            out2 = pipe(
+                simple_prompt,
+                max_new_tokens=max_tokens,
+                do_sample=False,
+                temperature=0.0,
+                return_full_text=True,
+            )
+        except Exception:
+            out2 = pipe(
+                simple_prompt,
+                max_new_tokens=max_tokens,
+                do_sample=False,
+                temperature=0.0,
+                return_full_text=False,
+            )
+        text2 = out2[0].get("generated_text", "") or out2[0].get("text", "") or ""
+        if text2.startswith(simple_prompt):
+            text2 = text2[len(simple_prompt):]
+        text2 = text2.strip()
+        if text2 and len(text2) > 10 and not any(ind in text2 for ind in low_quality_indicators):
+            return text2
+        # final fallback:
+        return "[LLM failed to generate a usable explanation — try a different model or reduce the prompt size]"
+
+    return text
+
 def llm_explain(record: dict) -> str:
     if not record.get("ok", False):
         return "Errors: " + "; ".join(record.get("errors", []))