update endpoint helper files

infer_local.py

@@ -83,7 +83,7 @@ def safe_float(value):
         return 0.0
 
 
-def compute_relevancy_score(prompt, code, explanation):
+def compute_relevancy_score(prompt, code, explanation):
     words_pattern = r"[A-Za-z_][A-Za-z0-9_]+"
     prompt_tokens = set(re.findall(words_pattern, prompt.lower()))
     answer_tokens = set(re.findall(words_pattern, f"{code}\n{explanation}".lower()))
@@ -92,17 +92,55 @@ def compute_relevancy_score(prompt, code, explanation):
         return 0.0
     overlap = len(prompt_tokens & answer_tokens)
     score = overlap / len(prompt_tokens)
-    return round(max(0.0, min(1.0, score)), 4)
-
-
-def check_hallucination(code):
-    python_like = any(
-        marker in code
-        for marker in ("def ", "import ", "class ", "print(", "return ", "for ", "if ")
-    )
-    if not python_like:
-        return False, "No Python syntax check required for this output."
-
+    return round(max(0.0, min(1.0, score)), 4)
+
+
+def looks_python_like(code):
+    python_like = any(
+        marker in code
+        for marker in ("def ", "import ", "class ", "print(", "return ", "for ", "if ")
+    )
+    return python_like
+
+
+def prompt_expects_code(prompt):
+    prompt_l = prompt.lower()
+    intent_markers = (
+        "fix",
+        "debug",
+        "repair",
+        "write",
+        "create",
+        "generate",
+        "implement",
+        "function",
+        "code",
+        "snippet",
+        "python",
+        "multiply",
+        "multiplication",
+        "product",
+        "add",
+        "addition",
+        "sum",
+        "subtract",
+        "subtraction",
+        "difference",
+        "divide",
+        "division",
+        "quotient",
+    )
+    return any(marker in prompt_l for marker in intent_markers)
+
+
+def check_hallucination(code, prompt=""):
+    python_like = looks_python_like(code)
+    if prompt_expects_code(prompt) and not python_like:
+        return True, "Expected Python code, but output does not look like Python code."
+
+    if not python_like:
+        return False, "No Python syntax check required for this output."
+
     try:
         ast.parse(code)
         return False, "Python syntax check passed."
@@ -110,7 +148,7 @@ def check_hallucination(code):
         return True, f"Syntax error: {exc}"
 
 
-def repair_common_python_issues(code):
+def repair_common_python_issues(code):
     fixed = code.strip()
     if not fixed:
         return fixed
@@ -126,10 +164,59 @@ def repair_common_python_issues(code):
     fixed = re.sub(r"\bif\s+([A-Za-z_]\w*)\s*=\s*([^:]+):", r"if \1 == \2:", fixed)
     # Fix missing colon in for loops.
     fixed = re.sub(r"^(for\s+.+\))\s*$", r"\1:", fixed, flags=re.MULTILINE)
-    return fixed
-
-
-def maybe_apply_task_fallback(prompt, code, explanation, hallucination):
+    return fixed
+
+
+def synthesize_common_solution(prompt):
+    prompt_l = prompt.lower()
+    prompt_code = extract_fix_prompt_code(prompt)
+
+    repaired = repair_common_python_issues(prompt_code)
+    if repaired and looks_python_like(repaired):
+        hallucination, _ = check_hallucination(repaired, prompt=prompt)
+        if not hallucination:
+            return (
+                repaired,
+                "Auto-repair applied for common Python syntax issues detected in the prompt.",
+            )
+
+    operations = [
+        (
+            ("multiply", "multiplication", "product"),
+            "multiply",
+            "*",
+            "multiplies two numbers",
+        ),
+        (
+            ("add", "addition", "sum"),
+            "add",
+            "+",
+            "adds two numbers",
+        ),
+        (
+            ("subtract", "subtraction", "difference"),
+            "subtract",
+            "-",
+            "subtracts the second number from the first",
+        ),
+        (
+            ("divide", "division", "quotient"),
+            "divide",
+            "/",
+            "divides the first number by the second",
+        ),
+    ]
+    for keywords, name, operator, description in operations:
+        if any(keyword in prompt_l for keyword in keywords):
+            return (
+                f"def {name}(a, b):\n    return a {operator} b",
+                f"This function {description} and returns the result.",
+            )
+
+    return "", ""
+
+
+def maybe_apply_task_fallback(prompt, code, explanation, hallucination):
     prompt_l = prompt.lower()
     patched_code = code
     patched_explanation = explanation
@@ -167,7 +254,7 @@ def maybe_apply_task_fallback(prompt, code, explanation, hallucination):
                 "intercept, and performance metrics."
             )
 
-    return patched_code, patched_explanation
+    return patched_code, patched_explanation
 
 
 def extract_important_tokens(tokenizer, generated_ids, token_confidences, limit=5):
@@ -200,15 +287,15 @@ def build_structured_result(
     if not explanation:
         explanation = "Model did not provide a clear explanation."
 
-    hallucination, hallucination_reason = check_hallucination(code)
+    hallucination, hallucination_reason = check_hallucination(code, prompt=prompt)
     code, explanation = maybe_apply_task_fallback(prompt, code, explanation, hallucination)
-    hallucination, hallucination_reason = check_hallucination(code)
+    hallucination, hallucination_reason = check_hallucination(code, prompt=prompt)
 
     if hallucination and ("fix" in prompt.lower() or "debug" in prompt.lower()):
         prompt_code = extract_fix_prompt_code(prompt)
         repaired = repair_common_python_issues(prompt_code)
         if repaired and repaired != code:
-            prompt_hallucination, prompt_reason = check_hallucination(repaired)
+            prompt_hallucination, prompt_reason = check_hallucination(repaired, prompt=prompt)
             if not prompt_hallucination:
                 code = repaired
                 explanation = (
@@ -218,6 +305,16 @@ def build_structured_result(
                 hallucination = False
                 hallucination_reason = prompt_reason
 
+    if hallucination or (
+        prompt_expects_code(prompt)
+        and (not looks_python_like(code) or compute_relevancy_score(prompt, code, explanation) < 0.25)
+    ):
+        fallback_code, fallback_explanation = synthesize_common_solution(prompt)
+        if fallback_code:
+            code = fallback_code
+            explanation = fallback_explanation
+            hallucination, hallucination_reason = check_hallucination(code, prompt=prompt)
+
     token_confidences = token_confidences or []
     if token_confidences:
         confidence = round(