Yev/text to sql v1.1

resources_servers/text_to_sql/README.md

@@ -41,7 +41,7 @@ Each data sample should include:
       },
       {
         "role": "user",
-        "content": "DIALECT: postgresql\n\nDATABASE CONTEXT:\nCREATE TABLE users (id SERIAL PRIMARY KEY, name VARCHAR(100));\nINSERT INTO users VALUES (1, 'Alice'), (2, 'Bob');\n\nQUESTION:\nList all user names ordered alphabetically"
+        "content": "<DIALECT>postgresql</DIALECT>\n\n<DATABASE_CONTEXT>\nCREATE TABLE users (id SERIAL PRIMARY KEY, name VARCHAR(100));\nINSERT INTO users VALUES (1, 'Alice'), (2, 'Bob');\n</DATABASE_CONTEXT>\n\n<QUESTION>\nList all user names ordered alphabetically\n</QUESTION>"
       }
     ]
   },

resources_servers/text_to_sql/app.py

@@ -97,25 +97,6 @@ def extract_sql_from_response(text: str) -> Optional[str]:
     return None
 
 
-def _extract_question_text(params: NeMoGymResponseCreateParamsNonStreaming) -> str:
-    """Extract the question text from the last user message."""
-    last_text: Optional[str] = None
-    for m in params.input or []:
-        if getattr(m, "role", None) == "user":
-            c = getattr(m, "content", None)
-            if isinstance(c, str):
-                last_text = c
-    return (last_text or "").strip()
-
-
-def _extract_dialect_from_prompt(text: str) -> Optional[str]:
-    """Extract SQL dialect from a structured prompt."""
-    if not text:
-        return None
-    match = re.search(r"^\s*DIALECT:\s*([a-zA-Z0-9_+-]+)\s*$", text, re.MULTILINE)
-    return match.group(1) if match else None
-
-
 def _normalize_dialect(dialect: Optional[str]) -> Optional[str]:
     if not dialect:
         return None
@@ -171,7 +152,7 @@ class TextToSqlRunRequest(BaseRunRequest):
     sql: str  # Ground truth SQL query (required)
     sql_dialect: str  # SQL dialect: mysql, postgresql, sqlite (required)
     sql_context: str = ""  # Database schema (CREATE/INSERT statements)
-    sql_prompt: Optional[str] = None  # Natural language question (optional, extracted from input if not provided)
+    sql_prompt: str  # Natural language question (required)
     metadata: Optional[dict[str, Any]] = None
 
 
@@ -196,7 +177,7 @@ class TextToSqlVerifyResponse(BaseVerifyResponse):
     extracted_sql: Optional[str] = None
     sql_dialect: str  # SQL dialect used
     sql_context: str  # Database schema provided
-    sql_prompt: Optional[str] = None  # May be extracted from input
+    sql_prompt: str  # Natural language question
     judge_passed: bool = False
     failure_reason: Optional[FailureCode] = None
     judge_evaluations: list[JudgeEvaluation] = []
@@ -235,8 +216,8 @@ async def verify(self, body: TextToSqlVerifyRequest) -> TextToSqlVerifyResponse:
         if sql_dialect not in SUPPORTED_DIALECTS:
             raise ValueError(f"Unsupported SQL dialect '{sql_dialect}'. Supported: {sorted(SUPPORTED_DIALECTS)}")
 
-        # Extract question from request field or from user message
-        sql_prompt = body.sql_prompt or _extract_question_text(body.responses_create_params)
+        # sql_prompt is a required field, validated by Pydantic
+        sql_prompt = body.sql_prompt
 
         # Get model output text directly from response
         generated = body.response.output_text or ""