ChangedModelForQueryGeneration

Co-authored-by: Copilot <copilot@github.com>

sql_generator.py

@@ -96,7 +96,7 @@ def _get_sql_generator() -> Any:
     if _SQL_GENERATOR is None:
         from transformers import pipeline
 
-        model_id = os.getenv("SQL_MODEL", "google/flan-t5-base")
+        model_id = os.getenv("SQL_MODEL", "gaussalgo/T5-LM-Large-text2sql-spider")
         _SQL_GENERATOR = pipeline(
             task="text2text-generation",
             model=model_id,
@@ -109,14 +109,13 @@ def _get_sql_generator() -> Any:
 
 def _build_prompt(payload: SqlGenerationRequest) -> str:
     return (
-        "You translate user requests into SQLite SELECT queries. "
-        "Return only SQL without explanations. "
-        "Use only tables and columns from the schema. "
-        "Never generate INSERT, UPDATE, DELETE, DROP, ALTER, PRAGMA, ATTACH or CREATE. "
-        "Prefer explicit JOIN conditions using foreign keys from the schema. "
-        f"Add LIMIT {payload.limit} when the query is not an aggregate result.\n\n"
-        f"Schema:\n{DEFAULT_DB_SCHEMA}\n\n"
-        f"User request:\n{payload.question}\n\n"
+        "Task: generate one SQLite SQL query for the question. "
+        "Return only SQL, no markdown and no explanation. "
+        "Use only schema objects provided below. "
+        "Never generate non-SELECT statements.\n\n"
+        f"Question: {payload.question}\n"
+        f"Schema: {DEFAULT_DB_SCHEMA}\n"
+        f"Constraints: output must start with SELECT; apply LIMIT {payload.limit} when not aggregate.\n"
         "SQL:"
     )
 
@@ -171,6 +170,7 @@ def generate_sql(question: str, limit: int = 200) -> str:
         prompt,
         max_new_tokens=256,
         do_sample=False,
+        num_beams=4,
         truncation=True,
     )