Text_to_sql / engine.py
bhavika24's picture
Upload 5 files
d036c34 verified
import json
import os
from functools import lru_cache
from openai import OpenAI
from datetime import datetime, date, timedelta
import re
# =========================
# CONFIG
# =========================
client = OpenAI(api_key=os.getenv("OPENAI_API_KEY"))
# =========================
# METADATA LOADING
# =========================
@lru_cache(maxsize=1)
def load_metadata():
with open("modules.json") as f:
modules = json.load(f)
with open("join_graph.json") as f:
joins = json.load(f)
with open("field_types.json") as f:
field_types = json.load(f)
with open("fields.json") as f:
fields = json.load(f)
return {
"modules": modules,
"joins": joins,
"field_types": field_types,
"fields": fields
}
# =========================
# OPERATOR RESOLUTION (COMPLETE FIXED VERSION)
# =========================
def resolve_operator(op, value, field_type=None):
"""
Resolve operator and format value based on data type
FIXED: Properly handles numeric types without quotes
"""
# Normalize operator input
op = op.lower().strip().replace(" ", "_")
# Extended operator aliases for all your operators
OPERATOR_ALIASES = {
"=": "equals",
"==": "equals",
"eq": "equals",
"!=": "not_equals",
"<>": "not_equals",
">": "greater_than",
"<": "less_than",
">=": "greater_or_equal",
"<=": "less_or_equal",
"greater than": "greater_than",
"less than": "less_than",
"greaterthan": "greater_than",
"lessthan": "less_than",
"greaterthanorequal": "greater_or_equal",
"lessthanorequal": "less_or_equal",
"does_not_contain": "not_contains",
"is_blank": "is_empty",
"is_not_blank": "is_not_empty",
"on": "equals",
"date_equals": "equals",
"date_between": "between",
"startswith": "starts_with",
"endswith": "ends_with"
}
op = OPERATOR_ALIASES.get(op, op)
# SQL operator mapping
mapping = {
"equals": "=",
"not_equals": "!=",
"greater_than": ">",
"less_than": "<",
"greater_or_equal": ">=",
"less_or_equal": "<=",
"contains": "LIKE",
"not_contains": "NOT LIKE",
"starts_with": "LIKE",
"ends_with": "LIKE",
"in": "IN",
"not_in": "NOT IN",
"is_empty": "IS NULL",
"is_not_empty": "IS NOT NULL",
"between": "BETWEEN",
"not_between": "NOT BETWEEN",
"before": "<",
"after": ">",
# Date relative operators
"today": "=",
"yesterday": "=",
"tomorrow": "=",
"this_week": "BETWEEN",
"last_week": "BETWEEN",
"next_week": "BETWEEN",
"this_month": "BETWEEN",
"last_month": "BETWEEN",
"next_month": "BETWEEN",
"this_quarter": "BETWEEN",
"last_quarter": "BETWEEN",
"next_quarter": "BETWEEN",
"this_year": "BETWEEN",
"last_year": "BETWEEN"
}
if op not in mapping:
raise ValueError(f"Unsupported operator: {op}")
sql_op = mapping[op]
# βœ… Determine if field is numeric
is_numeric = field_type in ['integer', 'decimal', 'float', 'number', 'int', 'bigint']
is_date = field_type in ['date', 'datetime', 'timestamp']
is_boolean = field_type in ['boolean', 'bool']
# Escape string values safely
def sql_escape(val):
if val is None:
return 'NULL'
return str(val).replace("'", "''")
# Handle NULL operators
if op in ("is_empty", "is_not_empty"):
return sql_op, ""
# Handle date relative operators
if op in ("today", "yesterday", "tomorrow", "this_week", "last_week", "next_week",
"this_month", "last_month", "next_month", "this_quarter", "last_quarter",
"next_quarter", "this_year", "last_year"):
today = date.today()
if op == "today":
return "=", f"'{today}'"
elif op == "yesterday":
return "=", f"'{today - timedelta(days=1)}'"
elif op == "tomorrow":
return "=", f"'{today + timedelta(days=1)}'"
elif op == "this_week":
start = today - timedelta(days=today.weekday())
end = start + timedelta(days=6)
return "BETWEEN", f"'{start}' AND '{end}'"
elif op == "this_month":
start = today.replace(day=1)
if today.month == 12:
end = today.replace(day=31)
else:
end = (today.replace(month=today.month+1, day=1) - timedelta(days=1))
return "BETWEEN", f"'{start}' AND '{end}'"
elif op == "this_year":
start = today.replace(month=1, day=1)
end = today.replace(month=12, day=31)
return "BETWEEN", f"'{start}' AND '{end}'"
# Add more as needed
# Handle LIKE operators
if op == "contains":
return sql_op, f"'%{sql_escape(value)}%'"
if op == "not_contains":
return sql_op, f"'%{sql_escape(value)}%'"
if op == "starts_with":
return sql_op, f"'{sql_escape(value)}%'"
if op == "ends_with":
return sql_op, f"'%{sql_escape(value)}'"
# Handle BETWEEN operator
if op in ("between", "not_between"):
if not isinstance(value, (list, tuple)) or len(value) != 2:
raise ValueError("BETWEEN operator requires array of 2 values")
if is_numeric:
return sql_op, f"{value[0]} AND {value[1]}"
else:
return sql_op, f"'{sql_escape(value[0])}' AND '{sql_escape(value[1])}'"
# βœ… Handle IN operators with type checking
if op in ("in", "not_in"):
if not isinstance(value, list):
value = [value]
if is_numeric:
escaped = [str(v) for v in value] # βœ… No quotes for numbers
else:
escaped = [f"'{sql_escape(v)}'" for v in value]
return sql_op, f"({', '.join(escaped)})"
# βœ… Handle regular comparison operators with type awareness
if is_numeric:
return sql_op, str(value) # βœ… No quotes for numbers
elif is_boolean:
if isinstance(value, bool):
return sql_op, "1" if value else "0"
return sql_op, str(value)
elif is_date:
return sql_op, f"'{sql_escape(value)}'"
else:
return sql_op, f"'{sql_escape(value)}'"
# =========================
# JOIN RESOLUTION (FIXED)
# =========================
def resolve_join_path(start_table, end_table):
"""
Find join path between two tables
FIXED: Handles your join_graph.json structure
"""
joins = load_metadata()["joins"]
# Try direct lookup with double underscore
key = f"{start_table}__{end_table}"
if key in joins:
return joins[key]
# Try searching by start and end table
for path_key, path in joins.items():
if path["start_table"] == start_table and path["end_table"] == end_table:
return path
raise ValueError(
f"No join path found from {start_table} to {end_table}"
)
def build_join_sql(base_table, join_path):
"""
Build JOIN SQL from join path
FIXED: Properly handles multi-step joins with from_previous_step flag
"""
steps = join_path["steps"]
sql = []
# Sort steps by step number
sorted_steps = sorted(steps, key=lambda x: x.get("step", 0))
for i, step in enumerate(sorted_steps):
alias = step["alias"]
table = step["table"]
join_type = step["join_type"].upper()
# βœ… Determine the left side of the join
if i == 0:
# First join always references base table
left_ref = base_table
else:
# Subsequent joins: check from_previous_step flag
if step.get("from_previous_step", False):
left_ref = sorted_steps[i-1]["alias"] # βœ… Use previous alias
else:
left_ref = base_table
# Build basic join condition
join_condition = f"{left_ref}.{step['base_column']} = {alias}.{step['foreign_column']}"
# βœ… Add extra conditions if present
if "extra_conditions" in step and step["extra_conditions"]:
for extra in step["extra_conditions"]:
condition = f"{alias}.{extra['column']} {extra['operator']} {extra['value']}"
join_condition += f" AND {condition}"
sql.append(
f"{join_type} JOIN {table} {alias} ON {join_condition}"
)
return "\n".join(sql)
# =========================
# FIELD RESOLUTION
# =========================
FIELD_ALIASES = {
"join_date": "date_of_joining",
"joining_date": "date_of_joining",
"joined": "date_of_joining",
"hire_date": "date_of_joining",
"emp_code": "employee_code",
"emp_name": "full_name",
"dept": "department"
}
def resolve_field(field_name, module):
meta = load_metadata()
fields = meta["fields"]
# πŸ”Ή Normalize field name
field_name = field_name.lower().strip().replace(" ", "_")
field_name = FIELD_ALIASES.get(field_name, field_name)
# πŸ”Ή Validate existence
if field_name not in fields:
raise ValueError(f"Unknown field: {field_name}")
field = fields[field_name]
# πŸ”Ή Validate module
if field["module"] != module:
raise ValueError(
f"Field '{field_name}' does not belong to module '{module}'"
)
# πŸ”Ή Validate mapping
if "table" not in field or "column" not in field:
raise ValueError(
f"Field '{field_name}' is missing table/column mapping"
)
return field
# =========================
# JSON SAFETY
# =========================
def safe_json_loads(text):
try:
return json.loads(text)
except json.JSONDecodeError:
# Try to extract JSON from markdown
match = re.search(r'```(?:json)?\s*(\{.*?\})\s*```', text, re.DOTALL)
if match:
return json.loads(match.group(1))
match = re.search(r"\{.*\}", text, re.DOTALL)
if match:
return json.loads(match.group())
raise ValueError("LLM returned invalid JSON")
# =========================
# INTENT PARSING (LLM)
# =========================
def parse_intent(question, retries=2):
meta = load_metadata()
# βœ… Build schema safely
schema_description = "\n".join([
f"{module}: {', '.join(fields)}"
for module in meta["modules"]
if (fields := [
f for f in meta["fields"]
if meta["fields"][f]["module"] == module
][:20]) # Limit to 20 fields per module for token efficiency
])
prompt = f"""
You are a Text-to-SQL engine.
Your task is to generate a SINGLE valid SQL query based ONLY on the metadata provided.
CRITICAL RULES (follow strictly):
1. Use ONLY the tables and columns explicitly listed in the metadata.
2. If the user asks for a field, table, or concept NOT present in the metadata, IGNORE that part.
3. Do NOT invent table names, column names, joins, or filters.
4. Do NOT explain the query.
5. Do NOT return anything except the SQL query.
6. If no valid SQL can be generated using the metadata, return a SQL query that explains the reason in a single text column named reason
Database Metadata:
{{METADATA_JSON}}
User Question:
{{USER_QUERY}}
Output:
- Return a single SQL query in {{SQL_DIALECT}} syntax.
- No markdown.
- No comments.
- No extra text.
"""
for attempt in range(retries):
try:
res = client.chat.completions.create(
model="gpt-4o-mini",
messages=[
{
"role": "system",
"content": "Return ONLY valid minified JSON. No text. No explanation."
},
{"role": "user", "content": prompt}
],
temperature=0
)
content = res.choices[0].message.content.strip()
plan = safe_json_loads(content)
# βœ… NORMALIZE + STABILIZE INTENT SHAPE
if "module" in plan:
plan["module"] = plan["module"].lower().strip()
plan.setdefault("filters", [])
plan.setdefault("select", [])
return plan
except Exception as e:
if attempt == retries - 1:
raise ValueError(f"LLM failed to return valid JSON: {str(e)}")
# =========================
# SQL GENERATOR (FIXED)
# =========================
def build_sql(plan):
meta = load_metadata()
# πŸ”΄ Defensive: normalize module
module = plan["module"].lower().strip()
if module not in meta["modules"]:
raise ValueError(f"Unknown module: {module}")
base_table = meta["modules"][module]["base_table"]
joins = []
joined_tables = {base_table} # βœ… Track all joined tables
where_clauses = []
# ---------- SELECT ----------
select_fields = plan.get("select", [])
if select_fields:
select_columns = []
for f in select_fields:
field = resolve_field(f, module)
select_columns.append(
f"{field['table']}.{field['column']} AS {f}"
)
select_sql = ", ".join(select_columns)
else:
select_sql = f"{base_table}.*"
# ---------- FILTERS ----------
for f in plan.get("filters", []):
field = resolve_field(f["field"], module)
table = field["table"]
column = field["column"]
field_type = field.get("type") # βœ… Get field type
# Add join if needed
if table != base_table and table not in joined_tables:
join_path = resolve_join_path(base_table, table)
joins.append(build_join_sql(base_table, join_path))
# βœ… Track all tables in join path
for step in join_path["steps"]:
joined_tables.add(step["table"])
# βœ… Pass field_type to resolve_operator
sql_op, sql_value = resolve_operator(f["operator"], f["value"], field_type)
if sql_value: # Has value
where_clauses.append(f"{table}.{column} {sql_op} {sql_value}")
else: # IS NULL / IS NOT NULL
where_clauses.append(f"{table}.{column} {sql_op}")
# πŸ”΄ FIX: safe WHERE clause
where_sql = f"WHERE {' AND '.join(where_clauses)}" if where_clauses else ""
# ---------- FINAL SQL ----------
sql_parts = [
f"SELECT {select_sql}",
f"FROM {base_table}"
]
if joins:
sql_parts.extend(joins)
if where_sql:
sql_parts.append(where_sql)
sql_parts.append("LIMIT 100")
sql = "\n".join(sql_parts)
return sql.strip()
# =========================
# VALIDATION
# =========================
def validate_sql(sql):
sql_lower = sql.lower()
if not sql_lower.strip().startswith("select"):
raise ValueError("Only SELECT allowed")
forbidden = ["drop", "delete", "update", "insert", "truncate", "alter", "create"]
for keyword in forbidden:
if re.search(rf'\b{keyword}\b', sql_lower):
raise ValueError(f"Unsafe SQL: '{keyword}' not allowed")
return sql
# =========================
# MAIN ENTRY POINT
# =========================
def run(question):
plan = parse_intent(question)
# πŸ”΄ REQUIRED: validate minimum intent
if not isinstance(plan, dict):
raise ValueError("Invalid intent format")
if "module" not in plan:
raise ValueError("Unable to determine module from question")
# Optional but safe defaults
plan.setdefault("filters", [])
plan.setdefault("select", [])
sql = build_sql(plan)
sql = validate_sql(sql)
return {
"query_plan": plan,
"sql": sql
}
# =========================
# TEST
# =========================
if __name__ == "__main__":
test_queries = [
"Show all employees",
"Find departments with more than 50 employees",
"Show employees in departments 1, 2, 3",
"List employees who joined this month"
]
for q in test_queries:
print(f"\n{'='*80}")
print(f"Q: {q}")
print('='*80)
try:
result = run(q)
print("SQL:", result["sql"])
except Exception as e:
print("ERROR:", e)