132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471 | @attrs
class SubqueryContextManager:
node: exp.Select = attrib()
prev_subquery_has_ingredient: bool = attrib()
tables_in_ingredients: set = attrib()
# Keep a running log of what aliases we've initialized so far, per subquery
alias_to_subquery: dict = attrib(default=None)
alias_to_tablename: dict = attrib(init=False)
tablename_to_alias: dict = attrib(init=False)
root: sqlglot.optimizer.scope.Scope = attrib(init=False)
def __attrs_post_init__(self):
self.alias_to_tablename = {}
self.tablename_to_alias = {}
# https://github.com/tobymao/sqlglot/blob/v20.9.0/posts/ast_primer.md#scope
self.root = build_scope(self.node)
def _reset_root(self):
self.root = build_scope(self.node)
def set_node(self, node):
self.node = node
self._reset_root()
def abstracted_table_selects(self) -> Generator[Tuple[str, bool, str], None, None]:
"""For each table in a given query, generates a `SELECT *` query where all unneeded predicates
are set to `TRUE`.
We say `unneeded` in the sense that to minimize the data that gets passed to an ingredient,
we don't need to factor in this operation at the moment.
Args:
node: exp.Select node from which to construct abstracted versions of queries for each table.
Returns:
abstracted_queries: Generator with (tablename, postprocess_columns, abstracted_query_str).
postprocess_columns tells us if we potentially executed a query with a `JOIN`, and need to apply some extra post-processing.
Examples:
```python
scm = SubqueryContextManager(
node=_parse_one(
"SELECT * FROM transactions WHERE {{Model('is this an italian restaurant?', 'transactions::merchant')}} = TRUE AND child_category = 'Restaurants & Dining'"
)
)
scm.abstracted_table_selects()
```
Returns:
```text
('transactions', False, 'SELECT * FROM transactions WHERE TRUE AND child_category = \'Restaurants & Dining\'')
```
"""
# TODO: don't really know how to optimize with 'CASE' queries right now
if self.node.find(exp.Case):
return
# Special condition: If...
# 1) We *only* have an ingredient in the top-level `SELECT` clause
# ... then we should execute entire rest of SQL first and assign to temporary session table.
# Example: """SELECT w.title, w."designer ( s )", {{LLMMap('How many animals are in this image?', 'images::title')}}
# FROM images JOIN w ON w.title = images.title
# WHERE "designer ( s )" = 'georgia gerber'"""
# Below, we need `self.node.find(exp.Table)` in case we get a QAIngredient on its own
# E.g. `SELECT A() AS _col_0` should be ignored
if (
self.node.find(exp.Table)
and check.ingredients_only_in_top_select(self.node)
and not check.ingredient_alias_in_query_body(self.node)
):
abstracted_query = to_select_star(self.node).transform(
transform.set_structs_to_true
)
abstracted_query_str = recover_blendsql(
abstracted_query.sql(dialect=FTS5SQLite)
)
for tablename in self.tables_in_ingredients:
yield (tablename, True, abstracted_query_str)
return
for tablename, table_star_query in self._table_star_queries():
# If this table_star_query doesn't have an ingredient at the top-level, we can safely ignore
if (
len(
list(
get_scope_nodes(
root=self.root, nodetype=exp.Struct, restrict_scope=True
)
)
)
== 0
):
continue
# If our previous subquery has an ingredient, we can't optimize with subquery condition
# So, remove this subquery constraint and run
if self.prev_subquery_has_ingredient:
table_star_query = table_star_query.transform(
transform.maybe_set_subqueries_to_true
)
# Substitute all ingredients with 'TRUE'
abstracted_query = table_star_query.transform(transform.set_structs_to_true)
# Check here to see if we have no other predicates other than 'WHERE TRUE'
# There's no point in creating a temporary table in this situation
where_node = abstracted_query.find(exp.Where)
if where_node:
if where_node.args["this"] == exp.true():
continue
elif isinstance(where_node.args["this"], exp.Column):
continue
elif check.all_terminals_are_true(where_node):
continue
elif not where_node:
continue
abstracted_query_str = recover_blendsql(
abstracted_query.sql(dialect=FTS5SQLite)
)
yield (tablename, False, abstracted_query_str)
def _table_star_queries(
self,
) -> Generator[Tuple[str, exp.Select], None, None]:
"""For each table in the select query, generates a new query
selecting all columns with the given predicates (Relationships like x = y, x > 1, x >= y).
Args:
node: The exp.Select node containing the query to extract table_star queries for
Returns:
table_star_queries: Generator with (tablename, exp.Select). The exp.Select is the table_star query
Examples:
```sql
SELECT "Run Date", Account, Action, ROUND("Amount ($)", 2) AS 'Total Dividend Payout ($$)', Name
FROM account_history
LEFT JOIN constituents ON account_history.Symbol = constituents.Symbol
WHERE constituents.Sector = 'Information Technology'
AND lower(Action) like "%dividend%"
```
Returns (after getting str representation of `exp.Select`):
```text
('account_history', 'SELECT * FROM account_history WHERE lower(Action) like "%dividend%')
('constituents', 'SELECT * FROM constituents WHERE sector = \'Information Technology\'')
```
"""
# Use `scope` to get all unique tablenodes in ast
tablenodes = set(
list(
get_scope_nodes(nodetype=exp.Table, root=self.root, restrict_scope=True)
)
)
# aliasnodes catch instances where we do something like
# `SELECT (SELECT * FROM x) AS w`
curr_alias_to_tablename = {}
curr_alias_to_subquery = {}
subquery_node = self.node.find(exp.Subquery)
if subquery_node is not None:
# Make a note here: we need to create a new table with the name of the alias,
# and set to results of this subquery
alias = None
if "alias" in subquery_node.args:
alias = subquery_node.args["alias"]
if alias is None:
# Try to get from parent
parent_node = subquery_node.parent
if parent_node is not None:
if "alias" in parent_node.args:
alias = parent_node.args["alias"]
if alias is not None:
if not any(x.name == alias.name for x in tablenodes):
tablenodes.add(exp.Table(this=exp.Identifier(this=alias.name)))
curr_alias_to_subquery = {alias.name: subquery_node.args["this"]}
for tablenode in tablenodes:
# Check to be sure this is in the top-level `SELECT`
if check.in_subquery(tablenode):
continue
# Check to see if we have a table alias
# e.g. `SELECT a FROM table AS w`
table_alias_node = tablenode.find(exp.TableAlias)
if table_alias_node is not None:
tablename_to_extract = table_alias_node.name
curr_alias_to_tablename = {tablename_to_extract: tablenode.name}
base_select_str = f'SELECT * FROM "{tablenode.name}" AS "{tablename_to_extract}" WHERE '
else:
tablename_to_extract = tablenode.name
base_select_str = f'SELECT * FROM "{tablenode.name}" WHERE '
table_conditions_str = self.get_table_predicates_str(
tablename=tablename_to_extract,
disambiguate_multi_tables=bool(len(tablenodes) > 1)
or (table_alias_node is not None),
)
self.alias_to_tablename = self.alias_to_tablename | curr_alias_to_tablename
self.tablename_to_alias = self.tablename_to_alias | {
v: k for k, v in curr_alias_to_tablename.items()
}
self.alias_to_subquery = self.alias_to_subquery | curr_alias_to_subquery
if table_conditions_str:
yield (
tablenode.name,
_parse_one(base_select_str + table_conditions_str),
)
def get_table_predicates_str(
self, tablename, disambiguate_multi_tables: bool
) -> str:
"""Returns str containing all predicates acting on a specific tablename.
Args:
tablename: The target tablename to search and extract predicates for
disambiguate_multi_tables: `True` if we have multiple tables in our subquery,
and need to be sure we're only fetching the predicates for the specified `tablename`
"""
# 2 places conditions can come in here
# 'WHERE' statement and predicate in a 'JOIN' statement
all_table_predicates = []
for table_predicates in get_scope_nodes(
nodetype=exp.Predicate, root=self.root, restrict_scope=True
):
if check.in_subquery(table_predicates):
continue
if disambiguate_multi_tables:
table_predicates = table_predicates.transform(
transform.extract_multi_table_predicates, tablename=tablename
)
if isinstance(table_predicates, exp.Expression):
all_table_predicates.append(table_predicates)
if len(all_table_predicates) == 0:
return ""
table_conditions_str = " AND ".join(
[c.sql(dialect=FTS5SQLite) for c in all_table_predicates]
)
return table_conditions_str
def infer_gen_constraints(self, start: int, end: int) -> dict:
"""Given syntax of BlendSQL query, infers a regex pattern (if possible) to guide
downstream Model generations.
For example:
```sql
SELECT * FROM w WHERE {{LLMMap('Is this true?', 'w::colname')}}
```
We can infer given the structure above that we expect `LLMMap` to return a boolean.
This function identifies that.
Arguments:
indices: The string indices pointing to the span within the overall BlendSQL query
containing our ingredient in question.
Returns:
dict, with keys:
- output_type
- 'boolean' | 'integer' | 'float' | 'string'
- regex: regular expression pattern lambda to use in constrained decoding with Model
- See `create_regex` for more info on these regex lambdas
- options: Optional str default to pass to `options` argument in a QAIngredient
- Will have the form '{table}::{column}'
"""
def create_regex(
output_type: Literal["boolean", "integer", "float"]
) -> Callable[[int], str]:
"""Helper function to create a regex lambda.
These regex lambdas take an integer (num_repeats) and return
a regex regex which is restricted to repeat exclusively num_repeats times.
"""
if output_type == "boolean":
base_regex = f"(t|f|{DEFAULT_NAN_ANS})"
elif output_type == "integer":
# SQLite max is 18446744073709551615
# This is 20 digits long, so to be safe, cap the generation at 19
base_regex = r"(\d{1,18}" + f"|{DEFAULT_NAN_ANS})"
elif output_type == "float":
base_regex = r"(\d(\d|\.)*" + f"|{DEFAULT_NAN_ANS})"
else:
raise ValueError(f"Unknown output_type {output_type}")
return base_regex
added_kwargs: Dict[str, Any] = {}
ingredient_node = _parse_one(self.sql()[start:end])
child = None
for child, _, _ in self.node.walk():
if child == ingredient_node:
break
if child is None:
raise ValueError
ingredient_node_in_context = child
start_node = ingredient_node_in_context.parent
# Below handles when we're in a function
# Example: CAST({{LLMMap('jump distance', 'w::notes')}} AS FLOAT)
while isinstance(start_node, exp.Func) and start_node is not None:
start_node = start_node.parent
output_type: Literal["boolean", "integer", "float"] = None
predicate_literals: List[str] = []
if start_node is not None:
predicate_literals = get_predicate_literals(start_node)
# Check for instances like `{column} = {QAIngredient}`
# where we can infer the space of possible options for QAIngredient
if isinstance(start_node, exp.EQ):
if isinstance(start_node.args["this"], exp.Column):
if "table" not in start_node.args["this"].args:
logger.debug(
"When inferring `options` in infer_gen_kwargs, encountered a column node with "
"no table specified!\nShould probably mark `schema_qualify` arg as True"
)
else:
# This is valid for a default `options` set
added_kwargs[
"options"
] = f"{start_node.args['this'].args['table'].name}::{start_node.args['this'].args['this'].name}"
if len(predicate_literals) > 0:
if all(isinstance(x, bool) for x in predicate_literals):
output_type = "boolean"
elif all(isinstance(x, float) for x in predicate_literals):
output_type = "float"
elif all(isinstance(x, int) for x in predicate_literals):
output_type = "integer"
else:
predicate_literals = [str(i) for i in predicate_literals]
added_kwargs["output_type"] = "string"
if len(predicate_literals) > 1:
added_kwargs["example_outputs"] = DEFAULT_ANS_SEP.join(
predicate_literals
)
else:
added_kwargs[
"example_outputs"
] = f"{predicate_literals[0]}{DEFAULT_ANS_SEP}{DEFAULT_NAN_ANS}"
return added_kwargs
elif isinstance(
ingredient_node_in_context.parent, (exp.Order, exp.Ordered, exp.AggFunc)
):
output_type = "float" # Use 'float' as default numeric regex, since it's more expressive than 'integer'
if output_type is not None:
added_kwargs["output_type"] = output_type
added_kwargs[IngredientKwarg.REGEX] = create_regex(output_type)
return added_kwargs
def sql(self, dialect: sqlglot.dialects.Dialect = FTS5SQLite):
return recover_blendsql(self.node.sql(dialect=dialect))
|