Introduce lf.track_queries.

- Revise the input preparation logic of `lf.query`.
- Add `lf.QueryInvocation` to represent `lf.query` invocations.
- Add `lf.track_queries` context manager to track `lf.query` invocations.

PiperOrigin-RevId: 702220910

langfun/__init__.py

@@ -35,6 +35,8 @@
 score = structured.score
 generate_class = structured.generate_class
 
+track_queries = structured.track_queries
+
 # Helper functions for input/output transformations based on
 # `lf.query` (e.g. jax-on-beam could use these for batch processing)
 query_prompt = structured.query_prompt

langfun/core/structured/__init__.py

@@ -69,6 +69,8 @@
 from langfun.core.structured.prompting import query_prompt
 from langfun.core.structured.prompting import query_output
 from langfun.core.structured.prompting import query_reward
+from langfun.core.structured.prompting import QueryInvocation
+from langfun.core.structured.prompting import track_queries
 
 from langfun.core.structured.description import DescribeStructure
 from langfun.core.structured.description import describe

langfun/core/structured/prompting.py

@@ -13,8 +13,9 @@
 # limitations under the License.
 """Symbolic query."""
 
+import contextlib
 import functools
-from typing import Any, Callable, Type, Union
+from typing import Annotated, Any, Callable, Iterator, Type, Union
 
 import langfun.core as lf
 from langfun.core.llms import fake
@@ -102,7 +103,7 @@ def _query_structure_cls(
 
 
 def query(
-    prompt: Union[str, pg.Symbolic],
+    prompt: Union[str, lf.Template, Any],
     schema: Union[
         schema_lib.Schema, Type[Any], list[Type[Any]], dict[str, Any], None
     ] = None,
@@ -119,7 +120,7 @@ def query(
     skip_lm: bool = False,
     **kwargs,
 ) -> Any:
-  """Parse a natural langugage message based on schema.
+  """Queries an language model for a (maybe) structured output.
 
   Examples:
 
@@ -189,55 +190,85 @@ class Flight(pg.Object):
   """
     # Internal usage logging.
 
+  # Normalize query schema.
   # When `lf.query` is used for symbolic completion, schema is automatically
   # inferred when it is None.
   if isinstance(prompt, pg.Symbolic) and prompt.sym_partial and schema is None:
     schema = prompt.__class__
 
-  # Create a copy of the prompt if it has a parent object, so all child modality
-  # objects could be referred by path relative to the prompt.
-  if isinstance(prompt, lf.Template) and prompt.sym_parent:
-    prompt = prompt.clone()
+  # Normalize query input.
+  if isinstance(prompt, (lf.Message, str)):
+    # Query with structured output.
+    prompt_kwargs = kwargs.copy()
+    prompt_kwargs.pop('template_str', None)
+    query_input = lf.Template.from_value(prompt, **prompt_kwargs)
+  elif isinstance(prompt, lf.Template):
+    # Create a copy of the prompt if it has a parent object, so all child
+    # modality objects could be referred by path relative to the prompt.
+    query_input = prompt.clone() if prompt.sym_parent is not None else prompt
+
+    # Attach template metadata from kwargs. This is used to pass through fields
+    # from kwargs to the rendered message.
+    template_metadata = {k: v for k, v in kwargs if k.startswith('metadata_')}
+    query_input.rebind(
+        template_metadata, skip_notification=True, raise_on_no_change=False
+    )
+  elif pg.MISSING_VALUE == prompt:
+    query_input = lf.UserMessage('Unused prompt.')
+  else:
+    query_input = schema_lib.mark_missing(prompt)
 
   if schema in (None, str):
     # Query with natural language output.
-    output = lf.LangFunc.from_value(prompt, **kwargs)(
+    output_message = lf.LangFunc.from_value(query_input)(
         lm=lm, cache_seed=cache_seed, skip_lm=skip_lm
     )
     if response_postprocess:
-      processed_text = response_postprocess(output.text)
-      if processed_text != output.text:
-        output = lf.AIMessage(processed_text, source=output)
-    return output if returns_message else output.text
-
-  # Query with structured output.
-  prompt_kwargs = kwargs.copy()
-
-  # NOTE(daiyip): when `template_str` is passed in, it's intended to modify the
-  # QueryStructure template string. Therefore, we pop out the argument for
-  # prompt rendering.
-  prompt_kwargs.pop('template_str', None)
-
-  if isinstance(prompt, (str, lf.Message, lf.Template)):
-    prompt = lf.Template.from_value(prompt, **prompt_kwargs).render(lm=lm)
+      processed_text = response_postprocess(output_message.text)
+      if processed_text != output_message.text:
+        output_message = lf.AIMessage(processed_text, source=output_message)
   else:
-    prompt = schema_lib.mark_missing(prompt)
-
-  output = _query_structure_cls(protocol)(
-      input=prompt,
-      schema=schema,
-      default=default,
-      examples=examples,
-      response_postprocess=response_postprocess,
-      autofix=autofix if protocol == 'python' else 0,
-      **kwargs,
-  )(
-      lm=lm,
-      autofix_lm=autofix_lm or lm,
-      cache_seed=cache_seed,
-      skip_lm=skip_lm,
-  )
-  return output if returns_message else output.result
+    # Query with structured output.
+    output_message = _query_structure_cls(protocol)(
+        input=(
+            query_input.render(lm=lm)
+            if isinstance(query_input, lf.Template)
+            else query_input
+        ),
+        schema=schema,
+        default=default,
+        examples=examples,
+        response_postprocess=response_postprocess,
+        autofix=autofix if protocol == 'python' else 0,
+        **kwargs,
+    )(
+        lm=lm,
+        autofix_lm=autofix_lm or lm,
+        cache_seed=cache_seed,
+        skip_lm=skip_lm,
+    )
+
+  def _result(message: lf.Message):
+    return message.text if schema in (None, str) else message.result
+
+  # Track the query invocations.
+  if pg.MISSING_VALUE != prompt and not skip_lm:
+    trackers = lf.context_value('__query_trackers__', [])
+    if trackers:
+      invocation = QueryInvocation(
+          input=pg.Ref(query_input),
+          schema=(
+              schema_lib.Schema.from_value(schema)
+              if schema not in (None, str) else None
+          ),
+          output=pg.Ref(_result(output_message)),
+          lm=pg.Ref(lm),
+          examples=pg.Ref(examples) if examples else [],
+      )
+      for i, (tracker, include_child_scopes) in enumerate(trackers):
+        if i == 0 or include_child_scopes:
+          tracker.append(invocation)
+  return output_message if returns_message else _result(output_message)
 
 
 def query_prompt(
@@ -264,7 +295,7 @@ def query_output(
   kwargs.pop('prompt', None)
   kwargs.pop('lm', None)
   return query(
-      'Unused prompt', schema, lm=fake.StaticResponse(response), **kwargs
+      pg.MISSING_VALUE, schema, lm=fake.StaticResponse(response), **kwargs
   )
 
 
@@ -320,3 +351,65 @@ def _reward(self, input, expected_output, metadata):  # pylint: disable=redefine
     args = [self, input, expected_output, metadata]
     return cls.__reward__(*args[:num_args])
   return _reward
+
+
+class QueryInvocation(pg.Object):
+  """A class to represent the invocation of `lf.query`."""
+
+  input: Annotated[
+      Union[lf.Template, pg.Symbolic],
+      'Mapping input of `lf.query`.'
+  ]
+  schema: pg.typing.Annotated[
+      schema_lib.schema_spec(noneable=True),
+      'Schema of `lf.query`.'
+  ]
+  output: Annotated[
+      Any,
+      'Mapping output of `lf.query`.'
+  ]
+  lm: Annotated[
+      lf.LanguageModel,
+      'Language model used for `lf.query`.'
+  ]
+  examples: Annotated[
+      list[mapping.MappingExample],
+      'Fewshot exemplars for `lf.query`.'
+  ] = []
+
+
+@contextlib.contextmanager
+def track_queries(
+    include_child_scopes: bool = True
+) -> Iterator[list[QueryInvocation]]:
+  """Track all queries made during the context.
+
+  Example:
+
+    ```
+    with lf.track_queries() as queries:
+      lf.query('hi', lm=lm)
+      lf.query('What is this {{image}}?', lm=lm, image=image)
+
+    print(queries)
+    ```
+
+  Args:
+    include_child_scopes: If True, the queries made in child scopes will be
+      included in the returned list. Otherwise, only the queries made in the
+      current scope will be included.
+
+  Yields:
+    A list of `QueryInvocation` objects representing the queries made during
+    the context.
+  """
+  trackers = lf.context_value('__query_trackers__', [])
+  tracker = []
+
+  with lf.context(
+      __query_trackers__=[(tracker, include_child_scopes)] + trackers
+  ):
+    try:
+      yield tracker
+    finally:
+      pass

langfun/core/structured/prompting_test.py

@@ -89,7 +89,7 @@ def test_call(self):
     )
     self.assertEqual(
         prompting.query(
-            lf.Template('what is {{x}} + {{y}}'), int, x=1, y=0, lm=lm.clone()
+            lf.Template('what is {{x}} + {{y}}', x=1, y=0), int, lm=lm.clone()
         ),
         1,
     )
@@ -945,5 +945,69 @@ def test_query(self):
     )
 
 
+class TrackQueriesTest(unittest.TestCase):
+
+  def test_include_child_scopes(self):
+    lm = fake.StaticSequence([
+        'bar',
+        'Activity(description="hi")',
+    ])
+    with prompting.track_queries() as queries:
+      prompting.query('foo', lm=lm)
+      with prompting.track_queries() as child_queries:
+        prompting.query('give me an activity', Activity, lm=lm)
+
+    self.assertEqual(len(queries), 2)
+    self.assertTrue(pg.eq(queries[0].input, lf.Template('foo')))
+    self.assertIsNone(queries[0].schema)
+    self.assertEqual(queries[0].output, 'bar')
+    self.assertIs(queries[0].lm, lm)
+
+    self.assertTrue(pg.eq(queries[1].input, lf.Template('give me an activity')))
+    self.assertEqual(queries[1].schema.spec.cls, Activity)
+    self.assertTrue(pg.eq(queries[1].output, Activity(description='hi')))
+    self.assertIs(queries[1].lm, lm)
+
+    self.assertEqual(len(child_queries), 1)
+    self.assertIs(child_queries[0], queries[1])
+
+  def test_exclude_child_scopes(self):
+    lm = fake.StaticSequence([
+        'bar',
+        'Activity(description="hi")',
+    ])
+    with prompting.track_queries(include_child_scopes=False) as queries:
+      prompting.query('foo', lm=lm)
+      with prompting.track_queries(include_child_scopes=False) as child_queries:
+        prompting.query('give me an activity', Activity, lm=lm)
+
+    self.assertEqual(len(queries), 1)
+    self.assertTrue(pg.eq(queries[0].input, lf.Template('foo')))
+    self.assertIsNone(queries[0].schema)
+    self.assertEqual(queries[0].output, 'bar')
+    self.assertIs(queries[0].lm, lm)
+
+    self.assertEqual(len(child_queries), 1)
+    self.assertTrue(
+        pg.eq(child_queries[0].input, lf.Template('give me an activity'))
+    )
+    self.assertEqual(child_queries[0].schema.spec.cls, Activity)
+    self.assertTrue(pg.eq(child_queries[0].output, Activity(description='hi')))
+    self.assertIs(child_queries[0].lm, lm)
+
+  def test_concurrent_map(self):
+
+    def make_query(prompt):
+      _ = prompting.query(prompt, lm=lm)
+
+    lm = fake.StaticSequence([
+        'foo',
+        'bar',
+    ])
+    with prompting.track_queries() as queries:
+      list(lf.concurrent_map(make_query, ['a', 'b']))
+    self.assertEqual(len(queries), 2)
+
+
 if __name__ == '__main__':
   unittest.main()