langfun 0.1.2.dev202502240805__tar.gz → 0.1.2.dev202502250804__tar.gz

{langfun-0.1.2.dev202502240805 → langfun-0.1.2.dev202502250804}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: langfun
-Version: 0.1.2.dev202502240805
+Version: 0.1.2.dev202502250804
 Summary: Langfun: Language as Functions.
 Home-page: https://github.com/google/langfun
 Author: Langfun Authors

{langfun-0.1.2.dev202502240805 → langfun-0.1.2.dev202502250804}/langfun/core/agentic/action.py

@@ -795,6 +795,84 @@ class Session(pg.Object, pg.views.html.HtmlTreeView.Extension):
           **kwargs
       )
 
+  def query_prompt(
+      self,
+      prompt: Union[str, lf.Template, Any],
+      schema: Union[
+          lf_structured.Schema, Type[Any], list[Type[Any]], dict[str, Any], None
+      ] = None,
+      **kwargs,
+  ) -> Any:
+    """Calls `lf.query_prompt` and associates it with the current invocation.
+
+    The following code are equivalent:
+
+      Code 1:
+      ```
+      session.query_prompt(...)
+      ```
+
+      Code 2:
+      ```
+      with session.track_queries() as queries:
+        output = lf.query_prompt(...)
+      ```
+    The former is preferred when `lf.query_prompt` is directly called by the
+    action.
+    If `lf.query_prompt` is called by a function that does not have access to
+    the
+    session, the latter should be used.
+
+    Args:
+      prompt: The prompt to query.
+      schema: The schema to use for the query.
+      **kwargs: Additional keyword arguments to pass to `lf.query_prompt`.
+
+    Returns:
+      The result of the query.
+    """
+    with self.track_queries():
+      return lf_structured.query_prompt(prompt, schema=schema, **kwargs)
+
+  def query_output(
+      self,
+      response: Union[str, lf.Template, Any],
+      schema: Union[
+          lf_structured.Schema, Type[Any], list[Type[Any]], dict[str, Any], None
+      ] = None,
+      **kwargs,
+  ) -> Any:
+    """Calls `lf.query_output` and associates it with the current invocation.
+
+    The following code are equivalent:
+
+      Code 1:
+      ```
+      session.query_output(...)
+      ```
+
+      Code 2:
+      ```
+      with session.track_queries() as queries:
+        output = lf.query_output(...)
+      ```
+    The former is preferred when `lf.query_output` is directly called by the
+    action.
+    If `lf.query_output` is called by a function that does not have access to
+    the
+    session, the latter should be used.
+
+    Args:
+      response: The response to query.
+      schema: The schema to use for the query.
+      **kwargs: Additional keyword arguments to pass to `lf.query_prompt`.
+
+    Returns:
+      The result of the query.
+    """
+    with self.track_queries():
+      return lf_structured.query_output(response, schema=schema, **kwargs)
+
   def _log(self, level: lf.logging.LogLevel, message: str, **kwargs):
     self._current_action.current_phase.append(
         lf.logging.LogEntry(

{langfun-0.1.2.dev202502240805 → langfun-0.1.2.dev202502250804}/langfun/core/eval/base_test.py

@@ -101,7 +101,7 @@ class EvaluationTest(unittest.TestCase):
     self.assertEqual(s.dir, os.path.join(s.root_dir, s.id))
     self.assertEqual(s.hash, s.clone().hash)
     # Test persistent hash.
-    self.assertEqual(s.hash, 'ae86c703')
+    self.assertEqual(s.hash, '31a1c20a')
     self.assertEqual(
         s.hash, s.clone(override={'max_workers': 2, 'lm.timeout': 20}).hash
     )
@@ -211,7 +211,7 @@ class EvaluationTest(unittest.TestCase):
         s.result,
         dict(
             experiment_setup=dict(
-                id='Evaluation@0fade07d',
+                id='Evaluation@e06b5967',
                 dir=s.dir,
                 model='StaticSequence',
                 prompt_template='{{example.question}}',
@@ -376,7 +376,7 @@ class EvaluationTest(unittest.TestCase):
         s.children[0].dir, os.path.join(s.root_dir, s.children[0].id)
     )
     # Test persistent hash.
-    self.assertEqual(s.hash, 'b66a4e88')
+    self.assertEqual(s.hash, '9f4bc85b')
 
     summary = s.run(verbose=True)
     self.assertEqual(len(summary.evaluations), 2)
@@ -526,10 +526,10 @@ class SuiteTest(unittest.TestCase):
         lm=lm
     )
     # Test for persistent hash.
-    self.assertEqual(s.hash, '26e6cc25')
+    self.assertEqual(s.hash, '2910c323')
     s.run()
     expected = {
-        'Evaluation@0fade07d': dict(
+        'Evaluation@e06b5967': dict(
             experiment_setup=dict(
                 id=s.children[0].id,
                 dir=s.children[0].dir,
@@ -555,7 +555,7 @@ class SuiteTest(unittest.TestCase):
             ),
             usage=s.children[0].result.usage,
         ),
-        'Evaluation@ae86c703': dict(
+        'Evaluation@31a1c20a': dict(
             experiment_setup=dict(
                 id=s.children[1].children[0].id,
                 dir=s.children[1].children[0].dir,

{langfun-0.1.2.dev202502240805 → langfun-0.1.2.dev202502250804}/langfun/core/eval/matching_test.py

@@ -103,7 +103,7 @@ class MatchingTest(unittest.TestCase):
         s.result,
         dict(
             experiment_setup=dict(
-                id='MyTask@739a174b',
+                id='MyTask@a98a284d',
                 dir=s.dir,
                 model='StaticSequence',
                 prompt_template='{{example.question}}',

{langfun-0.1.2.dev202502240805 → langfun-0.1.2.dev202502250804}/langfun/core/eval/scoring_test.py

@@ -81,7 +81,7 @@ class ScoringTest(unittest.TestCase):
         s.result,
         dict(
             experiment_setup=dict(
-                id='ConstraintFollowing@5c88a5eb',
+                id='ConstraintFollowing@90671d5e',
                 dir=s.dir,
                 model='StaticSequence',
                 prompt_template='{{example}}',

{langfun-0.1.2.dev202502240805 → langfun-0.1.2.dev202502250804}/langfun/core/langfunc_test.py

@@ -106,10 +106,10 @@ class LangFuncCallTest(unittest.TestCase):
         "LangFunc(template_str='Hello', clean=True,"
         ' lm=ExcitedEchoer(sampling_options=LMSamplingOptions(temperature=None,'
         ' max_tokens=None, n=1, top_k=40, top_p=None, stop=None,'
-        ' random_seed=None, logprobs=False, top_logprobs=None), cache=None,'
-        ' max_concurrency=None, timeout=120.0, max_attempts=5,'
-        ' retry_interval=(5, 60), exponential_backoff=True,'
-        ' max_retry_interval=300, debug=False))',
+        ' random_seed=None, logprobs=False, top_logprobs=None,'
+        ' max_thinking_tokens=None), cache=None, max_concurrency=None,'
+        ' timeout=120.0, max_attempts=5, retry_interval=(5, 60),'
+        ' exponential_backoff=True, max_retry_interval=300, debug=False))',
     )
 
     l = LangFunc('Hello')

{langfun-0.1.2.dev202502240805 → langfun-0.1.2.dev202502250804}/langfun/core/language_model.py

@@ -560,6 +560,10 @@ class LMSamplingOptions(component.Component):
       ),
   ] = None
 
+  max_thinking_tokens: Annotated[
+      int | None, 'Number of max thinking tokens.'
+  ] = None
+
   def cache_key(self) -> tuple[Any, ...]:
     """Returns a tuple of current values as cache key."""
     return (

{langfun-0.1.2.dev202502240805 → langfun-0.1.2.dev202502250804}/langfun/core/llms/__init__.py

@@ -127,6 +127,8 @@ from langfun.core.llms.openai import Gpt35
 
 # Anthropic models.
 
+from langfun.core.llms.anthropic import Claude37
+from langfun.core.llms.anthropic import Claude37Sonnet_20250219
 from langfun.core.llms.anthropic import Claude35Sonnet
 from langfun.core.llms.anthropic import Claude35Sonnet_20241022
 from langfun.core.llms.anthropic import Claude35Haiku
@@ -139,6 +141,7 @@ from langfun.core.llms.anthropic import Claude3Haiku
 from langfun.core.llms.anthropic import Claude3Haiku_20240307
 
 from langfun.core.llms.vertexai import VertexAIAnthropic
+from langfun.core.llms.vertexai import VertexAIClaude37Sonnet_20250219
 from langfun.core.llms.vertexai import VertexAIClaude35Sonnet_20241022
 from langfun.core.llms.vertexai import VertexAIClaude35Haiku_20241022
 from langfun.core.llms.vertexai import VertexAIClaude3Opus_20240229

{langfun-0.1.2.dev202502240805 → langfun-0.1.2.dev202502250804}/langfun/core/llms/anthropic.py

@@ -88,6 +88,32 @@ SUPPORTED_MODELS = [
             max_output_tokens_per_minute=80_000,
         ),
     ),
+    AnthropicModelInfo(
+        model_id='claude-3-7-sonnet-20250219',
+        provider='Anthropic',
+        in_service=True,
+        description='Claude 3.7 Sonnet model (2/19/2025).',
+        release_date=datetime.datetime(2025, 2, 19),
+        input_modalities=(
+            AnthropicModelInfo.INPUT_IMAGE_TYPES
+            + AnthropicModelInfo.INPUT_DOC_TYPES
+        ),
+        context_length=lf.ModelInfo.ContextLength(
+            max_input_tokens=200_000,
+            max_output_tokens=128_000,
+        ),
+        pricing=lf.ModelInfo.Pricing(
+            cost_per_1m_cached_input_tokens=0.3,
+            cost_per_1m_input_tokens=3,
+            cost_per_1m_output_tokens=15,
+        ),
+        rate_limits=AnthropicModelInfo.RateLimits(
+            # Tier 4 rate limits
+            max_requests_per_minute=4000,
+            max_input_tokens_per_minute=400_000,
+            max_output_tokens_per_minute=80_000,
+        ),
+    ),
     AnthropicModelInfo(
         model_id='claude-3-5-sonnet-20241022',
         provider='Anthropic',
@@ -139,9 +165,34 @@ SUPPORTED_MODELS = [
             cost_per_1m_output_tokens=15,
         ),
         rate_limits=AnthropicModelInfo.RateLimits(
-            # Tier 4 rate limits
-            max_requests_per_minute=4000,
-            max_input_tokens_per_minute=400_000,
+            max_requests_per_minute=100,
+            max_input_tokens_per_minute=1_000_000,
+            max_output_tokens_per_minute=80_000,
+        ),
+    ),
+    AnthropicModelInfo(
+        model_id='claude-3-7-sonnet@20250219',
+        alias_for='claude-3-7-sonnet-20250219',
+        provider='VertexAI',
+        in_service=True,
+        description='Claude 3.7 Sonnet model served on VertexAI (02/19/2025).',
+        release_date=datetime.datetime(2025, 2, 19),
+        input_modalities=(
+            AnthropicModelInfo.INPUT_IMAGE_TYPES
+            + AnthropicModelInfo.INPUT_DOC_TYPES
+        ),
+        context_length=lf.ModelInfo.ContextLength(
+            max_input_tokens=200_000,
+            max_output_tokens=128_000,
+        ),
+        pricing=lf.ModelInfo.Pricing(
+            cost_per_1m_cached_input_tokens=0.3,
+            cost_per_1m_input_tokens=3,
+            cost_per_1m_output_tokens=15,
+        ),
+        rate_limits=AnthropicModelInfo.RateLimits(
+            max_requests_per_minute=100,
+            max_input_tokens_per_minute=1_000_000,
             max_output_tokens_per_minute=80_000,
         ),
     ),
@@ -457,8 +508,7 @@ class Anthropic(rest.REST):
         'x-api-key': self._api_key,
         'anthropic-version': self.api_version,
         'content-type': 'application/json',
-        # TODO(yifenglu): Remove beta flag once the feature is fully supported.
-        'anthropic-beta': 'pdfs-2024-09-25',
+        'anthropic-beta': 'output-128k-2025-02-19',
     }
 
   @functools.cached_property
@@ -506,6 +556,17 @@ class Anthropic(rest.REST):
       args['top_k'] = options.top_k
     if options.top_p is not None:
       args['top_p'] = options.top_p
+    if options.max_thinking_tokens is not None:
+      args['thinking'] = {
+          'type': 'enabled',
+          # Minimum budget is 1,024 tokens.
+          'budget_tokens': options.max_thinking_tokens,
+      }
+      # Thinking isn’t compatible with temperature, top_p, or top_k.
+      # https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking#important-considerations-when-using-extended-thinking
+      args.pop('temperature', None)
+      args.pop('top_k', None)
+      args.pop('top_p', None)
     return args
 
   def _content_from_message(self, prompt: lf.Message) -> list[dict[str, Any]]:
@@ -559,9 +620,27 @@ class Anthropic(rest.REST):
   def _message_from_content(self, content: list[dict[str, Any]]) -> lf.Message:
     """Converts Anthropic's content protocol to message."""
     # Refer: https://docs.anthropic.com/claude/reference/messages-examples
-    return lf.AIMessage.from_chunks(
+    # Thinking: https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking#implementing-extended-thinking # pylint: disable=line-too-long
+    response = lf.AIMessage.from_chunks(
         [x['text'] for x in content if x['type'] == 'text']
     )
+    thinking = lf.AIMessage.from_chunks(
+        [x['thinking'] for x in content if x['type'] == 'thinking']
+    )
+    # thinking is added into the metadata.thinking field.
+    response.set('thinking', thinking)
+    return response
+
+
+class Claude37(Anthropic):
+  """Base class for Claude 3.7 models."""
+
+
+# pylint: disable=invalid-name
+class Claude37Sonnet_20250219(Claude37):
+  """Claude 3.7 Sonnet model (latest)."""
+
+  model = 'claude-3-7-sonnet-20250219'
 
 
 class Claude35(Anthropic):

{langfun-0.1.2.dev202502240805 → langfun-0.1.2.dev202502250804}/langfun/core/llms/vertexai.py

@@ -302,6 +302,12 @@ class VertexAIAnthropic(VertexAI, anthropic.Anthropic):
 # pylint: disable=invalid-name
 
 
+class VertexAIClaude37Sonnet_20250219(VertexAIAnthropic):
+  """Anthropic's Claude 3.7 model on VertexAI."""
+  model = 'claude-3-7-sonnet@20250219'
+  location = 'us-east5'
+
+
 class VertexAIClaude35Sonnet_20241022(VertexAIAnthropic):
   """Anthropic's Claude 3.5 Sonnet model on VertexAI."""
   model = 'claude-3-5-sonnet-v2@20241022'

{langfun-0.1.2.dev202502240805 → langfun-0.1.2.dev202502250804}/langfun.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: langfun
-Version: 0.1.2.dev202502240805
+Version: 0.1.2.dev202502250804
 Summary: Langfun: Language as Functions.
 Home-page: https://github.com/google/langfun
 Author: Langfun Authors