langfun 0.1.2.dev202504090804__tar.gz → 0.1.2.dev202504100804__tar.gz

{langfun-0.1.2.dev202504090804 → langfun-0.1.2.dev202504100804}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: langfun
-Version: 0.1.2.dev202504090804
+Version: 0.1.2.dev202504100804
 Summary: Langfun: Language as Functions.
 Home-page: https://github.com/google/langfun
 Author: Langfun Authors

{langfun-0.1.2.dev202504090804 → langfun-0.1.2.dev202504100804}/langfun/core/llms/anthropic.py

@@ -509,13 +509,18 @@ class Anthropic(rest.REST):
           raise ValueError(f'Unsupported modality: {chunk!r}.')
       return chunk
 
-    request.update(
-        dict(
-            messages=[
-                prompt.as_format('anthropic', chunk_preprocessor=modality_check)
-            ]
-        )
+    messages = []
+    if system_message := prompt.get('system_message'):
+      assert isinstance(system_message, lf.SystemMessage), type(system_message)
+      messages.append(
+          system_message.as_format(
+              'anthropic', chunk_preprocessor=modality_check
+          )
+      )
+    messages.append(
+        prompt.as_format('anthropic', chunk_preprocessor=modality_check)
     )
+    request.update(messages=messages)
     return request
 
   def _request_args(self, options: lf.LMSamplingOptions) -> dict[str, Any]:

{langfun-0.1.2.dev202504090804 → langfun-0.1.2.dev202504100804}/langfun/core/llms/anthropic_test.py

@@ -35,7 +35,8 @@ def mock_requests_post(url: str, json: dict[str, Any], **kwargs):
       'content': [{
           'type': 'text',
           'text': (
-              f'hello with temperature={json.get("temperature")}, '
+              '\n'.join(c['content'][0]['text'] for c in json['messages']) +
+              f' with temperature={json.get("temperature")}, '
               f'top_k={json.get("top_k")}, '
               f'top_p={json.get("top_p")}, '
               f'max_tokens={json.get("max_tokens")}, '
@@ -140,11 +141,11 @@ class AnthropicTest(unittest.TestCase):
       mock_request.side_effect = mock_requests_post
 
       lm = anthropic.Claude3Haiku(api_key='fake key')
-      self.assertRegex(lm('hi').text, 'hello.*')
+      self.assertRegex(lm('hello').text, 'hello.*')
 
       os.environ['ANTHROPIC_API_KEY'] = 'abc'
       lm = anthropic.Claude3Haiku()
-      self.assertRegex(lm('hi').text, 'hello.*')
+      self.assertRegex(lm('hello').text, 'hello.*')
       del os.environ['ANTHROPIC_API_KEY']
 
   def test_call(self):
@@ -165,6 +166,27 @@ class AnthropicTest(unittest.TestCase):
       self.assertIsNotNone(response.usage.total_tokens, 3)
       self.assertGreater(response.usage.estimated_cost, 0)
 
+  def test_call_with_system_message(self):
+    with mock.patch('requests.Session.post') as mock_request:
+      mock_request.side_effect = mock_requests_post
+      lm = anthropic.Claude3Haiku(api_key='fake_key')
+      response = lm(
+          lf.UserMessage(
+              'hello', system_message=lf.SystemMessage('system')
+          ),
+          temperature=0.0,
+          top_k=0.1,
+          top_p=0.2,
+          stop=['\n'],
+      )
+      self.assertEqual(
+          response.text,
+          (
+              'system\nhello with temperature=0.0, top_k=0.1, top_p=0.2, '
+              "max_tokens=4096, stop=['\\n']."
+          ),
+      )
+
   def test_mm_call(self):
     with mock.patch('requests.Session.post') as mock_mm_request:
       mock_mm_request.side_effect = mock_mm_requests_post

{langfun-0.1.2.dev202504090804 → langfun-0.1.2.dev202504100804}/langfun/core/llms/gemini.py

@@ -582,9 +582,17 @@ class Gemini(rest.REST):
           raise lf.ModalityError(f'Unsupported modality: {chunk!r}') from e
       return chunk
 
-    request['contents'] = [
+    contents = []
+    if system_message := prompt.get('system_message'):
+      assert isinstance(system_message, lf.SystemMessage), type(system_message)
+      contents.append(
+          system_message.as_format(
+              'gemini', chunk_preprocessor=modality_conversion)
+      )
+    contents.append(
         prompt.as_format('gemini', chunk_preprocessor=modality_conversion)
-    ]
+    )
+    request['contents'] = contents
     return request
 
   def _generation_config(

{langfun-0.1.2.dev202504090804 → langfun-0.1.2.dev202504100804}/langfun/core/llms/gemini_test.py

@@ -38,7 +38,7 @@ example_image = (
 def mock_requests_post(url: str, json: dict[str, Any], **kwargs):
   del url, kwargs
   c = pg.Dict(json['generationConfig'])
-  content = json['contents'][0]['parts'][0]['text']
+  content = '\n'.join(c['parts'][0]['text'] for c in json['contents'])
   response = requests.Response()
   response.status_code = 200
   response._content = pg.to_json_str({
@@ -178,6 +178,27 @@ class GeminiTest(unittest.TestCase):
       self.assertEqual(r.metadata.usage.prompt_tokens, 3)
       self.assertEqual(r.metadata.usage.completion_tokens, 4)
 
+  def test_call_model_with_system_message(self):
+    with mock.patch('requests.Session.post') as mock_generate:
+      mock_generate.side_effect = mock_requests_post
+
+      lm = gemini.Gemini('gemini-1.5-pro', api_endpoint='')
+      r = lm(
+          lf.UserMessage('hello', system_message=lf.SystemMessage('system')),
+          temperature=2.0,
+          top_p=1.0,
+          top_k=20,
+          max_tokens=1024,
+          stop='\n',
+      )
+      self.assertEqual(
+          r.text,
+          (
+              'This is a response to system\nhello with temperature=2.0, '
+              'top_p=1.0, top_k=20, max_tokens=1024, stop=\n.'
+          ),
+      )
+
 
 if __name__ == '__main__':
   unittest.main()

{langfun-0.1.2.dev202504090804 → langfun-0.1.2.dev202504100804}/langfun/core/llms/openai_compatible.py

@@ -113,8 +113,9 @@ class OpenAICompatible(rest.REST):
     # Users could use `metadata_system_message` to pass system message.
     system_message = prompt.metadata.get('system_message')
     if system_message:
+      assert isinstance(system_message, lf.SystemMessage), type(system_message)
       messages.append(
-          lf.SystemMessage.from_value(system_message).as_format(
+          system_message.as_format(
               'openai', chunk_preprocessor=modality_check
           )
       )

{langfun-0.1.2.dev202504090804 → langfun-0.1.2.dev202504100804}/langfun/core/llms/openai_compatible_test.py

@@ -465,7 +465,7 @@ class OpenAIComptibleTest(unittest.TestCase):
           lm(
               lf.UserMessage(
                   'hello',
-                  system_message='hi',
+                  system_message=lf.SystemMessage('hi'),
               ),
               sampling_options=lf.LMSamplingOptions(n=2)
           ),

{langfun-0.1.2.dev202504090804 → langfun-0.1.2.dev202504100804}/langfun/core/structured/querying.py

@@ -109,6 +109,7 @@ def query(
     *,
     lm: lf.LanguageModel | list[lf.LanguageModel] | None = None,
     num_samples: int | list[int] = 1,
+    system_message: str | lf.Template | None = None,
     examples: list[mapping.MappingExample] | None = None,
     cache_seed: int | None = 0,
     response_postprocess: Callable[[str], str] | None = None,
@@ -244,6 +245,8 @@ def query(
       If `None`, the LM from `lf.context` will be used.
     num_samples: Number of samples to generate. If a list is provided, its
       length must match the number of models in `lm`.
+    system_message: System instructions to guide the model output. If None,
+      no system message will be used.
     examples: Few-shot examples to guide the model output. Defaults to `None`.
     cache_seed: Seed for caching the query. Queries with the same
       `(lm, prompt, cache_seed)` will use cached responses. If `None`,
@@ -282,6 +285,7 @@ def query(
           schema,
           default=default,
           lm=lm,
+          system_message=system_message,
           examples=examples,
           # Usually num_examples should not be large, so we multiple the user
           # provided cache seed by 100 to avoid collision.
@@ -326,6 +330,13 @@ def query(
   if isinstance(prompt, pg.Symbolic) and prompt.sym_partial and schema is None:
     schema = prompt.__class__
 
+  # Attach system message as input template metadata, which will be passed
+  # through to the rendered message metadata under key `system_message`.
+  if system_message is not None:
+    kwargs['metadata_system_message'] = lf.Template.from_value(
+        system_message
+    ).render(message_cls=lf.SystemMessage)
+
   # Normalize query input.
   if isinstance(prompt, (lf.Message, str)):
     # Query with structured output.

{langfun-0.1.2.dev202504090804 → langfun-0.1.2.dev202504100804}/langfun/core/structured/querying_test.py

@@ -45,14 +45,16 @@ class QueryTest(unittest.TestCase):
       prompt,
       schema,
       examples: list[mapping.MappingExample] | None = None,
+      system_message: str | None = None,
       *,
       expected_snippet: str,
       exact_match: bool = False,
       expected_modalities: int = 0,
+      exepcted_system_message: str | None = None,
       **kwargs,
   ):
     m = querying.query(
-        prompt, schema=schema, examples=examples,
+        prompt, schema=schema, system_message=system_message, examples=examples,
         **kwargs, returns_message=True
     )
     self.assertIsNotNone(m.lm_input)
@@ -64,6 +66,11 @@ class QueryTest(unittest.TestCase):
         len([c for c in m.lm_input.chunk() if isinstance(c, lf.Modality)]),
         expected_modalities,
     )
+    if system_message is not None:
+      self.assertEqual(
+          m.lm_input.system_message.text,
+          exepcted_system_message,
+      )
 
   def test_call(self):
     lm = fake.StaticSequence(['1'])
@@ -114,6 +121,19 @@ class QueryTest(unittest.TestCase):
         Activity(description='hello'),
     )
 
+  def test_render_with_system_message(self):
+    lm = fake.StaticResponse('1')
+    self.assert_render(
+        'What is {{x}} + {{y}}?',
+        schema=None,
+        system_message='You are a helpful assistant.',
+        x=1,
+        y=2,
+        lm=lm.clone(),
+        expected_snippet='What is 1 + 2?',
+        exepcted_system_message='You are a helpful assistant.',
+    )
+
   def test_str_to_structure_render(self):
     lm = fake.StaticResponse('1')
     self.assert_render(

{langfun-0.1.2.dev202504090804 → langfun-0.1.2.dev202504100804}/langfun.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: langfun
-Version: 0.1.2.dev202504090804
+Version: 0.1.2.dev202504100804
 Summary: Langfun: Language as Functions.
 Home-page: https://github.com/google/langfun
 Author: Langfun Authors