langfun 0.1.2.dev202510230805__py3-none-any.whl → 0.1.2.dev202511270805__py3-none-any.whl

langfun/core/llms/cache/in_memory.py

@@ -24,7 +24,32 @@ import pyglove as pg
 
 @pg.use_init_args(['filename', 'ttl', 'key'])
 class InMemory(base.LMCacheBase):
-  """In memory cache."""
+  """An in-memory cache for language model lookups.
+
+  `InMemory` stores LM prompts and their corresponding responses in memory,
+  providing a simple and fast caching mechanism for a single session.
+  Optionally, it can persist the cache to a JSON file on disk, allowing
+  results to be reused across sessions.
+
+  When a filename is provided, the cache will be loaded from the file upon
+  initialization and saved to the file when `save()` is called. This is
+  useful for caching results in interactive environments like Colab or
+  when running batch jobs.
+
+  Example:
+
+  ```python
+  import langfun as lf
+  # Using in-memory cache without persistence
+  lm = lf.llms.GeminiPro(cache=lf.llms.cache.InMemory())
+  r = lm.query('hello')
+
+  # Using in-memory cache with persistence
+  lm = lf.llms.GeminiPro(cache=lf.llms.cache.InMemory('cache.json'))
+  r = lm.query('hello')
+  lm.cache.save()
+  ```
+  """
 
   filename: Annotated[
       str | None,
@@ -144,17 +169,33 @@ class InMemory(base.LMCacheBase):
 
 @contextlib.contextmanager
 def lm_cache(filename: str | None = None) -> Iterator[InMemory]:
-  """Context manager to enable cache for LMs under the context.
+  """Context manager to enable in-memory cache for LMs in the current context.
+
+  This context manager sets an `InMemory` cache as the default cache for
+  any Langfun language model instantiated within its scope, unless a model
+  is explicitly configured with a different cache.
+
+  If a `filename` is provided, the cache will be loaded from the specified
+  file at the beginning of the context and automatically saved back to the
+  file upon exiting the context. This is a convenient way to manage
+  persistent caching for a block of code.
+
+  Example:
 
-  If LMs under the context manager have explicitly specified cache, they will
-  use their own cache. Otherwise they will use the cache created by the context
-  manager.
+  ```python
+  import langfun as lf
+  with lf.lm_cache('my_cache.json'):
+    # LMs created here will use 'my_cache.json' for caching.
+    lm = lf.llms.GeminiPro()
+    print(lm.query('hello'))
+  ```
 
   Args:
-    filename: If not None, JSON file to load and save the cache.
+    filename: If provided, specifies the JSON file for loading and saving
+      the cache.
 
   Yields:
-    A cache object created.
+    The `InMemory` cache instance created for this context.
   """
   cache = InMemory(filename)
   try:

langfun/core/llms/cache/in_memory_test.py

@@ -175,18 +175,28 @@ class InMemoryLMCacheTest(unittest.TestCase):
 
     cache = in_memory.InMemory()
     lm = fake.StaticSequence(['1', '2', '3', '4', '5', '6'], cache=cache)
-    lm(lf.UserMessage('hi <<[[image]]>>', image=CustomModality('foo')))
-    lm(lf.UserMessage('hi <<[[image]]>>', image=CustomModality('bar')))
+    image_foo = CustomModality('foo')
+    image_bar = CustomModality('bar')
+    lm(
+        lf.UserMessage(
+            f'hi <<[[{image_foo.id}]]>>', referred_modalities=[image_foo]
+        )
+    )
+    lm(
+        lf.UserMessage(
+            f'hi <<[[{image_bar.id}]]>>', referred_modalities=[image_bar]
+        )
+    )
     self.assertEqual(
         list(cache.keys()),
         [
             (
-                'hi <<[[image]]>><image>acbd18db</image>',
+                f'hi <<[[{image_foo.id}]]>>',
                 (None, None, 1, 40, None, None),
                 0,
             ),
             (
-                'hi <<[[image]]>><image>37b51d19</image>',
+                f'hi <<[[{image_bar.id}]]>>',
                 (None, None, 1, 40, None, None),
                 0,
             ),

langfun/core/llms/compositional.py

@@ -21,7 +21,31 @@ import pyglove as pg
 
 @pg.use_init_args(['candidates', 'seed'])
 class RandomChoice(lf.LanguageModel):
-  """Random choice of a list of LLM models."""
+  """A composite language model that randomly selects from a list of candidates.
+
+  `RandomChoice` acts as a proxy that forwards each request (`sample`, `score`,
+  `tokenize`, or `__call__`) to one of the `candidates` selected randomly.
+  This can be useful for load balancing across multiple LLM endpoints,
+  for A/B testing different models, or for ensembling model outputs
+  by calling it multiple times.
+
+  The selection is determined by the provided `seed`, ensuring reproducibility
+  if needed.
+
+  Example:
+
+  ```python
+  import langfun as lf
+
+  lm = lf.llms.RandomChoice([
+      lf.llms.GeminiPro(),
+      lf.llms.GPT4(),
+  ])
+
+  # This call will be handled by either GeminiPro or GPT4, chosen randomly.
+  r = lm.sample('hello')
+  ```
+  """
 
   candidates: Annotated[
       list[lf.LanguageModel],

langfun/core/llms/deepseek.py

@@ -93,8 +93,36 @@ _SUPPORTED_MODELS_BY_ID = {m.model_id: m for m in SUPPORTED_MODELS}
 # DeepSeek API uses an API format compatible with OpenAI.
 # Reference: https://api-docs.deepseek.com/
 @lf.use_init_args(['model'])
-class DeepSeek(openai_compatible.OpenAICompatible):
-  """DeepSeek model."""
+class DeepSeek(openai_compatible.OpenAIChatCompletionAPI):
+  """DeepSeek models.
+
+  **Quick Start:**
+
+  ```python
+  import langfun as lf
+
+  # Call DeepSeek-V3 using API key from environment variable
+  # 'DEEPSEEK_API_KEY'.
+  lm = lf.llms.DeepSeekV3()
+  r = lm('Who are you?')
+  print(r)
+  ```
+
+  **Setting up API key:**
+
+  The DeepSeek API key can be specified in following ways:
+
+  1. At model instantiation:
+
+     ```python
+     lm = lf.llms.DeepSeekV3(api_key='MY_API_KEY')
+     ```
+  2. via environment variable `DEEPSEEK_API_KEY`.
+
+  **References:**
+
+  *   https://api-docs.deepseek.com/
+  """
 
   model: pg.typing.Annotated[
       pg.typing.Enum(

langfun/core/llms/fake.py

@@ -20,7 +20,38 @@ import langfun.core as lf
 
 
 class Fake(lf.LanguageModel):
-  """The base class for all fake language models."""
+  """Base class for fake language models, used for testing.
+
+  Fake models simulate the behavior of real language models but return
+  pre-defined responses, making them useful for testing prompts,
+  data processing logic, and agent behavior without incurring API costs
+  or relying on external services.
+
+  Langfun provides several fake models:
+  * `lf.llms.Echo`: Echoes the prompt back as the response.
+  * `lf.llms.StaticResponse`: Returns a fixed, pre-defined response for
+    any prompt.
+  * `lf.llms.StaticMapping`: Returns responses based on a prompt-to-response
+    dictionary.
+  * `lf.llms.StaticSequence`: Returns responses from a pre-defined sequence
+    in order.
+
+  **Example:**
+
+  ```python
+  import langfun as lf
+
+  # Use Echo model for testing
+  lm = lf.llms.Echo()
+  response = lm('hello')
+  assert response.text == 'hello'
+
+  # Use StaticResponse model
+  lm = lf.llms.StaticResponse('world')
+  response = lm('hello')
+  assert response.text == 'world'
+  ```
+  """
 
   def _score(self, prompt: lf.Message| list[lf.Message],
              completions: list[lf.Message]):

langfun/core/llms/gemini.py

@@ -151,6 +151,32 @@ SUPPORTED_MODELS = [
     #
     # Production models.
     #
+    # Gemini 3 Pro Preview
+    GeminiModelInfo(
+        model_id='gemini-3-pro-preview',
+        in_service=True,
+        provider=pg.oneof(['Google GenAI', 'VertexAI']),
+        model_type='instruction-tuned',
+        description='Gemini 3 Pro Preview.',
+        release_date=datetime.datetime(2025, 11, 18),
+        input_modalities=GeminiModelInfo.ALL_SUPPORTED_INPUT_TYPES,
+        context_length=lf.ModelInfo.ContextLength(
+            max_input_tokens=1_048_576,
+            max_output_tokens=65_536,
+        ),
+        pricing=GeminiModelInfo.Pricing(
+            cost_per_1m_cached_input_tokens=0.2,
+            cost_per_1m_input_tokens=2.0,
+            cost_per_1m_output_tokens=12.0,
+            cost_per_1m_cached_input_tokens_with_prompt_longer_than_128k=0.4,
+            cost_per_1m_input_tokens_with_prompt_longer_than_128k=4.0,
+            cost_per_1m_output_tokens_with_prompt_longer_than_128k=18.0,
+        ),
+        rate_limits=lf.ModelInfo.RateLimits(
+            max_requests_per_minute=2000,
+            max_tokens_per_minute=4_000_000,
+        ),
+    ),
     # Gemini 2.5 Flash
     GeminiModelInfo(
         model_id='gemini-2.5-flash',
@@ -696,7 +722,15 @@ _SUPPORTED_MODELS_BY_ID = {m.model_id: m for m in SUPPORTED_MODELS}
 
 @pg.use_init_args(['model'])
 class Gemini(rest.REST):
-  """Language models provided by Google GenAI."""
+  """Base class for Gemini models served on Google GenAI and Vertex AI.
+
+  This class implements the Gemini API protocol, shared by
+  `lf.llms.GoogleGenAI` and `lf.llms.VertexAI`, providing common request
+  formatting and response parsing for Gemini models.
+
+  It is not intended to be used directly. Please use `lf.llms.GoogleGenAI` or
+  `lf.llms.VertexAI` instead.
+  """
 
   model: pg.typing.Annotated[
       pg.typing.Enum(
@@ -752,11 +786,8 @@ class Gemini(rest.REST):
         prompt.as_format('gemini', chunk_preprocessor=modality_conversion)
     )
     request['contents'] = contents
-    # Users could use `metadata_gemini_tools` to pass Gemini tools. For example,
-    # for enabling Search Grounding, users could pass:
-    # metadata_gemini_tools=[{'google_search': {}}]
-    if tools := prompt.metadata.get('gemini_tools'):
-      request['tools'] = tools
+    if sampling_options.extras:
+      request.update(sampling_options.extras)
     return request
 
   def _generation_config(
@@ -788,11 +819,14 @@ class Gemini(rest.REST):
           + '\n\n [RESPONSE FORMAT (not part of prompt)]\n'
           + pg.to_json_str(json_schema, json_indent=2)
       )
+    thinking_config_data = {}
     if options.max_thinking_tokens is not None:
-      config['thinkingConfig'] = {
-          'includeThoughts': options.max_thinking_tokens > 0,
-          'thinkingBudget': options.max_thinking_tokens,
-      }
+      thinking_config_data['includeThoughts'] = options.max_thinking_tokens > 0
+      thinking_config_data['thinkingBudget'] = options.max_thinking_tokens
+    if options.thinking_level is not None:
+      thinking_config_data['thinkingLevel'] = options.thinking_level
+    if thinking_config_data:
+      config['thinkingConfig'] = thinking_config_data
 
     if self.response_modalities:
       config['responseModalities'] = self.response_modalities
@@ -808,10 +842,14 @@ class Gemini(rest.REST):
           'No candidates found in response. This is a Gemini API issue that '
           'happens occasionally, and retrying should fix it. '
       )
-    messages = [
-        lf.Message.from_value(candidate['content'], format='gemini')
-        for candidate in candidates
-    ]
+
+    messages = []
+    for candidate in candidates:
+      message = lf.Message.from_value(candidate['content'], format='gemini')
+      if finish_reason := candidate.get('finishReason'):
+        message.metadata['finish_reason'] = finish_reason
+      messages.append(message)
+
     usage = json['usageMetadata']
     input_tokens = usage['promptTokenCount']
     # NOTE(daiyip): We saw cases that `candidatesTokenCount` is not present.
@@ -833,9 +871,9 @@ class Gemini(rest.REST):
     )
 
   def _error(self, status_code: int, content: str) -> lf.LMError:
-    if (
-        status_code == 400
-        and b'exceeds the maximum number of tokens' in content
+    if status_code == 400 and (
+        b'exceeds the maximum number of tokens' in content
+        or b'Reduce the input token count and try again.' in content
     ):
       return lf.ContextLimitError(f'{status_code}: {content}')
     return super()._error(status_code, content)

langfun/core/llms/gemini_test.py

@@ -177,6 +177,58 @@ class GeminiTest(unittest.TestCase):
         ),
     )
 
+    # Add test for thinkingConfig with thinking_level.
+    actual = model._generation_config(
+        lf.UserMessage('hi'),
+        lf.LMSamplingOptions(
+            thinking_level='high',
+        ),
+    )
+    self.assertEqual(
+        actual,
+        dict(
+            candidateCount=1,
+            temperature=None,
+            topP=None,
+            topK=40,
+            maxOutputTokens=None,
+            stopSequences=None,
+            responseLogprobs=False,
+            logprobs=None,
+            seed=None,
+            thinkingConfig={'thinkingLevel': 'high'},
+        ),
+    )
+
+    # Add test for thinkingConfig with both max_thinking_tokens and
+    # thinking_level.
+    actual = model._generation_config(
+        lf.UserMessage('hi'),
+        lf.LMSamplingOptions(
+            max_thinking_tokens=100,
+            thinking_level='low',
+        ),
+    )
+    self.assertEqual(
+        actual,
+        dict(
+            candidateCount=1,
+            temperature=None,
+            topP=None,
+            topK=40,
+            maxOutputTokens=None,
+            stopSequences=None,
+            responseLogprobs=False,
+            logprobs=None,
+            seed=None,
+            thinkingConfig={
+                'includeThoughts': True,
+                'thinkingBudget': 100,
+                'thinkingLevel': 'low',
+            },
+        ),
+    )
+
     with self.assertRaisesRegex(
         ValueError, '`json_schema` must be a dict, got'
     ):
@@ -225,6 +277,38 @@ class GeminiTest(unittest.TestCase):
       ):
         lm('hello')
 
+  def test_call_model_with_max_tokens_error(self):
+    def mock_requests_post_error(*args, **kwargs):
+      del args, kwargs
+      response = requests.Response()
+      response.status_code = 200
+      response._content = pg.to_json_str({
+          'candidates': [
+              {
+                  'finishReason': 'MAX_TOKENS',
+                  'content': {
+                      'parts': [
+                          {
+                              'text': 'This is'
+                          }
+                      ]
+                  }
+              },
+          ],
+          'usageMetadata': {
+              'promptTokenCount': 3,
+              'candidatesTokenCount': 4,
+          }
+      }).encode()
+      return response
+
+    with mock.patch('requests.Session.post') as mock_generate:
+      mock_generate.side_effect = mock_requests_post_error
+      lm = gemini.Gemini('gemini-1.5-pro', api_endpoint='')
+      m = lm('hello')
+      self.assertEqual(m.metadata.finish_reason, 'MAX_TOKENS')
+      self.assertEqual(m.text, 'This is')
+
   def test_call_model_with_system_message(self):
     with mock.patch('requests.Session.post') as mock_generate:
       mock_generate.side_effect = mock_requests_post

langfun/core/llms/google_genai.py

@@ -25,7 +25,35 @@ import pyglove as pg
 @lf.use_init_args(['model'])
 @pg.members([('api_endpoint', pg.typing.Str().freeze(''))])
 class GenAI(gemini.Gemini):
-  """Language models provided by Google GenAI."""
+  """Google GenAI models.
+
+  **Quick Start:**
+
+  ```python
+  import langfun as lf
+
+  # Call Gemini 1.5 Flash using API key from environment variable
+  # 'GOOGLE_API_KEY'.
+  lm = lf.llms.Gemini15Flash()
+  r = lm('Who are you?')
+  print(r)
+  ```
+
+  **Setting up API key:**
+
+  The Google API key can be specified in following ways:
+
+  1. At model instantiation:
+
+     ```python
+     lm = lf.llms.Gemini15Flash(api_key='MY_API_KEY')
+     ```
+  2. via environment variable `GOOGLE_API_KEY`.
+
+  **References:**
+
+  *   https://ai.google.dev/docs
+  """
 
   model: pg.typing.Annotated[
       pg.typing.Enum(
@@ -87,9 +115,14 @@ class GenAI(gemini.Gemini):
 
 # pylint: disable=invalid-name
 
+
 #
 # Experimental models.
 #
+class Gemini3ProPreview(GenAI):
+  """Gemini 3 Pro Preview model."""
+
+  model = 'gemini-3-pro-preview'
 
 
 class Gemini25FlashImagePreview(GenAI):

langfun/core/llms/groq.py

@@ -259,10 +259,35 @@ _SUPPORTED_MODELS_BY_ID = {m.model_id: m for m in SUPPORTED_MODELS}
 
 
 @lf.use_init_args(['model'])
-class Groq(openai_compatible.OpenAICompatible):
-  """Groq LLMs through REST APIs (OpenAI compatible).
+class Groq(openai_compatible.OpenAIChatCompletionAPI):
+  """Groq models.
 
-  See https://platform.openai.com/docs/api-reference/chat
+  **Quick Start:**
+
+  ```python
+  import langfun as lf
+
+  # Call Llama 3.3 70B on Groq using API key from environment variable
+  # 'GROQ_API_KEY'.
+  lm = lf.llms.GroqLlama33_70B_Versatile()
+  r = lm('Who are you?')
+  print(r)
+  ```
+
+  **Setting up API key:**
+
+  The Groq API key can be specified in following ways:
+
+  1. At model instantiation:
+
+     ```python
+     lm = lf.llms.GroqLlama33_70B_Versatile(api_key='MY_API_KEY')
+     ```
+  2. via environment variable `GROQ_API_KEY`.
+
+  **References:**
+
+  *   https://console.groq.com/docs
   """
 
   model: pg.typing.Annotated[

langfun/core/llms/llama_cpp.py

@@ -20,11 +20,30 @@ import pyglove as pg
 
 @pg.use_init_args(['url', 'model'])
 @pg.members([('api_endpoint', pg.typing.Str().freeze(''))])
-class LlamaCppRemote(openai_compatible.OpenAICompatible):
-  """The remote LLaMA C++ model.
+class LlamaCppRemote(openai_compatible.OpenAIChatCompletionAPI):
+  """LLaMA C++ models served via a remote server.
 
-  The Remote LLaMA C++ models can be launched via
-  https://github.com/ggerganov/llama.cpp/tree/master/examples/server
+  This class provides an interface to interact with language models
+  hosted on a LLaMA C++ server, which is compatible with the OpenAI
+  Chat Completions API format.
+
+  **Quick Start:**
+
+  Assuming a LLaMA C++ server is running at `http://localhost:8080`,
+  you can interact with it as follows:
+
+  ```python
+  import langfun as lf
+
+  # If model name is not specified, it will use server's default.
+  lm = lf.llms.LlamaCppRemote(url='http://localhost:8080')
+  r = lm('Who are you?')
+  print(r)
+  ```
+
+  **References:**
+
+  *   https://github.com/ggerganov/llama.cpp/tree/master/examples/server
   """
   url: Annotated[
       str,

langfun/core/llms/openai.py

@@ -1031,8 +1031,36 @@ _SUPPORTED_MODELS_BY_MODEL_ID = {m.model_id: m for m in SUPPORTED_MODELS}
 
 
 @lf.use_init_args(['model'])
-class OpenAI(openai_compatible.OpenAICompatible):
-  """OpenAI model."""
+class OpenAI(openai_compatible.OpenAIResponsesAPI):
+  """OpenAI models.
+
+  **Quick Start:**
+
+  ```python
+  import langfun as lf
+
+  # Call GPT-4o using API key from environment variable 'OPENAI_API_KEY'.
+  lm = lf.llms.Gpt4o()
+  r = lm('Who are you?')
+  print(r)
+  ```
+
+  **Setting up API key:**
+
+  The OpenAI API key can be specified in following ways:
+
+  1. At model instantiation:
+
+     ```python
+     lm = lf.llms.Gpt4o(api_key='MY_API_KEY')
+     ```
+  2. via environment variable `OPENAI_API_KEY`.
+
+  **References:**
+
+  *   https://platform.openai.com/docs/models
+  *   https://platform.openai.com/docs/api-reference
+  """
 
   model: pg.typing.Annotated[
       pg.typing.Enum(
@@ -1041,7 +1069,12 @@ class OpenAI(openai_compatible.OpenAICompatible):
       'The name of the model to use.',
   ]
 
-  api_endpoint: str = 'https://api.openai.com/v1/chat/completions'
+  # Disable message storage by default.
+  sampling_options = lf.LMSamplingOptions(
+      extras={'store': False}
+  )
+
+  api_endpoint: str = 'https://api.openai.com/v1/responses'
 
   api_key: Annotated[
       str | None,