docent-python 0.1.41a0__tar.gz → 0.1.43a0__tar.gz

{docent_python-0.1.41a0 → docent_python-0.1.43a0}/.gitignore

@@ -198,3 +198,6 @@ inspect_evals
 
 # test data cache
 data/cache
+
+# dont commit package lock, force use of bun lock
+package-lock.json

{docent_python-0.1.41a0 → docent_python-0.1.43a0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: docent-python
-Version: 0.1.41a0
+Version: 0.1.43a0
 Summary: Docent SDK
 Project-URL: Homepage, https://github.com/TransluceAI/docent
 Project-URL: Issues, https://github.com/TransluceAI/docent/issues

{docent_python-0.1.41a0 → docent_python-0.1.43a0}/docent/_llm_util/data_models/exceptions.py

@@ -1,7 +1,24 @@
+from typing import Any
+
+
 class LLMException(Exception):
     error_type_id = "other"
     user_message = "The model failed to respond. Please try again later."
 
+    def serialize(self) -> dict[str, Any]:
+        data: dict[str, Any] = {
+            "type": self.__class__.__name__,
+            "user_message": getattr(self, "user_message", None),
+            "error_type_id": getattr(self, "error_type_id", None),
+        }
+        if failed_output := getattr(self, "failed_output", None):
+            data["failed_output"] = str(failed_output)
+        return data
+
+    @classmethod
+    def serialize_llm_errors(cls, errors: list["LLMException"]) -> list[dict[str, Any]]:
+        return [error.serialize() for error in errors]
+
 
 class CompletionTooLongException(LLMException):
     error_type_id = "completion_too_long"

{docent_python-0.1.41a0 → docent_python-0.1.43a0}/docent/_llm_util/llm_svc.py

@@ -433,40 +433,3 @@ class BaseLLMService:
                 break
 
         return outputs
-
-
-async def get_llm_completions_async(
-    inputs: list[MessagesInput],
-    model_options: list[ModelOption],
-    tools: list[ToolInfo] | None = None,
-    tool_choice: Literal["auto", "required"] | None = None,
-    max_new_tokens: int = 1024,
-    temperature: float = 1.0,
-    logprobs: bool = False,
-    top_logprobs: int | None = None,
-    timeout: float = 120.0,
-    streaming_callback: AsyncLLMOutputStreamingCallback | None = None,
-    validation_callback: AsyncLLMOutputStreamingCallback | None = None,
-    completion_callback: AsyncLLMOutputStreamingCallback | None = None,
-    use_cache: bool = False,
-    _api_key_overrides: dict[str, str] = dict(),
-) -> list[LLMOutput]:
-    """Convenience method for backward compatibility"""
-
-    svc = BaseLLMService()
-    return await svc.get_completions(
-        inputs=inputs,
-        model_options=model_options,
-        tools=tools,
-        tool_choice=tool_choice,
-        max_new_tokens=max_new_tokens,
-        temperature=temperature,
-        logprobs=logprobs,
-        top_logprobs=top_logprobs,
-        timeout=timeout,
-        streaming_callback=streaming_callback,
-        validation_callback=validation_callback,
-        completion_callback=completion_callback,
-        use_cache=use_cache,
-        _api_key_overrides=_api_key_overrides,
-    )

{docent_python-0.1.41a0 → docent_python-0.1.43a0}/docent/_llm_util/model_registry.py

@@ -78,11 +78,27 @@ _REGISTRY: list[tuple[str, ModelInfo]] = [
     ),
     (
         "gemini-2.5-pro",
+        # TODO(mengk, ryan): this is wrong for prompts > 200k
         ModelInfo(
             rate={"input": 1.25, "output": 10.00},
             context_window=1_000_000,
         ),
     ),
+    (
+        "gemini-3-pro-preview",
+        # TODO(mengk, ryan): this is wrong for prompts > 200k
+        ModelInfo(
+            rate={"input": 2.00, "output": 12.00},
+            context_window=1_048_576,
+        ),
+    ),
+    (
+        "gemini-3-flash-preview",
+        ModelInfo(
+            rate={"input": 0.50, "output": 3.00},
+            context_window=1_048_576,
+        ),
+    ),
     (
         "grok-4-fast",
         ModelInfo(

{docent_python-0.1.41a0 → docent_python-0.1.43a0}/docent/_llm_util/providers/openai.py

@@ -465,7 +465,7 @@ def chunk_and_tokenize(
     chunk_to_doc: list[int] = []
 
     for i, item in enumerate(text):
-        tokens = encoding.encode(item)
+        tokens = encoding.encode(item, disallowed_special=())
         if len(tokens) <= window_size:
             chunks = [tokens]
         else:

{docent_python-0.1.41a0 → docent_python-0.1.43a0}/docent/_llm_util/providers/preference_types.py

@@ -98,6 +98,10 @@ class PublicProviderPreferences(BaseModel):
                 model_name="claude-sonnet-4-5",
                 reasoning_effort="medium",
             ),
+            ModelOption(
+                provider="google",
+                model_name="gemini-3-flash-preview",
+            ),
         ]
 
 

{docent_python-0.1.41a0 → docent_python-0.1.43a0}/docent/data_models/_tiktoken_util.py

@@ -6,13 +6,13 @@ MAX_TOKENS = 100_000
 def get_token_count(text: str, model: str = "gpt-4") -> int:
     """Get the number of tokens in a text under the GPT-4 tokenization scheme."""
     encoding = tiktoken.encoding_for_model(model)
-    return len(encoding.encode(text))
+    return len(encoding.encode(text, disallowed_special=()))
 
 
 def truncate_to_token_limit(text: str, max_tokens: int, model: str = "gpt-4") -> str:
     """Truncate text to stay within the specified token limit."""
     encoding = tiktoken.encoding_for_model(model)
-    tokens = encoding.encode(text)
+    tokens = encoding.encode(text, disallowed_special=())
 
     if len(tokens) <= max_tokens:
         return text