- Stripping whitespace from custom LLM provider and model names when checking for "ollama_chat".
- Enhancing `_flatten_ollama_content` to correctly handle content that is None, a string, a dictionary, or an iterable (like a tuple) of content blocks, not just lists. This aligns with LiteLLM's `OpenAIMessageContent` type being an `Iterable`.
Close#3928
Co-authored-by: George Weale <gweale@google.com>
PiperOrigin-RevId: 845848017
This change ensures that file URI parts passed to LiteLLM always include a "format" field. If `mime_type` is not explicitly provided in `FileData`, the system attempts to infer it from the URI's file extension. If inference fails, a default "application/octet-stream" is used. This is necessary because LiteLLM's Vertex AI backend requires the "format" field for GCS URIs.
Close#3787
Co-authored-by: George Weale <gweale@google.com>
PiperOrigin-RevId: 843753810
LiteLLM's StreamHandlers output to stderr by default. In cloud environments like GCP, stderr output is treated as ERROR severity regardless of actual log level, causing INFO-level logs to be incorrectly classified as errors.
This change redirects LiteLLM loggers to stdout in two places:
- In `lite_llm.py`: Immediately after litellm import
- In `logs.py`: When `setup_adk_logger()` is called (with guard to check if litellm is imported)
Close#3824
Co-authored-by: George Weale <gweale@google.com>
PiperOrigin-RevId: 843393874
LiteLLM's `ollama_chat` provider does not accept array-based content in messages. This change flattens multipart content by joining text parts or JSON-serializing non-text parts before sending the request to the LiteLLM completion API. This ensures compatibility with Ollama's chat endpoint.
Close#3727
Co-authored-by: George Weale <gweale@google.com>
PiperOrigin-RevId: 843382361
The `_to_litellm_response_format` function now adapts the output format based on the provided model. Gemini models continue to use the "response_schema" key, while OpenAI-compatible models (including Azure OpenAI and Anthropic) now use the "json_schema" key as per LiteLLM's documentation for JSON mode. The schema name is also included in the "json_schema" format.
Close#3713Close#3890
Co-authored-by: George Weale <gweale@google.com>
PiperOrigin-RevId: 843326850
When converting `types.Content` with a `function_response` to LiteLLM's `ChatCompletionToolMessage`, if the response is already a string, use it directly. Otherwise, serialize the response to JSON. This prevents double-serialization of string payloads
Close#3676
Co-authored-by: George Weale <gweale@google.com>
PiperOrigin-RevId: 840013822
This change introduces `Gemma3Ollama`, a new LLM model class for running Gemma 3 models locally via Ollama, leveraging LiteLLM. The function calling logic previously in the `Gemma` class has been refactored into a `GemmaFunctionCallingMixin` and is now used by both `Gemma` and `Gemma3Ollama`. A new sample application, `hello_world_gemma3_ollama`, is added to demonstrate using `Gemma3Ollama` with an agent. Unit tests for `Gemma3Ollama` are also included.
Merge: https://github.com/google/adk-python/pull/3120
Co-authored-by: George Weale <gweale@google.com>
PiperOrigin-RevId: 839996879
This change expands the supported file MIME types and introduces provider-specific handling for file uploads. For providers like OpenAI and Azure, inline file data is now uploaded via `litellm.acreate_file` to obtain a `file_id`, which is then used in the message content. Other providers continue to use base64 encoded file data. Affected functions have been updated to be asynchronous
Merge:https://github.com/google/adk-python/pull/2863
Co-authored-by: George Weale <gweale@google.com>
PiperOrigin-RevId: 839996848
The Gemini API may not always send an explicit transcription finished signal. This change ensures that any buffered input or output transcription text is yielded as a finished transcription when a turn is completed, generation is complete, or the session is interrupted.
Also, refined the check for `event.partial` in runners.py to be more explicit.
Co-authored-by: Hangfei Lin <hangfei@google.com>
PiperOrigin-RevId: 839008606
This change introduces an `AnthropicLlm` base class for direct Anthropic API calls using `AsyncAnthropic`. The existing `Claude` class now inherits from `AnthropicLlm` and is specialized to use `AsyncAnthropicVertex` for models hosted on Vertex AI. The `messages.create` call is now properly awaited
Merge: https://github.com/google/adk-python/pull/2904
Co-authored-by: George Weale <gweale@google.com>
PiperOrigin-RevId: 838851026
Previously, image parts were always filtered out when converting content to Anthropic message parameters. This change updates the logic to only filter out image parts and log a warning when the content role is not "user". This enables sending image data as part of user prompts to Claude models
Merges: https://github.com/google/adk-python/pull/3286
Co-authored-by: George Weale <gweale@google.com>
PiperOrigin-RevId: 836725196
This change replaces the use of `ChatCompletionDeveloperMessage` with `ChatCompletionSystemMessage` and sets the role to "system" for providing system instructions to LiteLLM models
Close#3657
Co-authored-by: George Weale <gweale@google.com>
PiperOrigin-RevId: 835388738
LlmResponse/Event now keep both provider reasoning output and the raw vendor payload so callbacks and loggers can inspect hidden “thoughts” or trace bugs without rewriting adapters.
LiteLLM’s adapter and streaming loop emit reasoning chunks alongside text and aggregate them into final events -> all responses now carry a JSON-safe copy of the source payload for debug. UnsafeLocalCodeExecutor uses the documented exec(code, globals, globals) form, letting helper functions defined inside snippets call each other.
Close#1749
Co-authored-by: George Weale <gweale@google.com>
PiperOrigin-RevId: 834956847
