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 `input_transcription` and `output_transcription` fields to session events, enabling the storage and retrieval of transcription data in both the database and Vertex AI session services.
Closes#3172
Co-authored-by: Hangfei Lin <hangfei@google.com>
PiperOrigin-RevId: 834366848
Fixes https://github.com/google/adk-python/issues/3174
The event compaction process, configured via `EventsCompactionConfig`, was
previously scheduled as a background task using `asyncio.create_task`.
Because Python's `asyncio.create_task` only holds a weak reference to
the created task and no strong reference was maintained by ADK, the
compaction task could be garbage-collected before it finished executing.
This resulted in event compaction failing silently or only partially
running, preventing session history from being summarized.
### Approaches Considered
Two approaches were considered to fix this:
1. **`asyncio.create_task` + Reference:** Create the task with `create_task` and store a strong reference to it (e.g., in a `set` on the [Runner](http://_vscodecontentref_/0) instance), removing it only when complete via `task.add_done_callback()`.
* **Pros:** The [run_async](http://_vscodecontentref_/1) async generator finishes immediately after yielding the last agent event.
* **Cons:** Adds complexity to the Runner state; background task failures are silent to the [run_async](http://_vscodecontentref_/2) caller; requires enhancement to `runner.close()` to correctly manage pending tasks on shutdown.
2. **`await`:** Directly `await` the compaction coroutine at the end of [run_async](http://_vscodecontentref_/3) after all agent events have been yielded.
* **Pros:** Simple to implement; ensures compaction runs to completion; failures during compaction propagate immediately to the [run_async](http://_vscodecontentref_/4) caller, making them visible and easier to debug.
* **Cons:** The `async for` loop iterating over [run_async](http://_vscodecontentref_/5) will not terminate until compaction finishes.
### Decision
This change implements the `await` approach. Although it means the `async for` loop takes longer to terminate when compaction occurs, it was chosen for its **simplicity and robustness**. Ensuring that compaction either succeeds or fails visibly is preferable to silent background failures.
All agent response events are yielded *before* compaction starts, so there is **no user-perceived delay in receiving the agent's answer** for the current turn.
### Integration Note for Users
Because compaction is now awaited, code consuming events via `async for event in runner.run_async(...)` will only finish iterating *after* compaction is complete (if compaction is triggered for that invocation).
If your application only needs the agent's response to proceed (e.g., displaying a message in a UI and allowing the user to reply), you can process events as they arrive and initiate the next turn without waiting for the `async for` loop to fully terminate. A new call to [run_async](http://_vscodecontentref_/6) for the next user query can be made immediately and will execute concurrently.
**Example:**
```python
async def handle_agent_turn(runner, message):
print("Agent is thinking...")
async for event in runner.run_async(user_id='...', session_id='...', new_message=message):
# Stream events to UI, log, etc.
if event.author == 'model' and event.content and event.content.parts[0].text:
print(f"Agent response: {event.content.parts[0].text}")
# The agent has provided a text response.
# The application can now enable user input for the next turn,
# even though this async for loop might not finish immediately
# if compaction is running.
print("Invocation complete (including compaction if any).")
# In your application:
# A new call to handle_agent_turn(runner, next_message) can be made
# as soon as the user provides the next input, without waiting for
# the previous call's generator to be exhausted.
Co-authored-by: Hangfei Lin <hangfei@google.com>
PiperOrigin-RevId: 833885375
This change:
* Deprecates the `save_live_audio` configuration option in `RunConfig`, introducing `save_live_blob` as its replacement. A warning is issued if `save_live_audio` is used.
* Updates `base_llm_flow.py` to use the new `save_live_blob` flag.
* Ensures that audio events generated by `audio_cache_manager.py` are properly appended to the session service.
* Adds a utility script `pcm_audio_player.py` for playing raw PCM audio files.
Input sample event: 14a5859f-6b6c-46ed-9f28-e5008793b1c6|live_bidi_streaming_multi_agent|user|1e867c7f-dbe1-4268-a7bc-7a9fa5fbd16c|e-7a28a060-29bf-4483-bc5c-17248698a897|1762916981.5932|{"content":{"parts":[{"file_data":{"file_uri":"artifact://live_bidi_streaming_multi_agent/user/1e867c7f-dbe1-4268-a7bc-7a9fa5fbd16c/_adk_live/adk_live_audio_storage_input_audio_1762916981593.pcm#0","mime_type":"audio/pcm"}}],"role":"user"},"invocation_id":"e-7a28a060-29bf-4483-bc5c-17248698a897","author":"user","actions":{"state_delta":{},"artifact_delta":{},"requested_auth_configs":{},"requested_tool_confirmations":{}},"id":"14a5859f-6b6c-46ed-9f28-e5008793b1c6","timestamp":1762916981.5932002}
output sample event:
506c9df4-e143-4ebc-90a5-2f5b2eb26754|live_bidi_streaming_multi_agent|user|1e867c7f-dbe1-4268-a7bc-7a9fa5fbd16c|e-7a28a060-29bf-4483-bc5c-17248698a897|1762916986.10579|{"content":{"parts":[{"file_data":{"file_uri":"artifact://live_bidi_streaming_multi_agent/user/1e867c7f-dbe1-4268-a7bc-7a9fa5fbd16c/_adk_live/adk_live_audio_storage_output_audio_1762916986105.pcm;rate=24000#0","mime_type":"audio/pcm;rate=24000"}}],"role":"model"},"invocation_id":"e-7a28a060-29bf-4483-bc5c-17248698a897","author":"model","actions":{"state_delta":{},"artifact_delta":{},"requested_auth_configs":{},"requested_tool_confirmations":{}},"id":"506c9df4-e143-4ebc-90a5-2f5b2eb26754","timestamp":1762916986.105794}
Co-authored-by: Hangfei Lin <hangfei@google.com>
PiperOrigin-RevId: 831512074
This CL adds debug logs to show the history being sent to the live connection and the live connect config used.
Co-authored-by: Hangfei Lin <hangfei@google.com>
PiperOrigin-RevId: 831473597
The agent's instructions are updated to guide it to guess a positive integer, starting from 50 and decreasing, using the `guess_number_tool` until the target is found. The maximum number of retries for the `CustomRetryPlugin` is increased from 6 to 10.
Co-authored-by: Hangfei Lin <hangfei@google.com>
PiperOrigin-RevId: 830984481
Enhanced `save_artifact` in `callback_context.py` to accept `custom_metadata` and added `get_artifact_version` to retrieve artifact details.
Introduced a new sample, `context_offloading_with_artifact`, demonstrating how to use ADK artifacts to offload large data from the LLM context. The sample includes:
- `QueryLargeDataTool`: Generates mock sales reports, saves them as artifacts with custom metadata, and injects the artifact content into the LLM request immediately after creation.
- `CustomLoadArtifactsTool`: Provides summaries of available artifacts to the LLM based on metadata and loads artifact content on demand when `load_artifacts` is called.
Co-authored-by: Hangfei Lin <hangfei@google.com>
PiperOrigin-RevId: 830592786
This change adds a debug log statement to output the function response IDs and the event list when a matching function call event is not found, which helps in debugging.
Co-authored-by: Hangfei Lin <hangfei@google.com>
PiperOrigin-RevId: 829059186
In `gemini_llm_connection.py`, accumulate partial transcription texts and emit `LlmResponse` with `partial=True` for each chunk. When the transcription is marked as `finished`, emit a final `LlmResponse` with the full accumulated text and `partial=False`.
In `runners.py`, modify `_should_append_to_history` to only add transcription events to the history when they are fully finished, preventing partial transcriptions from being added.
Co-authored-by: Hangfei Lin <hangfei@google.com>
PiperOrigin-RevId: 829029715
The `_run_compaction_default` method is no longer called and can be safely removed.
Co-authored-by: Hangfei Lin <hangfei@google.com>
PiperOrigin-RevId: 828760157
The blob content is often large and binary, which makes the logs unreadable and can cause excessive logging.
Co-authored-by: Hangfei Lin <hangfei@google.com>
PiperOrigin-RevId: 828523413
The sample agent now uses updated model names for Gemini Live, including a new Vertex model as the default and a new AI Studio model option.
Co-authored-by: Hangfei Lin <hangfei@google.com>
PiperOrigin-RevId: 828515811
This change introduces a new section in the README.md to highlight the `adk-python-community` GitHub repository, describing it as a place for community-contributed tools and integrations.
Co-authored-by: Hangfei Lin <hangfei@google.com>
PiperOrigin-RevId: 828155205
This change adds a new section to the README.md, detailing past community events. The first entry is for the completed ADK's 1st community meeting, with links to the recording and deck.
PiperOrigin-RevId: 823322924
The "What's new" section has been updated to highlight: Custom API Registration, Rewind functionality, and a new CodeExecutor utilizing the Vertex AI Code Execution Sandbox API. Previous updates on context compaction, resumability, ReflectRetryToolPlugin, and search tool support have been removed.
PiperOrigin-RevId: 822848112
The section detailing the first ADK community call scheduled for Oct 15, 2025, including date, time, meeting links, and agenda, has been removed.
The section detailing the first ADK community call scheduled for Oct 15, 2025, including date, time, meeting links, and agenda, has been removed.
PiperOrigin-RevId: 822843889
This release includes updates to the changelog with new features across Core, Evals, Integrations, Observability, Services, Tools, and UI, along with various bug fixes and improvements. The base CL number has also been updated.
PiperOrigin-RevId: 822730595
This CL updates the "What's new" section to include Resumability, ReflectRetryToolPlugin, Context compaction, and Search tool support. It also moves "Agent Config" and "Tool Confirmation" from "What's new" to "Key Features".
PiperOrigin-RevId: 817469210
The added section provides details for the community call on Oct 15, 2025, including the agenda and links to join and add to calendars.
PiperOrigin-RevId: 817457276
When there are multiple intervals and compactions, the original implementation only keep the last one. The right implementation is to keep as many compaction events/summary as the requested internals.
PiperOrigin-RevId: 816516662
If `EventsCompactionConfig` is provided without a `compactor`, a `SlidingWindowCompactor` is now automatically instantiated using the `root_agent`'s LLM. This simplifies configuration by providing a sensible default.
PiperOrigin-RevId: 816038579
The class is now named `LlmEventSummarizer` to better reflect that its primary function is to use an LLM to summarize events. The docstring has been updated to clarify that this class is responsible *only* for the LLM-based summarization of a given set of events, while the logic for determining *when* and *which* events form the sliding window is handled by an external component, such as an ADK Runner.
PiperOrigin-RevId: 815976264
The docstrings for `compaction_range` and `compacted_content` are updated to reflect that compaction is based on timestamp ranges rather than sequence IDs, and to use consistent terminology ("compacted" instead of "summarized").
PiperOrigin-RevId: 808770610
Provide a more efficient way to compact LLM context for better agentic performance.
* `app`: the top level abstraction for an ADK application. It contains an root agent, and plugins.
* `content_strategy`: the abstraction for selecting the contents for LLM request.
* `compaction_strategy`: the abstraction for compacting the events.
* Added `sequence_id` and `summary_range` in event class.
PiperOrigin-RevId: 808634224
Corrected `CountInvocationPlugin` to be a class reference and added `ContextFilterPlugin` to limit the number of tool invocations kept in the context to 3.
PiperOrigin-RevId: 808591608
This commit introduces a new ContextFilterPlugin which allows for filtering the LlmRequest contents before they are sent to the LLM. This helps in managing and potentially reducing the size of the LLM context.
The plugin provides two primary filtering mechanisms:
num_invocations_to_keep: Keeps only the specified number of the most recent user-model invocations. An invocation is defined as one or more user messages followed by a model response.
custom_filter: Allows for a user-defined callable to be applied to the contents for more flexible filtering.
Unit tests have been added to cover the different filtering scenarios, including:
Filtering by the last N invocations.
Filtering using a custom function.
Combining both filtering methods.
Handling cases with multiple user turns in a single invocation.
Ensuring no filtering occurs when options are not provided.
Gracefully handling exceptions from custom filter functions."
For example, when num_of_innovacations=2:
-----------------------------------------------------------
Contents:
{"parts":[{"text":"9"}],"role":"user"}
{"parts":[{"text":"I am sorry, I cannot fulfill this request. I need more information on what you would like me to do. I can roll a die or check prime numbers.\n"}],"role":"model"}
{"parts":[{"text":"1"}],"role":"user"}
{"parts":[{"text":"I am sorry, I cannot fulfill this request. I need more information on what you would like me to do. I can roll a die or check prime numbers.\n"}],"role":"model"}
{"parts":[{"text":"10"}],"role":"user"}
-----------------------------------------------------------
PiperOrigin-RevId: 808355316
Update the bug report issue template to request minimal reproducible examples, error/stacktrace, clarify OS options, and include questions about LiteLLM usage and specific model details.
PiperOrigin-RevId: 806435953
The new test verifies that `output_audio_transcription` and `input_audio_transcription` attributes are unique to each `RunConfig` instance, preventing unintended side effects from modifying one instance.
PiperOrigin-RevId: 806405671
Switched the active model from `gemini-live-2.5-flash-preview` (for AI Studio) to `gemini-2.0-flash-live-preview-04-09` (for Vertex).
PiperOrigin-RevId: 806348640
The old live/bidi agents are using a cache to store context/history during agent transfer etc. As we have added support for session for live/bidi, we are now migrating the context/history cache to it. This improves scalability, efficiency and maintainability.
It introduces several changes:
* AudioTranscriber support is removed as now we are using native transcription from models.
* Transcription is returned as input_transcription/output_transcription fields and no longer as contents.
* We will return a new event with artifact references of file type of audio/pcm.(in addition to existing audio response event. So the users of this api need to do proper filtering here.)
PiperOrigin-RevId: 805997675
Changes references from `gemini-1.5-flash` and `gemini-1.5-pro` to `gemini-2.5-flash` and `gemini-2.5-pro` in docstrings, default values, sample agents, and tests.
PiperOrigin-RevId: 805536434
These tests verify that `ValueError` is raised when `Runner` is initialized without providing either an `app` instance or both `app_name` and `agent`.
PiperOrigin-RevId: 805427256
- Added `tests/unittests/apps/test_apps.py` with basic tests for `App` initialization.
- Modified `tests/unittests/test_runners.py` to include a test that verifies `Runner` raises a `ValueError` when both `app` and `app_name` are provided during initialization.
PiperOrigin-RevId: 803556826
The convention:
- If some fields(like plugin) are defined both at root_agent and app, then a error will be raised.
- app code should be located within agent.py.
- an instance named app should be created
PiperOrigin-RevId: 803155804
The convention:
- If some fields(like plugin) are defined both at root_agent and app, then a error will be raised.
- app code should be located within agent.py.
- an instance named app should be created
PiperOrigin-RevId: 801252329
The convention:
- If some fields(like plugin) are defined both at root_agent and app, then a error will be raised.
- app code should be located within agent.py.
- an instance named app should be created
PiperOrigin-RevId: 801084463
The transcription change breaks the multi-agent transfer during live/bidi.
Updates `GeminiLlmConnection` to populate the `content` field of `LlmResponse` with `types.Content` and `types.Part` objects for both input and output transcriptions, instead of using dedicated transcription fields. Also removes a debug print from `audio_cache_manager.py`.
the transcription is not fully ready to be used yet so roll back the transcription change.
PiperOrigin-RevId: 799851950
For Vertex model backend, we send response back. This doesn't work for streaming tools that the return type is AsyncGenerator. So the fix here is to ignore the return type when it's AsyncGenerator.
We can't distinguish streaming vs non-streaming tool with AsyncGenerator though as LiveRequestQueue is optional in streaming tool.
Adds an `ignore_response` option to `build_function_declaration` to skip including the return type in the function declaration. This is enabled for tools that return `AsyncGenerator`, as the model does not yet support understanding these return types, while streaming tools can still handle them. Also, removes redundant return statements in `_get_mandatory_params`.
PiperOrigin-RevId: 794392846
Previous implementation doesn't pass the actual handle to server. Now we cache the handle and pass it over when reconnection happens.
To enable:
run_config = RunConfig(
session_resumption=types.SessionResumptionConfig(transparent=True)
)
PiperOrigin-RevId: 791308462
This commit adds support for the session resumption configuration in the run_config.
The SessionResumptionConfig is added to RunConfig to allow the user to set up a configuration for session resumption(only transparent mode for now).
There are two modes of session resumption: manual and transparent. In manual mode, you have to manually bookkeeping the session information and restarts the session which is tricky to do right now. In transparent mode, the server does the bookkeeping for you and no hassle on ADK side. For now, the transparent mode should be enough.
Also, added the relevant unit tests to check that every possible configuration is set properly and the run_config is correctly populated.
This is needed for supporting the new session resumption feature.
PiperOrigin-RevId: 786549455
Now the LangchainTool can wrap:
* Langchain StructuredTool (sync and async).
* Langchain @Tool (sync and async).
This enhance the flexibility for user and enables async functionalities.
PiperOrigin-RevId: 784728061
**Highlights:**
- **Callback Chaining:** Now supports a list of before and after callbacks, as in the non-live version. Callbacks are executed in order, stopping when one returns a response.
- **Unit Tests:** Added new unit tests for before and after callbacks, async and sync versions, callback chains and mixed callbacks.
- **Sample Agent:** Introduced a new example agent with multiple callbacks showing various behaviors: audit, security, validation, and enhancement. This provides practical usage examples.
PiperOrigin-RevId: 783884562
This change adds activity start and end signals to the LiveRequestQueue,
allowing clients to manually control the start and end of user input in
streaming sessions when automatic voice activity detection is disabled.
The LiveRequestQueue allows users to send messages to the model with the following semantics:
- `content`: sends turn-by-turn content.
- `blob`: sends a media blob for realtime streaming (e.g., audio).
- `activity_start`: indicates the beginning of an activity.
- `activity_end`: indicates the end of an activity.
- `close`: closes the connection.
GeminiLLMConnection has been updated to send the new activity signals to the backend.
This change is a necessary to support clients (e.g. voice assistants) that do not want to use automatic voice activity detection. In this case, the client will be responsible to send the `activity_start` signal when the user starts talking, and `activity_end` when the user finishes talking.
To test the change:
run_config = RunConfig(
realtime_input_config=types.RealtimeInputConfig(
automatic_activity_detection=types.AutomaticActivityDetection(
disabled=True,
),
)
)
import threading # Add this import
def thread_target():
# Define the async operations to run in the background.
async def background_task():
live_request_queue.send_activity_start()
# live_request_queue.send_content(
# content=types.Content(
# role='user',
# parts=[types.Part.from_text(text="hi, what's the time?")],
# )
# )
await asyncio.sleep(3)
live_request_queue.send_activity_end()
PiperOrigin-RevId: 783882447
This commit includes a number of new tests for live streaming with function calls. These tests cover various scenarios:
- Single function calls
- Multiple function calls
- Parallel function calls
- Function calls with errors
- Synchronous function calls
- Simple streaming tools
- Video streaming tools
- Stopping a streaming tool
- Multiple streaming tools simultaneously
The tests use mock models and custom runners to simulate the interaction between the agent, model, and tools. They verify that function calls are correctly generated, executed, and that the expected data is returned.
PiperOrigin-RevId: 781318483
This PR extends the existing Streaming tests to consider new configurations that have been added, including:
- `output_audio_transcription`
- `input_audio_transcription`
- `realtime_input_config`
- `enable_affective_dialog`
- `proactivity`
These configurations are tested individually and also combined, to cover all the possibilities, increasing the testing coverage and ensuring everything is working as expected.
In addition, the new configuration values are also validated to be sure they are properly initialized.
PiperOrigin-RevId: 780178334
This explains the high-level architecture and philosophy of the ADK project.
It can also be feed into LLMs for vibe-coding.
PiperOrigin-RevId: 780178125
fix: Use `inspect.signature()` instead of `typing.get_type_hints()` to find the LiveRequestQueue.
Motivation: `typing.get_type_hints()` may raise errors in complex scenarios where type annotation is not available.
Add live_bidi_streaming_tools_agent example to show how to use the live streaming feature.
PiperOrigin-RevId: 780160921
Adds a description for docstring and comments in the AGENTS.md and adds a section for Versioning that describes how ADK follows Semantic Versioning 2.0.0
PiperOrigin-RevId: 778667398
Bug: When a model emits a stream of tokens, it sometimes emits a final chunk of whitespace or no content. The agent was trying to parse that content into JSON, causing a validation error.
Fix: If a model is expected to return JSON and the last streamed token is empty/whitespace, the agent will no longer try to parse it, and exit gracefully.
New unit tests confirm the scenario and the fix.
PiperOrigin-RevId: 777609415
This example include multi agents:
- Root agent.
- Sub agent for Rolling Dice.
- Sub agent for checking primes.
Added README.md to demonstrate how to use it.
PiperOrigin-RevId: 777599625
Merge https://github.com/google/adk-python/pull/1250
The `Args:` section in the docstring of the `cli_deploy_agent_engine` function was causing formatting issues in the help output, with line breaks not being rendered correctly.
This commit removes the redundant `Args:` section from the docstring. The help text for options is already comprehensively covered by the `help` attributes in the `@click.option` decorators, and `click` automatically lists the command's arguments.
This change ensures that the help output for
`adk deploy agent_engine --help` is clean, readable, and correctly formatted, relying on `click`'s standard help generation mechanisms.
After the fix:
(adk_test234) (base) hangfeilin@Hangfeis-MBP adk-python % adk deploy agent_engine --help
Usage: adk deploy agent_engine [OPTIONS] AGENT
Deploys an agent to Agent Engine.
Args: agent (str): Required. The path to the agent to be deloyed.
Example:
adk deploy agent_engine --project=[project] --region=[region] --staging_bucket=[staging_bucket] path/to/my_agent
Options:
--project TEXT Required. Google Cloud project to deploy the agent.
--region TEXT Required. Google Cloud region to deploy the agent.
--staging_bucket TEXT Required. GCS bucket for staging the deployment artifacts.
--trace_to_cloud Optional. Whether to enable Cloud Trace for Agent Engine.
--adk_app TEXT Optional. Python file for defining the ADK application (default: a file named agent_engine_app.py)
--temp_folder TEXT Optional. Temp folder for the generated Agent Engine source files. If the folder already exists, its
contents will be removed. (default: a timestamped folder in the system temp directory).
--env_file TEXT Optional. The filepath to the `.env` file for environment variables. (default: the `.env` file in
the `agent` directory, if any.)
--requirements_file TEXT Optional. The filepath to the `requirements.txt` file to use. (default: the `requirements.txt` file
in the `agent` directory, if any.)
--help Show this message and exit.
COPYBARA_INTEGRATE_REVIEW=https://github.com/google/adk-python/pull/1250 from google:fix-1191-agent-engine-help d2d0e89ed1af6ace11abe4e3bed89335dbcf9be5
PiperOrigin-RevId: 769182740
--
618add7d297cbe26f26f45aa01b39c3d086a13e8 by Hangfei Lin <hangfei@google.com>:
doc: Create readme.md
--
5ba85d653cac11b2858ee5d53d175c1c16d933ec by Hangfei Lin <hangfei@google.com>:
doc: Update CONTRIBUTING.md
--
02606a34babba6a660886a073332979fb2b12fc3 by Wei Sun (Jack) <weisun@google.com>:
Rename readme.md to README.md
--
08a38bd5727bf554f6fb043c73623d367e9b138e by Hangfei Lin <hangfei@google.com>:
Update README.md
--
92e7ee6d498dfce35f1c6df44c1ec0f86ae5d513 by Wei Sun (Jack) <weisun@google.com>:
Update README.md
--
69c3a44e6946b3541746ded43dae8a70d47d538f by Wei Sun (Jack) <weisun@google.com>:
Update README.md
--
9de40783990ae9935463b37225c186e91c93025d by Wei Sun (Jack) <weisun@google.com>:
Update README.md
--
0f8a011ddc5b22ba8361ce7b34413a34cfcf15ba by Wei Sun (Jack) <weisun@google.com>:
Update README.md
COPYBARA_INTEGRATE_REVIEW=https://github.com/google/adk-python/pull/1058 from google:hangfei-patch-2 81eacedef69468fa554312a187880ccf4c30c559
PiperOrigin-RevId: 766234622
--
76f579c8f75e28c79e322dc60c9dca56ac96d0fa by Hangfei Lin <hangfei@google.com>:
doc: Update README.md
COPYBARA_INTEGRATE_REVIEW=https://github.com/google/adk-python/pull/901 from google:hangfei-patch-3 6c5e264c8134e8eb164114992f3d8f2f2efe83fc
PiperOrigin-RevId: 762603928
--
cd8c580cbd4a2dc1c49c690ea81c4111860aa52c by Hangfei Lin <hangfei@google.com>:
doc: Update README.md
COPYBARA_INTEGRATE_REVIEW=https://github.com/google/adk-python/pull/842 from google:hangfei-patch-2 aa9842a43d8c464a8999beb2f6f5d491aa63ef6c
PiperOrigin-RevId: 761722188
Add a `task_completed` function to the agent so when a model finished the task, it can send a signal and the program knows it can go to next agent.
This cl include:
* Implements the `_run_live_impl` in `sequential_agent` so it can handle live case.
* Add an example for sequential agent.
* Improve error message for unimplemented _run_live_impl in other agents.
Note:
1. Compared to non-live case, live agents process a continuous streams of audio
or video, so it doesn't have a native way to tell if it's finished and should pass
to next agent or not. So we introduce a task_compelted() function so the
model can call this function to signal that it's finished the task and we
can move on to next agent.
2. live agents doesn't seems to be very useful or natural in parallel or loop agents so we don't implement it for now. If there is user demand, we can implement it easily using similar approach.
PiperOrigin-RevId: 758315430
The old implementation:
1. We only started transcription at the beginning of agent transferring.
2. The transcription service we used is not as good / fast as the model/native transcription.
In the current implementation, the live agent will rely on the llm's transcription, instead of our transcription when llm support audio transcription in the input. And in that case, the live agent won't use our own audio transcriber. This reduces the latency from 5secs to 2 secs during agent transferring. It also improves the transcription quality.
When the llm doesn't support audio transcription, we still use our audio transcriber to transcribe audio input.
PiperOrigin-RevId: 758296647
