Move the planning_re_act_planner python implementation to java version. #503
Conversation
Summary of ChangesHello @leweii, I'm Gemini Code Assist1! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed! This pull request introduces a comprehensive planning framework for LLM agents in Java, effectively porting a previously existing Python implementation. It establishes a Highlights
Using Gemini Code AssistThe full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips. Invoking Gemini You can request assistance from Gemini at any point by creating a comment using either
Customization To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a Limitations & Feedback Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here. You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension. Footnotes
|
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request successfully ports a Python-based Plan-ReAct planner to Java, introducing a BasePlanner interface and integrating it into the LlmAgent and processing flows. The implementation includes the PlanReActPlanner with its specific prompting and response parsing logic, a BuiltInPlanner for models with native cognitive features, and new flow processors (NLPlanning) to handle the planning lifecycle. The changes are well-structured, but I've identified a few areas for improvement regarding code clarity, robustness, and testing. Specifically, some redundant checks can be removed, passing null to a constructor can be avoided, and a new test is flawed and requires correction to properly validate the intended behavior.
| @Test | ||
| public void builtInPlanner_contentListUnchanged() { | ||
|
|
||
| Content content = Content.fromParts(Part.fromText("LLM response")); | ||
| TestLlm testLlm = createTestLlm(createLlmResponse(content)); | ||
| BuiltInPlanner planner = BuiltInPlanner.buildPlanner(ThinkingConfig.builder().build()); | ||
| LlmAgent agent = createTestAgent(testLlm, planner); | ||
| var invocationContext = com.google.adk.testing.TestUtils.createInvocationContext(agent); | ||
|
|
||
| List<Content> contents = List.of( | ||
| Content.builder() | ||
| .role("user") | ||
| .parts(Part.fromText("Hello")) | ||
| .build(), | ||
| Content.builder() | ||
| .role("model") | ||
| .parts( | ||
| Part.builder().thought(true).text("thinking....").build(), | ||
| Part.fromText("Here is my response") | ||
| ) | ||
| .build(), | ||
| Content.builder() | ||
| .role("user") | ||
| .parts(Part.fromText("Follow up")) | ||
| .build() | ||
| ); | ||
| LlmRequest llmRequest = LlmRequest.builder().contents(contents).build(); | ||
| List<Content> originalContents = List.copyOf(llmRequest.contents()); | ||
|
|
||
| // Act | ||
| agent.runAsync(invocationContext).blockingSubscribe(); | ||
|
|
||
| // Assert | ||
| assertThat(llmRequest.contents()).isEqualTo(originalContents); | ||
| } | ||
| } |
There was a problem hiding this comment.
This test is flawed and does not validate the intended behavior. The LlmRequest object created on line 47 is a local variable that is never passed into the agent's execution flow initiated by agent.runAsync(). The BaseLlmFlow creates its own LlmRequest internally, so the assertion on line 54 will always pass vacuously.
To properly test that the request processing logic does not mutate its input, you should unit test NlPlanningRequestProcessor.processRequest directly. Here is an example of a more effective test:
@Test
public void nlPlanningRequestProcessor_doesNotMutateOriginalRequest() {
// Arrange
var processor = new NLPlanning.NlPlanningRequestProcessor();
BuiltInPlanner planner = BuiltInPlanner.buildPlanner(ThinkingConfig.builder().build());
LlmAgent agent = createTestAgent(createTestLlm(createLlmResponse(Content.fromText(""))), planner);
var invocationContext = com.google.adk.testing.TestUtils.createInvocationContext(agent);
List<Content> contents = List.of(
Content.builder()
.role("model")
.parts(Part.builder().thought(true).text("thinking....").build())
.build()
);
LlmRequest originalRequest = LlmRequest.builder().contents(contents).build();
List<Content> originalContents = List.copyOf(originalRequest.contents());
// Act
RequestProcessor.RequestProcessingResult result = processor.processRequest(invocationContext, originalRequest).blockingGet();
// Assert
// Check that the original request object is not mutated.
assertThat(originalRequest.contents()).isEqualTo(originalContents);
assertThat(originalRequest.contents().get(0).parts().get().get(0).thought()).isTrue();
// Check that the new request has the thoughts removed.
LlmRequest processedRequest = result.updatedRequest();
assertThat(processedRequest.contents().get(0).parts().get().get(0).thought()).isFalse();
}| if (!(context.agent() instanceof LlmAgent)) { | ||
| throw new IllegalArgumentException( | ||
| "Agent in InvocationContext is not an instance of LlmAgent."); | ||
| } |
There was a problem hiding this comment.
This instanceof check for LlmAgent is redundant. The getPlanner method called on line 34 already performs this check and returns an empty Optional if the agent is not an LlmAgent. The subsequent check if (plannerOpt.isEmpty()) on line 35 is sufficient to handle this case. Removing these lines will simplify the code and avoid duplication, as the same pattern is used in NlPlanningResponseProcessor.
| if (!(context.agent() instanceof LlmAgent)) { | ||
| throw new IllegalArgumentException( | ||
| "Agent in InvocationContext is not an instance of LlmAgent."); | ||
| } |
There was a problem hiding this comment.
Similar to the NlPlanningRequestProcessor, this instanceof check is redundant. The getPlanner method on line 79 handles the case where the agent is not an LlmAgent by returning an empty Optional, which is correctly handled by the check on line 80. You can remove these lines to improve code clarity and reduce duplication.
| LlmResponse.Builder responseBuilder = llmResponse.toBuilder(); | ||
|
|
||
| // Process the planning response | ||
| CallbackContext callbackContext = new CallbackContext(context, null); |
There was a problem hiding this comment.
Passing null for the eventActions parameter in the CallbackContext constructor is not ideal, even though the constructor currently handles it. It's better to be explicit to make the code more robust and readable, especially if the constructor's null-handling logic were to change in the future. Please pass an empty EventActions object instead.
| CallbackContext callbackContext = new CallbackContext(context, null); | |
| CallbackContext callbackContext = new CallbackContext(context, EventActions.builder().build()); |
| public class BuiltInPlanner implements BasePlanner { | ||
| private ThinkingConfig cognitiveConfig; | ||
|
|
||
| private BuiltInPlanner() {} |
There was a problem hiding this comment.
This private no-argument constructor appears to be unused. The only way to create an instance of BuiltInPlanner is through the static factory method buildPlanner, which uses the other private constructor. You can remove this unused constructor to simplify the class.
leweii
force-pushed
the
feat/planner-support-oct
branch
from
750d6e5 to
1e0bed3
Compare
1 participant
No description provided.