Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Allow llmQuestion to be optional when llmMessages is used. (Issue #3… #3072

Merged
merged 2 commits into from
Oct 9, 2024

Conversation

austintlee
Copy link
Collaborator

…067)

Description

Remove the check on llmQuestion being present in RAG request parameters.

Related Issues

#3067

Check List

  • [x ] New functionality includes testing.
  • New functionality has been documented.
  • API changes companion pull request created.
  • [x ] Commits are signed per the DCO using --signoff.
  • Public documentation issue/PR created.

By submitting this pull request, I confirm that my contribution is made under the terms of the Apache 2.0 license.
For more information on following Developer Certificate of Origin and signing off your commits, please check here.

@@ -185,7 +187,7 @@ public GenerativeQAParameters(
public GenerativeQAParameters(StreamInput input) throws IOException {
this.conversationId = input.readOptionalString();
this.llmModel = input.readOptionalString();
this.llmQuestion = input.readString();
this.llmQuestion = input.readOptionalString();
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Hi @austintlee this does not quite make sense to me, why you only make llmQuestion into optional while keeping llmMessage as mandatory field, if you are trying to make user to choose one of them?

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@b4sjoo I do make them both optional in the first constructor:

public GenerativeQAParameters(
        String conversationId,
        String llmModel,
        String llmQuestion,
        String systemPrompt,
        String userInstructions,
        Integer contextSize,
        Integer interactionSize,
        Integer timeout,
        String llmResponseField,
        List<MessageBlock> llmMessages
    ) {
        this.conversationId = conversationId;
        this.llmModel = llmModel;

        Preconditions
            .checkArgument(
                !(Strings.isNullOrEmpty(llmQuestion) && (llmMessages == null || llmMessages.isEmpty())),
                "At least one of " + LLM_QUESTION + " or " + LLM_MESSAGES_FIELD + " must be provided."
            );
        this.llmQuestion = llmQuestion;
        this.systemPrompt = systemPrompt;
        this.userInstructions = userInstructions;
        this.contextSize = (contextSize == null) ? SIZE_NULL_VALUE : contextSize;
        this.interactionSize = (interactionSize == null) ? SIZE_NULL_VALUE : interactionSize;
        this.timeout = (timeout == null) ? SIZE_NULL_VALUE : timeout;
        this.llmResponseField = llmResponseField;
        if (llmMessages != null) {
            this.llmMessages.addAll(llmMessages);
        }
    }

But internally, llmMessages is never null and by default is an empty array.

So, when we write out to StreamOut, we don't need to do a null check:

public void writeTo(StreamOutput out) throws IOException {
        out.writeOptionalString(conversationId);
        out.writeOptionalString(llmModel);
        out.writeOptionalString(llmQuestion);
        out.writeOptionalString(systemPrompt);
        out.writeOptionalString(userInstructions);
        out.writeInt(contextSize);
        out.writeInt(interactionSize);
        out.writeInt(timeout);
        out.writeOptionalString(llmResponseField);
        out.writeList(llmMessages);
    }

Which is why I always expect it to be present (at least as an empty list) when I read it back:

public GenerativeQAParameters(StreamInput input) throws IOException {
        this.conversationId = input.readOptionalString();
        this.llmModel = input.readOptionalString();
        this.llmQuestion = input.readOptionalString();
        this.systemPrompt = input.readOptionalString();
        this.userInstructions = input.readOptionalString();
        this.contextSize = input.readInt();
        this.interactionSize = input.readInt();
        this.timeout = input.readInt();
        this.llmResponseField = input.readOptionalString();
        this.llmMessages.addAll(input.readList(MessageBlock::new));
    }

Is this an incorrect assumption? Does the StreamInput constructor need to consider llmMessages not being present in input?

You can also take a look at stream roundtrip test cases I have in GenerativeQAParamExtBuilderTests.

Copy link
Collaborator

@b4sjoo b4sjoo Oct 9, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It's just because I saw a null check above and then you make here mandatory makes me confused. I think your answer makes sense to me, that llmMessage should never be null due to an empty list created

Copy link
Collaborator

@b4sjoo b4sjoo Oct 9, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

BTW, changing a readString() into readOptionalString() could potentially introduce a bwc issue when we have a mixed cluster

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Going from required to optional should be OK, but not the other way around. How do we test it?

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@pyek-bot is currently testing it, he should have a result by tomorrow. Basically we create a lower version cluster (e.g. 2.16) with dedicated master node, then we upgrade the data node to the current version to test. After this we perform the test again, but we upgrade master this time. Does this make sense?

Copy link
Contributor

@pyek-bot pyek-bot Oct 9, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I have tested this scenario. It seems to work fine when all nodes are eventually upgraded to 2.17.

  • When only the data node is upgraded, the NPE [https://github.com/[BUG] RAG processor throws null pointer exception #2983] comes into play since the master cannot serialize the data and send it to the data node.
  • When only the master node is upgraded, the data node cannot de-serialize due to new format and throws unexpected byte error.

However, when both are upgraded it works as expected with both llmQuestion and llmMessages.

@@ -359,7 +359,7 @@ public class RestMLRAGSearchProcessorIT extends MLCommonsRestTestCase {
+ " \"ext\": {\n"
+ " \"generative_qa_parameters\": {\n"
+ " \"llm_model\": \"%s\",\n"
+ " \"llm_question\": \"%s\",\n"
// + " \"llm_question\": \"%s\",\n"
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

why not removing this line?

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Will do.

@@ -378,7 +378,7 @@ public class RestMLRAGSearchProcessorIT extends MLCommonsRestTestCase {
+ " \"ext\": {\n"
+ " \"generative_qa_parameters\": {\n"
+ " \"llm_model\": \"%s\",\n"
+ " \"llm_question\": \"%s\",\n"
// + " \"llm_question\": \"%s\",\n"
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

same here?

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Will remove it.

@austintlee
Copy link
Collaborator Author

REPRODUCE WITH: ./gradlew ':opensearch-ml-plugin:integTest' --tests "org.opensearch.ml.rest.RestMLInferenceSearchResponseProcessorIT.testMLInferenceProcessorLocalModel" -Dtests.seed=E6490143B0860730 -Dtests.security.manager=false -Dtests.locale=sw-TZ -Dtests.timezone=Asia/Aden -Druntime.java=21
RestMLInferenceSearchResponseProcessorIT > testMLInferenceProcessorLocalModel STANDARD_ERROR
    REPRODUCE WITH: ./gradlew ':opensearch-ml-plugin:integTest' --tests "org.opensearch.ml.rest.RestMLInferenceSearchResponseProcessorIT.testMLInferenceProcessorLocalModel" -Dtests.seed=E6490143B0860730 -Dtests.security.manager=false -Dtests.locale=sw-TZ -Dtests.timezone=Asia/Aden -Druntime.java=21

RestMLInferenceSearchResponseProcessorIT > testMLInferenceProcessorLocalModel FAILED
    org.opensearch.client.ResponseException: method [POST], host [http://127.0.0.1:32939/], URI [/_plugins/_ml/models/null/_deploy], status line [HTTP/1.1 404 Not Found]
    {"error":{"root_cause":[{"type":"status_exception","reason":"Failed to find model"}],"type":"status_exception","reason":"Failed to find model"},"status":404}

Not related to my change.

Signed-off-by: Austin Lee <[email protected]>
@b4sjoo
Copy link
Collaborator

b4sjoo commented Oct 9, 2024

REPRODUCE WITH: ./gradlew ':opensearch-ml-plugin:integTest' --tests "org.opensearch.ml.rest.RestMLInferenceSearchResponseProcessorIT.testMLInferenceProcessorLocalModel" -Dtests.seed=E6490143B0860730 -Dtests.security.manager=false -Dtests.locale=sw-TZ -Dtests.timezone=Asia/Aden -Druntime.java=21
RestMLInferenceSearchResponseProcessorIT > testMLInferenceProcessorLocalModel STANDARD_ERROR
    REPRODUCE WITH: ./gradlew ':opensearch-ml-plugin:integTest' --tests "org.opensearch.ml.rest.RestMLInferenceSearchResponseProcessorIT.testMLInferenceProcessorLocalModel" -Dtests.seed=E6490143B0860730 -Dtests.security.manager=false -Dtests.locale=sw-TZ -Dtests.timezone=Asia/Aden -Druntime.java=21

RestMLInferenceSearchResponseProcessorIT > testMLInferenceProcessorLocalModel FAILED
    org.opensearch.client.ResponseException: method [POST], host [http://127.0.0.1:32939/], URI [/_plugins/_ml/models/null/_deploy], status line [HTTP/1.1 404 Not Found]
    {"error":{"root_cause":[{"type":"status_exception","reason":"Failed to find model"}],"type":"status_exception","reason":"Failed to find model"},"status":404}

Not related to my change.

Yeah this is a know issue. I think it should be flaky

@@ -723,8 +720,12 @@ public void testBM25WithBedrock() throws Exception {
public void testBM25WithBedrockConverse() throws Exception {
// Skip test if key is null
if (AWS_ACCESS_KEY_ID == null) {
System.out.println("Skipping testBM25WithBedrockConverse because AWS_ACCESS_KEY_ID is null");
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

minor: can we use log ?

Copy link
Collaborator

@ylwu-amzn ylwu-amzn left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM. left minor comment for test code. But I'm ok to fix later

@b4sjoo b4sjoo merged commit 48d275d into opensearch-project:main Oct 9, 2024
7 of 8 checks passed
opensearch-trigger-bot bot pushed a commit that referenced this pull request Oct 9, 2024
…#3072)

* Allow llmQuestion to be optional when llmMessages is used.  (Issue #3067)

Signed-off-by: Austin Lee <[email protected]>

* Remove unused lines.

Signed-off-by: Austin Lee <[email protected]>

---------

Signed-off-by: Austin Lee <[email protected]>
(cherry picked from commit 48d275d)
dhrubo-os pushed a commit that referenced this pull request Oct 9, 2024
…#3072) (#3082)

* Allow llmQuestion to be optional when llmMessages is used.  (Issue #3067)

Signed-off-by: Austin Lee <[email protected]>

* Remove unused lines.

Signed-off-by: Austin Lee <[email protected]>

---------

Signed-off-by: Austin Lee <[email protected]>
(cherry picked from commit 48d275d)

Co-authored-by: Austin Lee <[email protected]>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Projects
None yet
Development

Successfully merging this pull request may close these issues.

5 participants