Skip to content

Refactor max step summary #4461

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.

Sign up for GitHub

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

Open
PauiC wants to merge 4 commits into
base: feature/3.4-release-prep
Choose a base branch
from
Open

Refactor max step summary #4461

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
122e77d
init: refactor
PauiC Nov 24, 2025
15a8995
fix: duplicated try catch
PauiC Nov 24, 2025
65f702c
fix: recover file
PauiC Nov 24, 2025
cc66d76
fix: bug
PauiC Nov 25, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
100 changes: 100 additions & 0 deletions ml-algorithms/src/main/java/org/opensearch/ml/engine/algorithms/agent/AgentUtils.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -75,16 +75,24 @@
import org.opensearch.core.xcontent.NamedXContentRegistry;
import org.opensearch.core.xcontent.XContentParser;
import org.opensearch.index.IndexNotFoundException;
import org.opensearch.ml.common.FunctionName;
import org.opensearch.ml.common.agent.LLMSpec;
import org.opensearch.ml.common.agent.MLAgent;
import org.opensearch.ml.common.agent.MLMemorySpec;
import org.opensearch.ml.common.agent.MLToolSpec;
import org.opensearch.ml.common.connector.Connector;
import org.opensearch.ml.common.connector.HttpConnector;
import org.opensearch.ml.common.connector.McpConnector;
import org.opensearch.ml.common.connector.McpStreamableHttpConnector;
import org.opensearch.ml.common.dataset.remote.RemoteInferenceInputDataSet;
import org.opensearch.ml.common.input.remote.RemoteInferenceMLInput;
import org.opensearch.ml.common.output.model.ModelTensor;
import org.opensearch.ml.common.output.model.ModelTensorOutput;
import org.opensearch.ml.common.output.model.ModelTensors;
import org.opensearch.ml.common.spi.tools.Tool;
import org.opensearch.ml.common.transport.MLTaskResponse;
import org.opensearch.ml.common.transport.prediction.MLPredictionTaskAction;
import org.opensearch.ml.common.transport.prediction.MLPredictionTaskRequest;
import org.opensearch.ml.common.utils.StringUtils;
import org.opensearch.ml.engine.MLEngineClassLoader;
import org.opensearch.ml.engine.algorithms.remote.McpConnectorExecutor;
Expand Down Expand Up @@ -1178,4 +1186,96 @@ private static Map<String, String> parseStringMapParameter(String rawValue, Stri
return null;
}
}

public static String extractSummaryFromResponse(MLTaskResponse response, Map<String, String> parameters) {
ModelTensorOutput output = (ModelTensorOutput) response.getOutput();
if (output == null || output.getMlModelOutputs() == null || output.getMlModelOutputs().isEmpty()) {
throw new IllegalStateException("No model output available in response");
}

ModelTensors tensors = output.getMlModelOutputs().getFirst();
if (tensors == null || tensors.getMlModelTensors() == null || tensors.getMlModelTensors().isEmpty()) {
throw new IllegalStateException("No model tensors available in response");
}

ModelTensor tensor = tensors.getMlModelTensors().getFirst();
if (tensor.getResult() != null) {
return tensor.getResult().trim();
}

if (tensor.getDataAsMap() == null) {
throw new IllegalStateException("No data map available in tensor");
}

Map<String, ?> dataMap = tensor.getDataAsMap();
if (dataMap.containsKey(RESPONSE_FIELD)) {
return String.valueOf(dataMap.get(RESPONSE_FIELD)).trim();
}

if (dataMap.containsKey("output")) {
try {
Object outputObj = JsonPath.read(dataMap, parameters.get(LLM_RESPONSE_FILTER));
if (outputObj != null) {
return String.valueOf(outputObj).trim();
}
} catch (PathNotFoundException e) {
throw new IllegalStateException("Failed to extract output using filter path", e);
}
}

throw new IllegalStateException("No result/response field found. Available fields: " + dataMap.keySet());
}

public static void generateMaxStepSummary(
Client client,
LLMSpec llmSpec,
String promptContent,
String systemPrompt,
Map<String, String> allParams,
String tenantId,
ActionListener<String> listener
) {
if (promptContent == null || promptContent.isEmpty()) {
listener.onFailure(new IllegalArgumentException("Prompt content cannot be null or empty"));
return;
}

try {
Map<String, String> summaryParams = new HashMap<>();
if (llmSpec.getParameters() != null) {
summaryParams.putAll(llmSpec.getParameters());
}
summaryParams.putAll(allParams);
summaryParams.put(MLPlanExecuteAndReflectAgentRunner.PROMPT_FIELD, promptContent);
summaryParams.put(MLPlanExecuteAndReflectAgentRunner.SYSTEM_PROMPT_FIELD, systemPrompt);

MLPredictionTaskRequest request = new MLPredictionTaskRequest(
llmSpec.getModelId(),
RemoteInferenceMLInput
.builder()
.algorithm(FunctionName.REMOTE)
.inputDataset(RemoteInferenceInputDataSet.builder().parameters(summaryParams).build())
.build(),
null,
tenantId
);

client.execute(MLPredictionTaskAction.INSTANCE, request, ActionListener.wrap(response -> {
try {
String summary = extractSummaryFromResponse(response, summaryParams);
if (summary == null || summary.isEmpty()) {
listener.onFailure(new RuntimeException("Empty LLM summary response"));
return;
}
listener.onResponse(summary);
} catch (IllegalStateException e) {
log.error("Failed to extract summary, using fallback", e);
listener.onFailure(e);
}
}, listener::onFailure));
} catch (IllegalStateException e) {
log.error("Failed to generate summary", e);
listener.onFailure(e);
}
}
}
95 changes: 18 additions & 77 deletions ml-algorithms/src/main/java/org/opensearch/ml/engine/algorithms/agent/MLChatAgentRunner.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19,7 +19,6 @@
import static org.opensearch.ml.common.utils.ToolUtils.parseResponse;
import static org.opensearch.ml.engine.algorithms.agent.AgentUtils.DISABLE_TRACE;
import static org.opensearch.ml.engine.algorithms.agent.AgentUtils.INTERACTIONS_PREFIX;
import static org.opensearch.ml.engine.algorithms.agent.AgentUtils.LLM_RESPONSE_FILTER;
import static org.opensearch.ml.engine.algorithms.agent.AgentUtils.PROMPT_CHAT_HISTORY_PREFIX;
import static org.opensearch.ml.engine.algorithms.agent.AgentUtils.PROMPT_PREFIX;
import static org.opensearch.ml.engine.algorithms.agent.AgentUtils.PROMPT_SUFFIX;
Expand Down Expand Up @@ -67,25 +66,20 @@
import org.opensearch.core.action.ActionListener;
import org.opensearch.core.common.Strings;
import org.opensearch.core.xcontent.NamedXContentRegistry;
import org.opensearch.ml.common.FunctionName;
import org.opensearch.ml.common.MLMemoryType;
import org.opensearch.ml.common.agent.LLMSpec;
import org.opensearch.ml.common.agent.MLAgent;
import org.opensearch.ml.common.agent.MLToolSpec;
import org.opensearch.ml.common.contextmanager.ContextManagerContext;
import org.opensearch.ml.common.conversation.Interaction;
import org.opensearch.ml.common.dataset.remote.RemoteInferenceInputDataSet;
import org.opensearch.ml.common.hooks.HookRegistry;
import org.opensearch.ml.common.input.remote.RemoteInferenceMLInput;
import org.opensearch.ml.common.memory.Memory;
import org.opensearch.ml.common.memory.Message;
import org.opensearch.ml.common.output.model.ModelTensor;
import org.opensearch.ml.common.output.model.ModelTensorOutput;
import org.opensearch.ml.common.output.model.ModelTensors;
import org.opensearch.ml.common.spi.tools.Tool;
import org.opensearch.ml.common.transport.MLTaskResponse;
import org.opensearch.ml.common.transport.prediction.MLPredictionTaskAction;
import org.opensearch.ml.common.transport.prediction.MLPredictionTaskRequest;
import org.opensearch.ml.common.utils.StringUtils;
import org.opensearch.ml.engine.agents.AgentContextUtil;
import org.opensearch.ml.engine.encryptor.Encryptor;
Expand All @@ -100,7 +94,6 @@

import com.google.common.annotations.VisibleForTesting;
import com.google.gson.reflect.TypeToken;
import com.jayway.jsonpath.JsonPath;

import lombok.Data;
import lombok.NoArgsConstructor;
Expand Down Expand Up @@ -502,7 +495,11 @@ private void runReAct(
);
toolParams.put(TENANT_ID_FIELD, tenantId);
lastToolParams.clear();
lastToolParams.putAll(toolParams);
toolParams
.entrySet()
.stream()
.filter(e -> e.getValue() != null)
.forEach(e -> lastToolParams.put(e.getKey(), e.getValue()));
runTool(
tools,
toolSpecMap,
Expand Down Expand Up @@ -1128,12 +1125,6 @@ void generateLLMSummary(
}

try {
Map<String, String> summaryParams = new HashMap<>();
if (llmSpec.getParameters() != null) {
summaryParams.putAll(llmSpec.getParameters());
}
summaryParams.putAll(parameter);

// Convert ModelTensors to strings before joining, skip session/interaction IDs
List<String> stepStrings = new ArrayList<>();
for (ModelTensors tensor : stepsSummary) {
Expand All @@ -1152,74 +1143,24 @@ void generateLLMSummary(
}
}
}
String steps = String.format(Locale.ROOT, "Question: %s\n\nCompleted Steps:\n%s", question, String.join("\n", stepStrings));
summaryParams.put(PROMPT, steps);
summaryParams.put(SYSTEM_PROMPT_FIELD, MAX_STEP_SUMMARY_CHAT_AGENT_SYSTEM_PROMPT);

ActionRequest request = new MLPredictionTaskRequest(
llmSpec.getModelId(),
RemoteInferenceMLInput
.builder()
.algorithm(FunctionName.REMOTE)
.inputDataset(RemoteInferenceInputDataSet.builder().parameters(summaryParams).build())
.build(),
null,
tenantId
);
client.execute(MLPredictionTaskAction.INSTANCE, request, ActionListener.wrap(response -> {
String summary = extractSummaryFromResponse(response, summaryParams);
if (summary == null) {
listener.onFailure(new RuntimeException("Empty or invalid LLM summary response"));
return;
}
listener.onResponse(summary);
}, listener::onFailure));
String promptContent = String
.format(Locale.ROOT, "Question: %s\n\nCompleted Steps:\n%s", question, String.join("\n", stepStrings));

AgentUtils
.generateMaxStepSummary(
client,
llmSpec,
promptContent,
MAX_STEP_SUMMARY_CHAT_AGENT_SYSTEM_PROMPT,
parameter,
tenantId,
listener
);
} catch (Exception e) {
listener.onFailure(e);
}
}

public String extractSummaryFromResponse(MLTaskResponse response, Map<String, String> params) {
try {
ModelTensorOutput output = (ModelTensorOutput) response.getOutput();
if (output == null || output.getMlModelOutputs() == null || output.getMlModelOutputs().isEmpty()) {
return null;
}

ModelTensors tensors = output.getMlModelOutputs().getFirst();
if (tensors == null || tensors.getMlModelTensors() == null || tensors.getMlModelTensors().isEmpty()) {
return null;
}

ModelTensor tensor = tensors.getMlModelTensors().getFirst();
if (tensor.getResult() != null) {
return tensor.getResult().trim();
}

if (tensor.getDataAsMap() == null) {
return null;
}

Map<String, ?> dataMap = tensor.getDataAsMap();
if (dataMap.containsKey("response")) {
return String.valueOf(dataMap.get("response")).trim();
}

if (dataMap.containsKey("output")) {
Object outputObj = JsonPath.read(dataMap, params.get(LLM_RESPONSE_FILTER));
if (outputObj != null) {
return String.valueOf(outputObj).trim();
}
}

log.error("Summary generate error. No result/response field found. Available fields: {}", dataMap.keySet());
return null;
} catch (Exception e) {
log.error("Failed to extract summary from response", e);
throw new RuntimeException("Failed to extract summary from response", e);
}
}

private void saveMessage(
Memory memory,
String question,
Expand Down
Loading