[Security solution] Bedrock streaming and token tracking (#170815)

package.json

@@ -842,6 +842,8 @@
     "@opentelemetry/semantic-conventions": "^1.4.0",
     "@reduxjs/toolkit": "1.7.2",
     "@slack/webhook": "^5.0.4",
+    "@smithy/eventstream-codec": "^2.0.12",
+    "@smithy/util-utf8": "^2.0.0",
     "@tanstack/react-query": "^4.29.12",
     "@tanstack/react-query-devtools": "^4.29.12",
     "@turf/along": "6.0.1",

x-pack/packages/kbn-elastic-assistant/impl/assistant/api.test.tsx

@@ -75,19 +75,20 @@ describe('API tests', () => {
       expect(mockHttp.fetch).toHaveBeenCalledWith(
         '/internal/elastic_assistant/actions/connector/foo/_execute',
         {
-          body: '{"params":{"subActionParams":{"model":"gpt-4","messages":[{"role":"user","content":"This is a test"}],"n":1,"stop":null,"temperature":0.2},"subAction":"invokeAI"},"assistantLangChain":false}',
-          headers: { 'Content-Type': 'application/json' },
+          body: '{"params":{"subActionParams":{"model":"gpt-4","messages":[{"role":"user","content":"This is a test"}],"n":1,"stop":null,"temperature":0.2},"subAction":"invokeStream"},"assistantLangChain":false}',
           method: 'POST',
+          asResponse: true,
+          rawResponse: true,
           signal: undefined,
         }
       );
     });
 
-    it('returns API_ERROR when the response status is not ok', async () => {
+    it('returns API_ERROR when the response status is error and langchain is on', async () => {
       (mockHttp.fetch as jest.Mock).mockResolvedValue({ status: 'error' });
 
       const testProps: FetchConnectorExecuteAction = {
-        assistantLangChain: false,
+        assistantLangChain: true,
         http: mockHttp,
         messages,
         apiConfig,
@@ -98,10 +99,50 @@ describe('API tests', () => {
       expect(result).toEqual({ response: API_ERROR, isStream: false, isError: true });
     });
 
+    it('returns API_ERROR when the response status is error, langchain is off, and response is not a reader', async () => {
+      (mockHttp.fetch as jest.Mock).mockResolvedValue({ status: 'error' });
+
+      const testProps: FetchConnectorExecuteAction = {
+        assistantLangChain: false,
+        http: mockHttp,
+        messages,
+        apiConfig,
+      };
+
+      const result = await fetchConnectorExecuteAction(testProps);
+
+      expect(result).toEqual({
+        response: `${API_ERROR}\n\nCould not get reader from response`,
+        isStream: false,
+        isError: true,
+      });
+    });
+
+    it('returns API_ERROR when the response is error, langchain is off, and response is a reader', async () => {
+      const mockReader = jest.fn();
+      (mockHttp.fetch as jest.Mock).mockRejectedValue({
+        response: { body: { getReader: jest.fn().mockImplementation(() => mockReader) } },
+      });
+      const testProps: FetchConnectorExecuteAction = {
+        assistantLangChain: false,
+        http: mockHttp,
+        messages,
+        apiConfig,
+      };
+
+      const result = await fetchConnectorExecuteAction(testProps);
+
+      expect(result).toEqual({
+        response: mockReader,
+        isStream: true,
+        isError: true,
+      });
+    });
+
     it('returns API_ERROR when there are no choices', async () => {
       (mockHttp.fetch as jest.Mock).mockResolvedValue({ status: 'ok', data: '' });
       const testProps: FetchConnectorExecuteAction = {
-        assistantLangChain: false,
+        assistantLangChain: true,
         http: mockHttp,
         messages,
         apiConfig,

x-pack/packages/kbn-elastic-assistant/impl/assistant/api.tsx

@@ -54,17 +54,16 @@ export const fetchConnectorExecuteAction = async ({
           messages: outboundMessages,
         };
 
-  // TODO: Remove in part 2 of streaming work for security solution
+  // TODO: Remove in part 3 of streaming work for security solution
   // tracked here: https://github.com/elastic/security-team/issues/7363
-  // My "Feature Flag", turn to false before merging
-  // In part 2 I will make enhancements to invokeAI to make it work with both openA, but to keep it to a Security Soltuion only review on this PR,
-  // I'm calling the stream action directly
-  const isStream = !assistantLangChain && false;
+  // In part 3 I will make enhancements to langchain to introduce streaming
+  // Once implemented, invokeAI can be removed
+  const isStream = !assistantLangChain;
   const requestBody = isStream
     ? {
         params: {
           subActionParams: body,
-          subAction: 'stream',
+          subAction: 'invokeStream',
         },
         assistantLangChain,
       }
@@ -105,7 +104,7 @@ export const fetchConnectorExecuteAction = async ({
       };
     }
 
-    // TODO: Remove in part 2 of streaming work for security solution
+    // TODO: Remove in part 3 of streaming work for security solution
     // tracked here: https://github.com/elastic/security-team/issues/7363
     // This is a temporary code to support the non-streaming API
     const response = await http.fetch<{
@@ -140,10 +139,19 @@ export const fetchConnectorExecuteAction = async ({
       isStream: false,
     };
   } catch (error) {
+    const reader = error?.response?.body?.getReader();
+
+    if (!reader) {
+      return {
+        response: `${API_ERROR}\n\n${error?.body?.message ?? error?.message}`,
+        isError: true,
+        isStream: false,
+      };
+    }
     return {
-      response: `${API_ERROR}\n\n${error?.body?.message ?? error?.message}`,
+      response: reader,
+      isStream: true,
       isError: true,
-      isStream: false,
     };
   }
 };

x-pack/plugins/actions/server/lib/action_executor.ts

@@ -8,12 +8,13 @@
 import type { PublicMethodsOf } from '@kbn/utility-types';
 import { Logger, KibanaRequest } from '@kbn/core/server';
 import { cloneDeep } from 'lodash';
+import { set } from '@kbn/safer-lodash-set';
 import { withSpan } from '@kbn/apm-utils';
 import { EncryptedSavedObjectsClient } from '@kbn/encrypted-saved-objects-plugin/server';
 import { SpacesServiceStart } from '@kbn/spaces-plugin/server';
 import { IEventLogger, SAVED_OBJECT_REL_PRIMARY } from '@kbn/event-log-plugin/server';
 import { SecurityPluginStart } from '@kbn/security-plugin/server';
-import { PassThrough, Readable } from 'stream';
+import { getGenAiTokenTracking, shouldTrackGenAiToken } from './gen_ai_token_tracking';
 import {
   validateParams,
   validateConfig,
@@ -38,7 +39,6 @@ import { RelatedSavedObjects } from './related_saved_objects';
 import { createActionEventLogRecordObject } from './create_action_event_log_record_object';
 import { ActionExecutionError, ActionExecutionErrorReason } from './errors/action_execution_error';
 import type { ActionsAuthorization } from '../authorization/actions_authorization';
-import { getTokenCountFromOpenAIStream } from './get_token_count_from_openai_stream';
 
 // 1,000,000 nanoseconds in 1 millisecond
 const Millis2Nanos = 1000 * 1000;
@@ -328,55 +328,33 @@ export class ActionExecutor {
           eventLogger.logEvent(event);
         }
 
-        // start openai extension
-        // add event.kibana.action.execution.openai to event log when OpenAI Connector is executed
-        if (result.status === 'ok' && actionTypeId === '.gen-ai') {
-          const data = result.data as unknown as {
-            usage: { prompt_tokens?: number; completion_tokens?: number; total_tokens?: number };
-          };
-          event.kibana = event.kibana || {};
-          event.kibana.action = event.kibana.action || {};
-          event.kibana = {
-            ...event.kibana,
-            action: {
-              ...event.kibana.action,
-              execution: {
-                ...event.kibana.action.execution,
-                gen_ai: {
-                  usage: {
-                    total_tokens: data.usage?.total_tokens,
-                    prompt_tokens: data.usage?.prompt_tokens,
-                    completion_tokens: data.usage?.completion_tokens,
-                  },
-                },
-              },
-            },
-          };
-
-          if (result.data instanceof Readable) {
-            getTokenCountFromOpenAIStream({
-              responseStream: result.data.pipe(new PassThrough()),
-              body: (validatedParams as { subActionParams: { body: string } }).subActionParams.body,
+        // start genai extension
+        if (result.status === 'ok' && shouldTrackGenAiToken(actionTypeId)) {
+          getGenAiTokenTracking({
+            actionTypeId,
+            logger,
+            result,
+            validatedParams,
+          })
+            .then((tokenTracking) => {
+              if (tokenTracking != null) {
+                set(event, 'kibana.action.execution.gen_ai.usage', {
+                  total_tokens: tokenTracking.total_tokens,
+                  prompt_tokens: tokenTracking.prompt_tokens,
+                  completion_tokens: tokenTracking.completion_tokens,
+                });
+              }
             })
-              .then(({ total, prompt, completion }) => {
-                event.kibana!.action!.execution!.gen_ai!.usage = {
-                  total_tokens: total,
-                  prompt_tokens: prompt,
-                  completion_tokens: completion,
-                };
-              })
-              .catch((err) => {
-                logger.error('Failed to calculate tokens from streaming response');
-                logger.error(err);
-              })
-              .finally(() => {
-                completeEventLogging();
-              });
-
-            return resultWithoutError;
-          }
+            .catch((err) => {
+              logger.error('Failed to calculate tokens from streaming response');
+              logger.error(err);
+            })
+            .finally(() => {
+              completeEventLogging();
+            });
+          return resultWithoutError;
         }
-        // end openai extension
+        // end genai extension
 
         completeEventLogging();