[8.x] [Inference] Inference CLI client (#214691)

.github/CODEOWNERS

@@ -543,6 +543,7 @@ x-pack/platform/plugins/private/index_lifecycle_management @elastic/kibana-manag
 x-pack/platform/plugins/shared/index_management @elastic/kibana-management
 x-pack/platform/packages/shared/index-management/index_management_shared_types @elastic/kibana-management
 src/platform/test/plugin_functional/plugins/index_patterns @elastic/kibana-data-discovery
+x-pack/platform/packages/shared/kbn-inference-cli @elastic/appex-ai-infra
 x-pack/platform/packages/shared/ai-infra/inference-common @elastic/appex-ai-infra
 x-pack/platform/plugins/shared/inference_endpoint @elastic/ml-ui
 x-pack/platform/packages/shared/kbn-inference-endpoint-ui-common @elastic/response-ops @elastic/appex-ai-infra @elastic/obs-ai-assistant @elastic/security-generative-ai
@@ -568,12 +569,14 @@ src/platform/packages/private/kbn-jest-serializers @elastic/kibana-operations
 src/platform/packages/private/kbn-journeys @elastic/kibana-operations @elastic/appex-qa
 packages/kbn-json-ast @elastic/kibana-operations
 x-pack/platform/packages/private/ml/json_schemas @elastic/ml-ui
+x-pack/solutions/observability/packages/kbn-genai-cli @elastic/obs-knowledge-team
 src/platform/test/health_gateway/plugins/status @elastic/kibana-core
 src/platform/test/plugin_functional/plugins/kbn_sample_panel_action @elastic/appex-sharedux
 src/platform/test/plugin_functional/plugins/kbn_top_nav @elastic/kibana-core
 src/platform/test/plugin_functional/plugins/kbn_tp_custom_visualizations @elastic/kibana-visualizations
 src/platform/test/interpreter_functional/plugins/kbn_tp_run_pipeline @elastic/kibana-core
 x-pack/platform/packages/shared/kbn-key-value-metadata-table @elastic/obs-ux-infra_services-team @elastic/obs-ux-logs-team
+x-pack/platform/packages/shared/kbn-kibana-api-cli @elastic/appex-ai-infra
 x-pack/test/functional_cors/plugins/kibana_cors_test @elastic/kibana-security
 packages/kbn-kibana-manifest-schema @elastic/kibana-operations
 src/platform/plugins/private/kibana_overview @elastic/appex-sharedux

package.json

@@ -615,6 +615,7 @@
     "@kbn/kbn-tp-custom-visualizations-plugin": "link:src/platform/test/plugin_functional/plugins/kbn_tp_custom_visualizations",
     "@kbn/kbn-tp-run-pipeline-plugin": "link:src/platform/test/interpreter_functional/plugins/kbn_tp_run_pipeline",
     "@kbn/key-value-metadata-table": "link:x-pack/platform/packages/shared/kbn-key-value-metadata-table",
+    "@kbn/kibana-api-cli": "link:x-pack/platform/packages/shared/kbn-kibana-api-cli",
     "@kbn/kibana-cors-test-plugin": "link:x-pack/test/functional_cors/plugins/kibana_cors_test",
     "@kbn/kibana-overview-plugin": "link:src/platform/plugins/private/kibana_overview",
     "@kbn/kibana-react-plugin": "link:src/platform/plugins/shared/kibana_react",
@@ -1471,10 +1472,12 @@
     "@kbn/get-repo-files": "link:src/platform/packages/private/kbn-get-repo-files",
     "@kbn/import-locator": "link:packages/kbn-import-locator",
     "@kbn/import-resolver": "link:src/platform/packages/private/kbn-import-resolver",
+    "@kbn/inference-cli": "link:x-pack/platform/packages/shared/kbn-inference-cli",
     "@kbn/inventory-e2e": "link:x-pack/solutions/observability/plugins/inventory/e2e",
     "@kbn/jest-serializers": "link:src/platform/packages/private/kbn-jest-serializers",
     "@kbn/journeys": "link:src/platform/packages/private/kbn-journeys",
     "@kbn/json-ast": "link:packages/kbn-json-ast",
+    "@kbn/kbn-genai-cli": "link:x-pack/solutions/observability/packages/kbn-genai-cli",
     "@kbn/kibana-manifest-schema": "link:packages/kbn-kibana-manifest-schema",
     "@kbn/lint-packages-cli": "link:packages/kbn-lint-packages-cli",
     "@kbn/lint-ts-projects-cli": "link:packages/kbn-lint-ts-projects-cli",

src/platform/packages/shared/kbn-sse-utils-server/src/supertest_to_observable.ts

@@ -36,7 +36,10 @@ export function supertestToObservable<T = any>(response: supertest.Test): Observ
   return new Observable<T>((subscriber) => {
     const parser = createParser({
       onEvent: (event) => {
-        subscriber.next(JSON.parse(event.data));
+        subscriber.next({
+          type: event.event ?? 'event',
+          ...JSON.parse(event.data),
+        });
       },
     });
 

src/platform/packages/shared/kbn-sse-utils/src/errors.ts

@@ -23,6 +23,13 @@ export class ServerSentEventError<
     super(message);
   }
 
+  public get status() {
+    if (typeof this.meta === 'object' && this.meta.status) {
+      return this.meta.status as number;
+    }
+    return undefined;
+  }
+
   toJSON(): ServerSentErrorEvent {
     return {
       type: ServerSentEventType.error,

tsconfig.base.json

@@ -1080,6 +1080,8 @@
       "@kbn/index-management-shared-types/*": ["x-pack/platform/packages/shared/index-management/index_management_shared_types/*"],
       "@kbn/index-patterns-test-plugin": ["src/platform/test/plugin_functional/plugins/index_patterns"],
       "@kbn/index-patterns-test-plugin/*": ["src/platform/test/plugin_functional/plugins/index_patterns/*"],
+      "@kbn/inference-cli": ["x-pack/platform/packages/shared/kbn-inference-cli"],
+      "@kbn/inference-cli/*": ["x-pack/platform/packages/shared/kbn-inference-cli/*"],
       "@kbn/inference-common": ["x-pack/platform/packages/shared/ai-infra/inference-common"],
       "@kbn/inference-common/*": ["x-pack/platform/packages/shared/ai-infra/inference-common/*"],
       "@kbn/inference-endpoint-plugin": ["x-pack/platform/plugins/shared/inference_endpoint"],
@@ -1130,6 +1132,8 @@
       "@kbn/json-ast/*": ["packages/kbn-json-ast/*"],
       "@kbn/json-schemas": ["x-pack/platform/packages/private/ml/json_schemas"],
       "@kbn/json-schemas/*": ["x-pack/platform/packages/private/ml/json_schemas/*"],
+      "@kbn/kbn-genai-cli": ["x-pack/solutions/observability/packages/kbn-genai-cli"],
+      "@kbn/kbn-genai-cli/*": ["x-pack/solutions/observability/packages/kbn-genai-cli/*"],
       "@kbn/kbn-health-gateway-status-plugin": ["src/platform/test/health_gateway/plugins/status"],
       "@kbn/kbn-health-gateway-status-plugin/*": ["src/platform/test/health_gateway/plugins/status/*"],
       "@kbn/kbn-sample-panel-action-plugin": ["src/platform/test/plugin_functional/plugins/kbn_sample_panel_action"],
@@ -1142,6 +1146,8 @@
       "@kbn/kbn-tp-run-pipeline-plugin/*": ["src/platform/test/interpreter_functional/plugins/kbn_tp_run_pipeline/*"],
       "@kbn/key-value-metadata-table": ["x-pack/platform/packages/shared/kbn-key-value-metadata-table"],
       "@kbn/key-value-metadata-table/*": ["x-pack/platform/packages/shared/kbn-key-value-metadata-table/*"],
+      "@kbn/kibana-api-cli": ["x-pack/platform/packages/shared/kbn-kibana-api-cli"],
+      "@kbn/kibana-api-cli/*": ["x-pack/platform/packages/shared/kbn-kibana-api-cli/*"],
       "@kbn/kibana-cors-test-plugin": ["x-pack/test/functional_cors/plugins/kibana_cors_test"],
       "@kbn/kibana-cors-test-plugin/*": ["x-pack/test/functional_cors/plugins/kibana_cors_test/*"],
       "@kbn/kibana-manifest-schema": ["packages/kbn-kibana-manifest-schema"],

x-pack/platform/packages/shared/ai-infra/inference-common/index.ts

@@ -120,3 +120,5 @@ export {
   InferenceEndpointProvider,
   elasticModelIds,
 } from './src/inference_endpoints';
+
+export { Tokenizer } from './src/utils/tokenizer';

x-pack/platform/packages/shared/ai-infra/inference-common/src/chat_complete/events.ts

@@ -22,16 +22,19 @@ export enum ChatCompletionEventType {
  * the whole text content and potential tool calls of the response.
  */
 export type ChatCompletionMessageEvent<TToolOptions extends ToolOptions = ToolOptions> =
-  InferenceTaskEventBase<ChatCompletionEventType.ChatCompletionMessage> & {
-    /**
-     * The text content of the LLM response.
-     */
-    content: string;
-    /**
-     * The eventual tool calls performed by the LLM.
-     */
-    toolCalls: ToolCallsOf<TToolOptions>['toolCalls'];
-  };
+  InferenceTaskEventBase<
+    ChatCompletionEventType.ChatCompletionMessage,
+    {
+      /**
+       * The text content of the LLM response.
+       */
+      content: string;
+      /**
+       * The eventual tool calls performed by the LLM.
+       */
+      toolCalls: ToolCallsOf<TToolOptions>['toolCalls'];
+    }
+  >;
 
 /**
  * Represent a partial tool call present in a chunk event.
@@ -64,8 +67,9 @@ export interface ChatCompletionChunkToolCall {
  * Chunk event, containing a fragment of the total content,
  * and potentially chunks of tool calls.
  */
-export type ChatCompletionChunkEvent =
-  InferenceTaskEventBase<ChatCompletionEventType.ChatCompletionChunk> & {
+export type ChatCompletionChunkEvent = InferenceTaskEventBase<
+  ChatCompletionEventType.ChatCompletionChunk,
+  {
     /**
      * The content chunk
      */
@@ -74,7 +78,8 @@ export type ChatCompletionChunkEvent =
      * The tool call chunks
      */
     tool_calls: ChatCompletionChunkToolCall[];
-  };
+  }
+>;
 
 /**
  * Token count structure for the chatComplete API.
@@ -98,13 +103,15 @@ export interface ChatCompletionTokenCount {
  * Token count event, send only once, usually (but not necessarily)
  * before the message event
  */
-export type ChatCompletionTokenCountEvent =
-  InferenceTaskEventBase<ChatCompletionEventType.ChatCompletionTokenCount> & {
+export type ChatCompletionTokenCountEvent = InferenceTaskEventBase<
+  ChatCompletionEventType.ChatCompletionTokenCount,
+  {
     /**
      * The token count structure
      */
     tokens: ChatCompletionTokenCount;
-  };
+  }
+>;
 
 /**
  * Events emitted from the {@link ChatCompleteResponse} observable

x-pack/platform/packages/shared/ai-infra/inference-common/src/errors.ts

@@ -5,6 +5,7 @@
  * 2.0.
  */
 
+import { ServerSentEventError } from '@kbn/sse-utils';
 import { InferenceTaskEventBase, InferenceTaskEventType } from './inference_task';
 
 /**
@@ -17,44 +18,22 @@ export enum InferenceTaskErrorCode {
   abortedError = 'requestAborted',
 }
 
-/**
- * Base class for all inference API errors.
- */
-export class InferenceTaskError<
+const InferenceTaskError = ServerSentEventError;
+type InferenceTaskError<
   TCode extends string,
   TMeta extends Record<string, any> | undefined
-> extends Error {
-  constructor(public code: TCode, message: string, public meta: TMeta) {
-    super(message);
-  }
-
-  public get status() {
-    if (typeof this.meta === 'object' && this.meta.status) {
-      return this.meta.status as number;
-    }
-    return undefined;
-  }
-
-  toJSON(): InferenceTaskErrorEvent {
-    return {
-      type: InferenceTaskEventType.error,
-      error: {
-        code: this.code,
-        message: this.message,
-        meta: this.meta,
-      },
+> = ServerSentEventError<TCode, TMeta>;
+
+export type InferenceTaskErrorEvent = InferenceTaskEventBase<
+  InferenceTaskEventType.error,
+  {
+    error: {
+      code: string;
+      message: string;
+      meta?: Record<string, any>;
     };
   }
-}
-
-export type InferenceTaskErrorEvent = InferenceTaskEventBase<InferenceTaskEventType.error> & {
-  error: {
-    code: string;
-    message: string;
-    meta?: Record<string, any>;
-  };
-};
-
+>;
 /**
  * Inference error thrown when an unexpected internal error occurs while handling the request.
  */
@@ -162,3 +141,5 @@ export function isInferenceRequestAbortedError(error: unknown): error is Inferen
 export function isInferenceProviderError(error: unknown): error is InferenceTaskProviderError {
   return isInferenceError(error) && error.code === InferenceTaskErrorCode.providerError;
 }
+
+export { InferenceTaskError };

x-pack/platform/packages/shared/ai-infra/inference-common/src/inference_task.ts

@@ -4,19 +4,18 @@
  * 2.0; you may not use this file except in compliance with the Elastic License
  * 2.0.
  */
+import type { ServerSentEventBase } from '@kbn/sse-utils';
 
 /**
  * Base interface for all inference events.
  */
-export interface InferenceTaskEventBase<TEventType extends string> {
-  /**
-   * Unique identifier of the event type.
-   */
-  type: TEventType;
-}
+export type InferenceTaskEventBase<
+  TEventType extends string,
+  TData extends Record<string, any>
+> = ServerSentEventBase<TEventType, TData>;
 
 export enum InferenceTaskEventType {
   error = 'error',
 }
 
-export type InferenceTaskEvent = InferenceTaskEventBase<string>;
+export type InferenceTaskEvent = InferenceTaskEventBase<string, Record<string, unknown>>;

x-pack/platform/packages/shared/ai-infra/inference-common/src/output/events.ts

@@ -23,8 +23,9 @@ export type Output = Record<string, any> | undefined | unknown;
 /**
  * Update (chunk) event for the {@link OutputAPI}
  */
-export type OutputUpdateEvent<TId extends string = string> =
-  InferenceTaskEventBase<OutputEventType.OutputUpdate> & {
+export type OutputUpdateEvent<TId extends string = string> = InferenceTaskEventBase<
+  OutputEventType.OutputUpdate,
+  {
     /**
      * The id of the operation, as provided as input
      */
@@ -33,29 +34,33 @@ export type OutputUpdateEvent<TId extends string = string> =
      * The text content of the chunk
      */
     content: string;
-  };
+  }
+>;
 
 /**
  * Completion (complete message) event for the {@link OutputAPI}
  */
 export type OutputCompleteEvent<
   TId extends string = string,
   TOutput extends Output = Output
-> = InferenceTaskEventBase<OutputEventType.OutputComplete> & {
-  /**
-   * The id of the operation, as provided as input
-   */
-  id: TId;
-  /**
-   * The task output, following the schema specified as input
-   */
-  output: TOutput;
-  /**
-   * Potential text content provided by the LLM,
-   * if it was provided in addition to the tool call
-   */
-  content: string;
-};
+> = InferenceTaskEventBase<
+  OutputEventType.OutputComplete,
+  {
+    /**
+     * The id of the operation, as provided as input
+     */
+    id: TId;
+    /**
+     * The task output, following the schema specified as input
+     */
+    output: TOutput;
+    /**
+     * Potential text content provided by the LLM,
+     * if it was provided in addition to the tool call
+     */
+    content: string;
+  }
+>;
 
 /**
  * Events emitted from the {@link OutputEvent}.

x-pack/platform/packages/shared/ai-infra/inference-common/src/utils/tokenizer.ts

@@ -0,0 +1,32 @@
+/*
+ * Copyright Elasticsearch B.V. and/or licensed to Elasticsearch B.V. under one
+ * or more contributor license agreements. Licensed under the Elastic License
+ * 2.0; you may not use this file except in compliance with the Elastic License
+ * 2.0.
+ */
+
+export class Tokenizer {
+  /**
+   * Approximates the number of tokens in a string,
+   * assuming 4 characters per token.
+   */
+  static count(input: string): number {
+    return Math.ceil(input.length / 4);
+  }
+
+  /**
+   * If the text is longer than the amount of tokens,
+   * truncate and mark as truncated.
+   */
+  static truncate(
+    input: string,
+    maxTokens: number
+  ): { tokens: number; truncated: boolean; text: string } {
+    const count = Tokenizer.count(input);
+    if (Tokenizer.count(input) > maxTokens) {
+      const maxChars = maxTokens * 4;
+      return { truncated: true, tokens: count, text: input.slice(0, maxChars) + '... <truncated>' };
+    }
+    return { truncated: false, tokens: count, text: input };
+  }
+}

x-pack/platform/packages/shared/ai-infra/inference-common/tsconfig.json

@@ -16,5 +16,6 @@
     "target/**/*"
   ],
   "kbn_references": [
+    "@kbn/sse-utils",
   ]
 }