Introduce the InferenceChatModel for langchain (#206429)

## Summary

Part of #206710

This PR introduces the `InferenceChatModel` class, which is a langchain
chatModel utilizing the inference APIs (`chatComplete`) under the hood.

Creating instances of `InferenceChatModel` can either be done by
manually importing the class from the new `@kbn/inference-langchain`
package, or by using the new `createChatModel` API exposes from the
inference plugin's start contract.

The main upside of using this chatModel is that the unification and
normalization layers are already being taken care of by the inference
plugin, making sure that the underlying models are being used with the
exact same capabilities. More details on the upsides and reasoning in
the associated issue.

### Usage

Usage is very straightforward

```ts
const chatModel = await inferenceStart.getChatModel({
  request,
  connectorId: myInferenceConnectorId,
  chatModelOptions: {
    temperature: 0.2,
  },
});

// just use it as another langchain chatModel, e.g.
const response = await chatModel.stream('What is Kibana?');
for await (const chunk of response) {
     // do something with the chunk
}
```

### Important

This PR is only adding the implementation, and not wiring it anywhere or
using it in any existing code. This is meant to be done in a later
stage. Merging that implementation first will allow to have distinct PRs
for the integration with search (playground) and security (assistant +
other workflows), with proper testing

---------

Co-authored-by: kibanamachine <[email protected]>
(cherry picked from commit 1c218f9)

# Conflicts:
#	.github/CODEOWNERS

package.json

@@ -581,6 +581,7 @@
     "@kbn/inference-common": "link:x-pack/platform/packages/shared/ai-infra/inference-common",
     "@kbn/inference-endpoint-plugin": "link:x-pack/platform/plugins/shared/inference_endpoint",
     "@kbn/inference-endpoint-ui-common": "link:x-pack/platform/packages/shared/kbn-inference-endpoint-ui-common",
+    "@kbn/inference-langchain": "link:x-pack/platform/packages/shared/ai-infra/inference-langchain",
     "@kbn/inference-plugin": "link:x-pack/platform/plugins/shared/inference",
     "@kbn/inference_integration_flyout": "link:x-pack/platform/packages/private/ml/inference_integration_flyout",
     "@kbn/infra-forge": "link:x-pack/platform/packages/private/kbn-infra-forge",
@@ -1304,7 +1305,8 @@
     "yaml": "^2.5.1",
     "yauzl": "^2.10.0",
     "yazl": "^2.5.1",
-    "zod": "^3.22.3"
+    "zod": "^3.22.3",
+    "zod-to-json-schema": "^3.23.0"
   },
   "devDependencies": {
     "@apidevtools/swagger-parser": "^10.1.1",
@@ -1877,8 +1879,7 @@
     "xml-crypto": "^6.0.0",
     "xmlbuilder": "13.0.2",
     "yargs": "^15.4.1",
-    "yarn-deduplicate": "^6.0.2",
-    "zod-to-json-schema": "^3.23.0"
+    "yarn-deduplicate": "^6.0.2"
   },
   "packageManager": "[email protected]"
 }

src/platform/packages/shared/kbn-router-to-openapispec/src/oas_converter/zod/lib.ts

@@ -8,7 +8,6 @@
  */
 
 import { z, isZod } from '@kbn/zod';
-// eslint-disable-next-line import/no-extraneous-dependencies
 import zodToJsonSchema from 'zod-to-json-schema';
 import type { OpenAPIV3 } from 'openapi-types';
 

tsconfig.base.json

@@ -1080,6 +1080,8 @@
       "@kbn/inference-endpoint-plugin/*": ["x-pack/platform/plugins/shared/inference_endpoint/*"],
       "@kbn/inference-endpoint-ui-common": ["x-pack/platform/packages/shared/kbn-inference-endpoint-ui-common"],
       "@kbn/inference-endpoint-ui-common/*": ["x-pack/platform/packages/shared/kbn-inference-endpoint-ui-common/*"],
+      "@kbn/inference-langchain": ["x-pack/platform/packages/shared/ai-infra/inference-langchain"],
+      "@kbn/inference-langchain/*": ["x-pack/platform/packages/shared/ai-infra/inference-langchain/*"],
       "@kbn/inference-plugin": ["x-pack/platform/plugins/shared/inference"],
       "@kbn/inference-plugin/*": ["x-pack/platform/plugins/shared/inference/*"],
       "@kbn/infra-forge": ["x-pack/platform/packages/private/kbn-infra-forge"],

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

@@ -96,13 +96,17 @@ export {
   isInferenceRequestError,
   isInferenceRequestAbortedError,
 } from './src/errors';
+export { generateFakeToolCallId } from './src/utils';
 export { elasticModelDictionary } from './src/const';
 
 export { truncateList } from './src/truncate_list';
 export {
   InferenceConnectorType,
   isSupportedConnectorType,
   isSupportedConnector,
+  getConnectorDefaultModel,
+  getConnectorProvider,
+  connectorToInference,
   type InferenceConnector,
 } from './src/connectors';
 export {

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

@@ -69,7 +69,7 @@ export type ToolMessage<
   TToolResponse extends Record<string, any> | unknown = Record<string, any> | unknown,
   TToolData extends Record<string, any> | undefined = Record<string, any> | undefined
 > = MessageBase<MessageRole.Tool> & {
-  /*
+  /**
    * The name of the tool called. Used for refining the type of the response.
    */
   name: TName;

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

@@ -0,0 +1,44 @@
+/*
+ * 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.
+ */
+
+import { type InferenceConnector, InferenceConnectorType } from './connectors';
+
+/**
+ * Returns the default model as defined in the connector's config, if available.
+ *
+ * Note: preconfigured connectors only expose their config if their `exposeConfig` flag
+ * is set to true.
+ */
+export const getConnectorDefaultModel = (connector: InferenceConnector): string | undefined => {
+  switch (connector.type) {
+    case InferenceConnectorType.OpenAI:
+    case InferenceConnectorType.Gemini:
+    case InferenceConnectorType.Bedrock:
+      return connector.config?.defaultModel ?? undefined;
+    case InferenceConnectorType.Inference:
+      return connector.config?.providerConfig?.model_id ?? undefined;
+  }
+};
+
+/**
+ * Returns the provider used for the given connector
+ *
+ * Inferred from the type for "legacy" connectors,
+ * and from the provider config field for inference connectors.
+ */
+export const getConnectorProvider = (connector: InferenceConnector): string => {
+  switch (connector.type) {
+    case InferenceConnectorType.OpenAI:
+      return 'openai';
+    case InferenceConnectorType.Gemini:
+      return 'gemini';
+    case InferenceConnectorType.Bedrock:
+      return 'bedrock';
+    case InferenceConnectorType.Inference:
+      return connector.config?.provider ?? 'unknown';
+  }
+};

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

@@ -0,0 +1,31 @@
+/*
+ * 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.
+ */
+
+import { createInferenceRequestError } from '../errors';
+import type { InferenceConnector, RawConnector } from './connectors';
+import { isSupportedConnector } from './is_supported_connector';
+
+/**
+ * Converts an action connector to the internal inference connector format.
+ *
+ * The function will throw if the provided connector is not compatible
+ */
+export const connectorToInference = (connector: RawConnector): InferenceConnector => {
+  if (!isSupportedConnector(connector)) {
+    throw createInferenceRequestError(
+      `Connector '${connector.id}' of type '${connector.actionTypeId}' not recognized as a supported connector`,
+      400
+    );
+  }
+
+  return {
+    connectorId: connector.id,
+    name: connector.name,
+    type: connector.actionTypeId,
+    config: connector.config ?? {},
+  };
+};

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

@@ -0,0 +1,54 @@
+/*
+ * 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.
+ */
+
+/**
+ * The list of connector types that can be used with the inference APIs
+ */
+export enum InferenceConnectorType {
+  OpenAI = '.gen-ai',
+  Bedrock = '.bedrock',
+  Gemini = '.gemini',
+  Inference = '.inference',
+}
+
+export const allSupportedConnectorTypes = Object.values(InferenceConnectorType);
+
+/**
+ * Represents a stack connector that can be used for inference.
+ */
+export interface InferenceConnector {
+  /** the type of the connector, see {@link InferenceConnectorType} */
+  type: InferenceConnectorType;
+  /** the name of the connector */
+  name: string;
+  /** the id of the connector */
+  connectorId: string;
+  /**
+   * configuration (without secrets) of the connector.
+   * the list of properties depends on the connector type (and subtype for inference)
+   */
+  config: Record<string, any>;
+}
+
+/**
+ * Connector types are living in the actions plugin and we can't afford
+ * having dependencies from this package to some mid-level plugin,
+ * so we're just using our own connector mixin type.
+ */
+export interface RawConnector {
+  id: string;
+  actionTypeId: string;
+  name: string;
+  config?: Record<string, any>;
+}
+
+export interface RawInferenceConnector {
+  id: string;
+  actionTypeId: InferenceConnectorType;
+  name: string;
+  config?: Record<string, any>;
+}

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

@@ -0,0 +1,11 @@
+/*
+ * 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 { isSupportedConnectorType, isSupportedConnector } from './is_supported_connector';
+export { connectorToInference } from './connector_to_inference';
+export { getConnectorDefaultModel, getConnectorProvider } from './connector_config';
+export { InferenceConnectorType, type InferenceConnector } from './connectors';

x-pack/platform/packages/shared/ai-infra/inference-common/src/connectors.test.ts → x-pack/platform/packages/shared/ai-infra/inference-common/src/connectors/is_supported_connector.test.ts

@@ -5,13 +5,12 @@
  * 2.0.
  */
 
+import { InferenceConnectorType, RawConnector } from './connectors';
 import {
-  InferenceConnectorType,
   isSupportedConnectorType,
   isSupportedConnector,
-  RawConnector,
   COMPLETION_TASK_TYPE,
-} from './connectors';
+} from './is_supported_connector';
 
 const createRawConnector = (parts: Partial<RawConnector>): RawConnector => {
   return {
@@ -36,8 +35,6 @@ describe('isSupportedConnectorType', () => {
 });
 
 describe('isSupportedConnector', () => {
-  // TODO
-
   it('returns true for OpenAI connectors', () => {
     expect(
       isSupportedConnector(createRawConnector({ actionTypeId: InferenceConnectorType.OpenAI }))

x-pack/platform/packages/shared/ai-infra/inference-common/src/connectors.ts → x-pack/platform/packages/shared/ai-infra/inference-common/src/connectors/is_supported_connector.ts

@@ -5,37 +5,15 @@
  * 2.0.
  */
 
-/**
- * The list of connector types that can be used with the inference APIs
- */
-export enum InferenceConnectorType {
-  OpenAI = '.gen-ai',
-  Bedrock = '.bedrock',
-  Gemini = '.gemini',
-  Inference = '.inference',
-}
+import {
+  InferenceConnectorType,
+  RawInferenceConnector,
+  RawConnector,
+  allSupportedConnectorTypes,
+} from './connectors';
 
 export const COMPLETION_TASK_TYPE = 'chat_completion';
 
-const allSupportedConnectorTypes = Object.values(InferenceConnectorType);
-
-/**
- * Represents a stack connector that can be used for inference.
- */
-export interface InferenceConnector {
-  /** the type of the connector, see {@link InferenceConnectorType} */
-  type: InferenceConnectorType;
-  /** the name of the connector */
-  name: string;
-  /** the id of the connector */
-  connectorId: string;
-  /**
-   * configuration (without secrets) of the connector.
-   * the list of properties depends on the connector type (and subtype for inference)
-   */
-  config: Record<string, any>;
-}
-
 /**
  * Checks if a given connector type is compatible for inference.
  *
@@ -67,22 +45,3 @@ export function isSupportedConnector(connector: RawConnector): connector is RawI
   }
   return true;
 }
-
-/**
- * Connector types are living in the actions plugin and we can't afford
- * having dependencies from this package to some mid-level plugin,
- * so we're just using our own connector mixin type.
- */
-export interface RawConnector {
-  id: string;
-  actionTypeId: string;
-  name: string;
-  config?: Record<string, any>;
-}
-
-interface RawInferenceConnector {
-  id: string;
-  actionTypeId: InferenceConnectorType;
-  name: string;
-  config?: Record<string, any>;
-}

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

@@ -27,6 +27,13 @@ export class InferenceTaskError<
     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,

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

@@ -0,0 +1,8 @@
+/*
+ * 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 { generateFakeToolCallId } from './tool_calls';

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

@@ -0,0 +1,12 @@
+/*
+ * 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.
+ */
+
+import { v4 } from 'uuid';
+
+export function generateFakeToolCallId() {
+  return v4().substr(0, 6);
+}

x-pack/platform/packages/shared/ai-infra/inference-langchain/README.md

@@ -0,0 +1,39 @@
+# @kbn/inference-langchain
+
+This package exposes utilities to use the inference APIs and plugin with langchain
+
+## InferenceChatModel
+
+The inference chat model is a langchain model leveraging the inference APIs under the hood.
+
+The main upside is that the unification and normalization layers are then fully handled
+by the inference plugin. The developer / consumer doesn't even need to know which provider
+is being used under the hood.
+
+The easiest way to create an `InferenceChatModel` is by using the inference APIs:
+
+```ts
+const chatModel = await inferenceStart.getChatModel({
+  request,
+  connectorId: myInferenceConnectorId,
+  chatModelOptions: {
+    temperature: 0.2,
+  },
+});
+
+// just use it as another langchain chatModel
+```
+
+But the chatModel can also be instantiated directly if needed:
+
+```ts
+import { connectorToInference } from '@kbn/inference-common';
+
+const chatModel = new InferenceChatModel({
+  chatComplete: inference.chatComplete,
+  connector: connectorToInference(someInferenceConnector),
+  logger: myPluginLogger,
+});
+
+// just use it as another langchain chatModel
+```

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

@@ -0,0 +1,12 @@
+/*
+ * 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 {
+  InferenceChatModel,
+  type InferenceChatModelParams,
+  type InferenceChatModelCallOptions,
+} from './src/chat_model';