feat(OpenAI Node): Use v2 assistants API and add support for memory (#9406)

Signed-off-by: Oleg Ivaniv <me@olegivaniv.com>

packages/@n8n/nodes-langchain/nodes/vendors/OpenAi/actions/assistant/create.operation.ts

@@ -133,6 +133,24 @@ const properties: INodeProperties[] = [
 		type: 'collection',
 		default: {},
 		options: [
+			{
+				displayName: 'Output Randomness (Temperature)',
+				name: 'temperature',
+				default: 1,
+				typeOptions: { maxValue: 1, minValue: 0, numberPrecision: 1 },
+				description:
+					'Controls randomness: Lowering results in less random completions. As the temperature approaches zero, the model will become deterministic and repetitive. We generally recommend altering this or temperature but not both.',
+				type: 'number',
+			},
+			{
+				displayName: 'Output Randomness (Top P)',
+				name: 'topP',
+				default: 1,
+				typeOptions: { maxValue: 1, minValue: 0, numberPrecision: 1 },
+				description:
+					'An alternative to sampling with temperature, controls diversity via nucleus sampling: 0.5 means half of all likelihood-weighted options are considered. We generally recommend altering this or temperature but not both.',
+				type: 'number',
+			},
 			{
 				displayName: 'Fail if Assistant Already Exists',
 				name: 'failIfExists',
@@ -176,7 +194,7 @@ export async function execute(this: IExecuteFunctions, i: number): Promise<INode
 		do {
 			const response = (await apiRequest.call(this, 'GET', '/assistants', {
 				headers: {
-					'OpenAI-Beta': 'assistants=v1',
+					'OpenAI-Beta': 'assistants=v2',
 				},
 				qs: {
 					limit: 100,
@@ -219,7 +237,6 @@ export async function execute(this: IExecuteFunctions, i: number): Promise<INode
 		name,
 		description: assistantDescription,
 		instructions,
-		file_ids,
 	};
 
 	const tools = [];
@@ -228,12 +245,28 @@ export async function execute(this: IExecuteFunctions, i: number): Promise<INode
 		tools.push({
 			type: 'code_interpreter',
 		});
+		body.tool_resources = {
+			...((body.tool_resources as object) ?? {}),
+			code_interpreter: {
+				file_ids,
+			},
+		};
 	}
 
 	if (knowledgeRetrieval) {
 		tools.push({
-			type: 'retrieval',
+			type: 'file_search',
 		});
+		body.tool_resources = {
+			...((body.tool_resources as object) ?? {}),
+			file_search: {
+				vector_stores: [
+					{
+						file_ids,
+					},
+				],
+			},
+		};
 	}
 
 	if (tools.length) {
@@ -243,7 +276,7 @@ export async function execute(this: IExecuteFunctions, i: number): Promise<INode
 	const response = await apiRequest.call(this, 'POST', '/assistants', {
 		body,
 		headers: {
-			'OpenAI-Beta': 'assistants=v1',
+			'OpenAI-Beta': 'assistants=v2',
 		},
 	});
 

packages/@n8n/nodes-langchain/nodes/vendors/OpenAi/actions/assistant/deleteAssistant.operation.ts

@@ -19,7 +19,7 @@ export async function execute(this: IExecuteFunctions, i: number): Promise<INode
 
 	const response = await apiRequest.call(this, 'DELETE', `/assistants/${assistantId}`, {
 		headers: {
-			'OpenAI-Beta': 'assistants=v1',
+			'OpenAI-Beta': 'assistants=v2',
 		},
 	});
 

packages/@n8n/nodes-langchain/nodes/vendors/OpenAi/actions/assistant/list.operation.ts

@@ -30,7 +30,7 @@ export async function execute(this: IExecuteFunctions, i: number): Promise<INode
 	do {
 		const response = await apiRequest.call(this, 'GET', '/assistants', {
 			headers: {
-				'OpenAI-Beta': 'assistants=v1',
+				'OpenAI-Beta': 'assistants=v2',
 			},
 			qs: {
 				limit: 100,

packages/@n8n/nodes-langchain/nodes/vendors/OpenAi/actions/assistant/message.operation.ts

@@ -4,9 +4,17 @@ import { OpenAIAssistantRunnable } from 'langchain/experimental/openai_assistant
 import type { OpenAIToolType } from 'langchain/dist/experimental/openai_assistant/schema';
 import { OpenAI as OpenAIClient } from 'openai';
 
-import { NodeOperationError, updateDisplayOptions } from 'n8n-workflow';
+import { NodeConnectionType, NodeOperationError, updateDisplayOptions } from 'n8n-workflow';
-import type { IExecuteFunctions, INodeExecutionData, INodeProperties } from 'n8n-workflow';
+import type {
+	IDataObject,
+	IExecuteFunctions,
+	INodeExecutionData,
+	INodeProperties,
+} from 'n8n-workflow';
 
+import type { BufferWindowMemory } from 'langchain/memory';
+import omit from 'lodash/omit';
+import type { BaseMessage } from '@langchain/core/messages';
 import { formatToOpenAIAssistantTool } from '../../helpers/utils';
 import { assistantRLC } from '../descriptions';
 
@@ -110,6 +118,12 @@ const displayOptions = {
 };
 
 export const description = updateDisplayOptions(displayOptions, properties);
+const mapChatMessageToThreadMessage = (
+	message: BaseMessage,
+): OpenAIClient.Beta.Threads.ThreadCreateParams.Message => ({
+	role: message._getType() === 'ai' ? 'assistant' : 'user',
+	content: message.content.toString(),
+});
 
 export async function execute(this: IExecuteFunctions, i: number): Promise<INodeExecutionData[]> {
 	const credentials = await this.getCredentials('openAiApi');
@@ -182,11 +196,47 @@ export async function execute(this: IExecuteFunctions, i: number): Promise<INode
 		tools: tools ?? [],
 	});
 
-	const response = await agentExecutor.withConfig(getTracingConfig(this)).invoke({
+	const memory = (await this.getInputConnectionData(NodeConnectionType.AiMemory, 0)) as
+		| BufferWindowMemory
+		| undefined;
+
+	const chainValues: IDataObject = {
 		content: input,
 		signal: this.getExecutionCancelSignal(),
 		timeout: options.timeout ?? 10000,
-	});
+	};
+	let thread: OpenAIClient.Beta.Threads.Thread;
+	if (memory) {
+		const chatMessages = await memory.chatHistory.getMessages();
+
+		// Construct a new thread from the chat history to map the memory
+		if (chatMessages.length) {
+			const first32Messages = chatMessages.slice(0, 32);
+			// There is a undocumented limit of 32 messages per thread when creating a thread with messages
+			const mappedMessages: OpenAIClient.Beta.Threads.ThreadCreateParams.Message[] =
+				first32Messages.map(mapChatMessageToThreadMessage);
+
+			thread = await client.beta.threads.create({ messages: mappedMessages });
+			const overLimitMessages = chatMessages.slice(32).map(mapChatMessageToThreadMessage);
+
+			// Send the remaining messages that exceed the limit of 32 sequentially
+			for (const message of overLimitMessages) {
+				await client.beta.threads.messages.create(thread.id, message);
+			}
+
+			chainValues.threadId = thread.id;
+		}
+	}
+
+	const response = await agentExecutor.withConfig(getTracingConfig(this)).invoke(chainValues);
+	if (memory) {
+		await memory.saveContext({ input }, { output: response.output });
+
+		if (response.threadId && response.runId) {
+			const threadRun = await client.beta.threads.runs.retrieve(response.threadId, response.runId);
+			response.usage = threadRun.usage;
+		}
+	}
 
 	if (
 		options.preserveOriginalTools !== false &&
@@ -197,6 +247,6 @@ export async function execute(this: IExecuteFunctions, i: number): Promise<INode
 			tools: assistantTools,
 		});
 	}
-
+	const filteredResponse = omit(response, ['signal', 'timeout']);
-	return [{ json: response, pairedItem: { item: i } }];
+	return [{ json: filteredResponse, pairedItem: { item: i } }];
 }

packages/@n8n/nodes-langchain/nodes/vendors/OpenAi/actions/assistant/update.operation.ts

@@ -84,6 +84,25 @@ const properties: INodeProperties[] = [
 				default: false,
 				description: 'Whether to remove all custom tools (functions) from the assistant',
 			},
+
+			{
+				displayName: 'Output Randomness (Temperature)',
+				name: 'temperature',
+				default: 1,
+				typeOptions: { maxValue: 1, minValue: 0, numberPrecision: 1 },
+				description:
+					'Controls randomness: Lowering results in less random completions. As the temperature approaches zero, the model will become deterministic and repetitive. We generally recommend altering this or temperature but not both.',
+				type: 'number',
+			},
+			{
+				displayName: 'Output Randomness (Top P)',
+				name: 'topP',
+				default: 1,
+				typeOptions: { maxValue: 1, minValue: 0, numberPrecision: 1 },
+				description:
+					'An alternative to sampling with temperature, controls diversity via nucleus sampling: 0.5 means half of all likelihood-weighted options are considered. We generally recommend altering this or temperature but not both.',
+				type: 'number',
+			},
 		],
 	},
 ];
@@ -109,6 +128,8 @@ export async function execute(this: IExecuteFunctions, i: number): Promise<INode
 		knowledgeRetrieval,
 		file_ids,
 		removeCustomTools,
+		temperature,
+		topP,
 	} = options;
 
 	const assistantDescription = options.description as string;
@@ -128,7 +149,19 @@ export async function execute(this: IExecuteFunctions, i: number): Promise<INode
 			);
 		}
 
-		body.file_ids = files;
+		body.tool_resources = {
+			...((body.tool_resources as object) ?? {}),
+			code_interpreter: {
+				file_ids,
+			},
+			file_search: {
+				vector_stores: [
+					{
+						file_ids,
+					},
+				],
+			},
+		};
 	}
 
 	if (modelId) {
@@ -147,11 +180,19 @@ export async function execute(this: IExecuteFunctions, i: number): Promise<INode
 		body.instructions = instructions;
 	}
 
+	if (temperature) {
+		body.temperature = temperature;
+	}
+
+	if (topP) {
+		body.topP = topP;
+	}
+
 	let tools =
 		((
 			await apiRequest.call(this, 'GET', `/assistants/${assistantId}`, {
 				headers: {
-					'OpenAI-Beta': 'assistants=v1',
+					'OpenAI-Beta': 'assistants=v2',
 				},
 			})
 		).tools as IDataObject[]) || [];
@@ -166,14 +207,14 @@ export async function execute(this: IExecuteFunctions, i: number): Promise<INode
 		tools = tools.filter((tool) => tool.type !== 'code_interpreter');
 	}
 
-	if (knowledgeRetrieval && !tools.find((tool) => tool.type === 'retrieval')) {
+	if (knowledgeRetrieval && !tools.find((tool) => tool.type === 'file_search')) {
 		tools.push({
-			type: 'retrieval',
+			type: 'file_search',
 		});
 	}
 
-	if (knowledgeRetrieval === false && tools.find((tool) => tool.type === 'retrieval')) {
+	if (knowledgeRetrieval === false && tools.find((tool) => tool.type === 'file_search')) {
-		tools = tools.filter((tool) => tool.type !== 'retrieval');
+		tools = tools.filter((tool) => tool.type !== 'file_search');
 	}
 
 	if (removeCustomTools) {
@@ -185,7 +226,7 @@ export async function execute(this: IExecuteFunctions, i: number): Promise<INode
 	const response = await apiRequest.call(this, 'POST', `/assistants/${assistantId}`, {
 		body,
 		headers: {
-			'OpenAI-Beta': 'assistants=v1',
+			'OpenAI-Beta': 'assistants=v2',
 		},
 	});
 

packages/@n8n/nodes-langchain/nodes/vendors/OpenAi/actions/versionDescription.ts

@@ -46,6 +46,7 @@ const configureNodeInputs = (resource: string, operation: string, hideTools: str
 	if (resource === 'assistant' && operation === 'message') {
 		return [
 			{ type: NodeConnectionType.Main },
+			{ type: NodeConnectionType.AiMemory, displayName: 'Memory', maxConnections: 1 },
 			{ type: NodeConnectionType.AiTool, displayName: 'Tools' },
 		];
 	}

packages/@n8n/nodes-langchain/nodes/vendors/OpenAi/methods/listSearch.ts

@@ -78,7 +78,7 @@ export async function assistantSearch(
 ): Promise<INodeListSearchResult> {
 	const { data, has_more, last_id } = await apiRequest.call(this, 'GET', '/assistants', {
 		headers: {
-			'OpenAI-Beta': 'assistants=v1',
+			'OpenAI-Beta': 'assistants=v2',
 		},
 		qs: {
 			limit: 100,

packages/@n8n/nodes-langchain/nodes/vendors/OpenAi/test/OpenAi.node.test.ts

@@ -84,13 +84,24 @@ describe('OpenAi, Assistant resource', () => {
 		expect(transport.apiRequest).toHaveBeenCalledWith('POST', '/assistants', {
 			body: {
 				description: 'description',
-				file_ids: [],
 				instructions: 'some instructions',
 				model: 'gpt-model',
 				name: 'name',
-				tools: [{ type: 'code_interpreter' }, { type: 'retrieval' }],
+				tool_resources: {
+					code_interpreter: {
+						file_ids: [],
 					},
-			headers: { 'OpenAI-Beta': 'assistants=v1' },
+					file_search: {
+						vector_stores: [
+							{
+								file_ids: [],
+							},
+						],
+					},
+				},
+				tools: [{ type: 'code_interpreter' }, { type: 'file_search' }],
+			},
+			headers: { 'OpenAI-Beta': 'assistants=v2' },
 		});
 	});
 
@@ -124,7 +135,7 @@ describe('OpenAi, Assistant resource', () => {
 		);
 
 		expect(transport.apiRequest).toHaveBeenCalledWith('DELETE', '/assistants/assistant-id', {
-			headers: { 'OpenAI-Beta': 'assistants=v1' },
+			headers: { 'OpenAI-Beta': 'assistants=v2' },
 		});
 	});
 
@@ -185,17 +196,28 @@ describe('OpenAi, Assistant resource', () => {
 
 		expect(transport.apiRequest).toHaveBeenCalledTimes(2);
 		expect(transport.apiRequest).toHaveBeenCalledWith('GET', '/assistants/assistant-id', {
-			headers: { 'OpenAI-Beta': 'assistants=v1' },
+			headers: { 'OpenAI-Beta': 'assistants=v2' },
 		});
 		expect(transport.apiRequest).toHaveBeenCalledWith('POST', '/assistants/assistant-id', {
 			body: {
-				file_ids: [],
 				instructions: 'some instructions',
 				model: 'gpt-model',
 				name: 'name',
-				tools: [{ type: 'existing_tool' }, { type: 'code_interpreter' }, { type: 'retrieval' }],
+				tool_resources: {
+					code_interpreter: {
+						file_ids: [],
 					},
-			headers: { 'OpenAI-Beta': 'assistants=v1' },
+					file_search: {
+						vector_stores: [
+							{
+								file_ids: [],
+							},
+						],
+					},
+				},
+				tools: [{ type: 'existing_tool' }, { type: 'code_interpreter' }, { type: 'file_search' }],
+			},
+			headers: { 'OpenAI-Beta': 'assistants=v2' },
 		});
 	});
 });