feat: AI nodes usability fixes + Summarization Chain V2 (#7949)

Fixes:
- Refactor connection snapping when dragging and enable it also for
non-main connection types
- Fix propagation of errors from sub-nodes
- Fix chat scrolling when sending/receiving messages
- Prevent empty chat messages
- Fix sub-node selected styles
- Fix output names text overflow

Usability improvements:
- Auto-add manual chat trigger for agents & chain nodes
- Various labels and description updates
- Make the output parser input optional for Basic LLM Chain
- Summarization Chain V2 with a simplified document loader & text
chunking mode

#### How to test the change:
Example workflow showcasing different operation mode of the new
summarization chain:

[Summarization_V2.json](https://github.com/n8n-io/n8n/files/13599901/Summarization_V2.json)


## Issues fixed
Include links to Github issue or Community forum post or **Linear
ticket**:
> Important in order to close automatically and provide context to
reviewers
-
https://www.notion.so/n8n/David-Langchain-Posthog-notes-7a9294938420403095f4508f1a21d31d
- https://linear.app/n8n/issue/N8N-7070/ux-fixes-batch
- https://linear.app/n8n/issue/N8N-7071/ai-sub-node-bugs


## Review / Merge checklist
- [x] PR title and summary are descriptive. **Remember, the title
automatically goes into the changelog. Use `(no-changelog)` otherwise.**
([conventions](https://github.com/n8n-io/n8n/blob/master/.github/pull_request_title_conventions.md))
- [x] [Docs updated](https://github.com/n8n-io/n8n-docs) or follow-up
ticket created.
- [ ] Tests included.
> A bug is not considered fixed, unless a test is added to prevent it
from happening again. A feature is not complete without tests.
  >
> *(internal)* You can use Slack commands to trigger [e2e
tests](https://www.notion.so/n8n/How-to-use-Test-Instances-d65f49dfc51f441ea44367fb6f67eb0a?pvs=4#a39f9e5ba64a48b58a71d81c837e8227)
or [deploy test
instance](https://www.notion.so/n8n/How-to-use-Test-Instances-d65f49dfc51f441ea44367fb6f67eb0a?pvs=4#f6a177d32bde4b57ae2da0b8e454bfce)
or [deploy early access version on
Cloud](https://www.notion.so/n8n/Cloudbot-3dbe779836004972b7057bc989526998?pvs=4#fef2d36ab02247e1a0f65a74f6fb534e).

---------

Signed-off-by: Oleg Ivaniv <me@olegivaniv.com>
Co-authored-by: Elias Meire <elias@meire.dev>
This commit is contained in:
oleg
2023-12-08 13:42:32 +01:00
committed by GitHub
parent dbd62a4992
commit dcf12867b3
32 changed files with 1235 additions and 436 deletions

View File

@@ -64,7 +64,7 @@ export class OpenAiAssistant implements INodeType {
default: 'existing',
options: [
{
name: 'Create New Assistant',
name: 'Use New Assistant',
value: 'new',
},
{
@@ -94,7 +94,6 @@ export class OpenAiAssistant implements INodeType {
typeOptions: {
rows: 5,
},
required: true,
displayOptions: {
show: {
'/mode': ['new'],
@@ -237,11 +236,28 @@ export class OpenAiAssistant implements INodeType {
value: 'code_interpreter',
},
{
name: 'Retrieval',
name: 'Knowledge Retrieval',
value: 'retrieval',
},
],
},
{
displayName: 'Connect your own custom tools to this node on the canvas',
name: 'noticeTools',
type: 'notice',
default: '',
},
{
displayName:
'Upload files for retrieval using the <a href="https://platform.openai.com/playground" target="_blank">OpenAI website<a/>',
name: 'noticeTools',
type: 'notice',
typeOptions: {
noticeTheme: 'info',
},
displayOptions: { show: { '/nativeTools': ['retrieval'] } },
default: '',
},
{
displayName: 'Options',
name: 'options',

View File

@@ -191,13 +191,33 @@ async function getChain(
return Array.isArray(response) ? response : [response];
}
function getInputs(parameters: IDataObject) {
const hasOutputParser = parameters?.hasOutputParser;
const inputs = [
{ displayName: '', type: NodeConnectionType.Main },
{
displayName: 'Model',
maxConnections: 1,
type: NodeConnectionType.AiLanguageModel,
required: true,
},
];
// If `hasOutputParser` is undefined it must be version 1.1 or earlier so we
// always add the output parser input
if (hasOutputParser === undefined || hasOutputParser === true) {
inputs.push({ displayName: 'Output Parser', type: NodeConnectionType.AiOutputParser });
}
return inputs;
}
export class ChainLlm implements INodeType {
description: INodeTypeDescription = {
displayName: 'Basic LLM Chain',
name: 'chainLlm',
icon: 'fa:link',
group: ['transform'],
version: [1, 1.1],
version: [1, 1.1, 1.2],
description: 'A simple chain to prompt a large language model',
defaults: {
name: 'Basic LLM Chain',
@@ -217,25 +237,11 @@ export class ChainLlm implements INodeType {
],
},
},
// eslint-disable-next-line n8n-nodes-base/node-class-description-inputs-wrong-regular-node
inputs: [
NodeConnectionType.Main,
{
displayName: 'Model',
maxConnections: 1,
type: NodeConnectionType.AiLanguageModel,
required: true,
},
{
displayName: 'Output Parser',
type: NodeConnectionType.AiOutputParser,
required: false,
},
],
inputs: `={{ ((parameter) => { ${getInputs.toString()}; return getInputs(parameter) })($parameter) }}`,
outputs: [NodeConnectionType.Main],
credentials: [],
properties: [
getTemplateNoticeField(1951),
getTemplateNoticeField(1978),
{
displayName: 'Prompt',
name: 'prompt',
@@ -256,7 +262,7 @@ export class ChainLlm implements INodeType {
default: '={{ $json.chat_input }}',
displayOptions: {
show: {
'@version': [1.1],
'@version': [1.1, 1.2],
},
},
},
@@ -400,6 +406,28 @@ export class ChainLlm implements INodeType {
},
],
},
{
displayName: 'Require Specific Output Format',
name: 'hasOutputParser',
type: 'boolean',
default: false,
displayOptions: {
show: {
'@version': [1.2],
},
},
},
{
displayName: `Connect an <a data-action='openSelectiveNodeCreator' data-action-parameter-connectiontype='${NodeConnectionType.AiOutputParser}'>output parser</a> on the canvas to specify the output format you require`,
name: 'notice',
type: 'notice',
default: '',
displayOptions: {
show: {
hasOutputParser: [true],
},
},
},
],
};
@@ -413,10 +441,14 @@ export class ChainLlm implements INodeType {
0,
)) as BaseLanguageModel;
const outputParsers = (await this.getInputConnectionData(
NodeConnectionType.AiOutputParser,
0,
)) as BaseOutputParser[];
let outputParsers: BaseOutputParser[] = [];
if (this.getNodeParameter('hasOutputParser', 0, true) === true) {
outputParsers = (await this.getInputConnectionData(
NodeConnectionType.AiOutputParser,
0,
)) as BaseOutputParser[];
}
for (let itemIndex = 0; itemIndex < items.length; itemIndex++) {
const prompt = this.getNodeParameter('prompt', itemIndex) as string;

View File

@@ -1,277 +1,39 @@
import {
NodeConnectionType,
type IExecuteFunctions,
type INodeExecutionData,
type INodeType,
type INodeTypeDescription,
} from 'n8n-workflow';
import type { INodeTypeBaseDescription, IVersionedNodeType } from 'n8n-workflow';
import { VersionedNodeType } from 'n8n-workflow';
import type { SummarizationChainParams } from 'langchain/chains';
import { loadSummarizationChain } from 'langchain/chains';
import type { BaseLanguageModel } from 'langchain/dist/base_language';
import type { Document } from 'langchain/document';
import { PromptTemplate } from 'langchain/prompts';
import { N8nJsonLoader } from '../../../utils/N8nJsonLoader';
import { N8nBinaryLoader } from '../../../utils/N8nBinaryLoader';
import { getTemplateNoticeField } from '../../../utils/sharedFields';
import { REFINE_PROMPT_TEMPLATE, DEFAULT_PROMPT_TEMPLATE } from './prompt';
import { ChainSummarizationV1 } from './V1/ChainSummarizationV1.node';
import { ChainSummarizationV2 } from './V2/ChainSummarizationV2.node';
export class ChainSummarization implements INodeType {
description: INodeTypeDescription = {
displayName: 'Summarization Chain',
name: 'chainSummarization',
icon: 'fa:link',
group: ['transform'],
version: 1,
description: 'Transforms text into a concise summary',
defaults: {
name: 'Summarization Chain',
color: '#909298',
},
codex: {
alias: ['LangChain'],
categories: ['AI'],
subcategories: {
AI: ['Chains'],
export class ChainSummarization extends VersionedNodeType {
constructor() {
const baseDescription: INodeTypeBaseDescription = {
displayName: 'Summarization Chain',
name: 'chainSummarization',
icon: 'fa:link',
group: ['transform'],
description: 'Transforms text into a concise summary',
codex: {
alias: ['LangChain'],
categories: ['AI'],
subcategories: {
AI: ['Chains'],
},
resources: {
primaryDocumentation: [
{
url: 'https://docs.n8n.io/integrations/builtin/cluster-nodes/root-nodes/n8n-nodes-langchain.chainsummarization/',
},
],
},
},
resources: {
primaryDocumentation: [
{
url: 'https://docs.n8n.io/integrations/builtin/cluster-nodes/root-nodes/n8n-nodes-langchain.chainsummarization/',
},
],
},
},
// eslint-disable-next-line n8n-nodes-base/node-class-description-inputs-wrong-regular-node
inputs: [
NodeConnectionType.Main,
{
displayName: 'Model',
maxConnections: 1,
type: NodeConnectionType.AiLanguageModel,
required: true,
},
{
displayName: 'Document',
maxConnections: 1,
type: NodeConnectionType.AiDocument,
required: true,
},
],
outputs: [NodeConnectionType.Main],
credentials: [],
properties: [
getTemplateNoticeField(1951),
{
displayName: 'Type',
name: 'type',
type: 'options',
description: 'The type of summarization to run',
default: 'map_reduce',
options: [
{
name: 'Map Reduce (Recommended)',
value: 'map_reduce',
description:
'Summarize each document (or chunk) individually, then summarize those summaries',
},
{
name: 'Refine',
value: 'refine',
description:
'Summarize the first document (or chunk). Then update that summary based on the next document (or chunk), and repeat.',
},
{
name: 'Stuff',
value: 'stuff',
description: 'Pass all documents (or chunks) at once. Ideal for small datasets.',
},
],
},
{
displayName: 'Options',
name: 'options',
type: 'collection',
default: {},
placeholder: 'Add Option',
options: [
{
displayName: 'Final Prompt to Combine',
name: 'combineMapPrompt',
type: 'string',
hint: 'The prompt to combine individual summaries',
displayOptions: {
show: {
'/type': ['map_reduce'],
},
},
default: DEFAULT_PROMPT_TEMPLATE,
typeOptions: {
rows: 6,
},
},
{
displayName: 'Individual Summary Prompt',
name: 'prompt',
type: 'string',
default: DEFAULT_PROMPT_TEMPLATE,
hint: 'The prompt to summarize an individual document (or chunk)',
displayOptions: {
show: {
'/type': ['map_reduce'],
},
},
typeOptions: {
rows: 6,
},
},
{
displayName: 'Prompt',
name: 'prompt',
type: 'string',
default: DEFAULT_PROMPT_TEMPLATE,
displayOptions: {
show: {
'/type': ['stuff'],
},
},
typeOptions: {
rows: 6,
},
},
{
displayName: 'Subsequent (Refine) Prompt',
name: 'refinePrompt',
type: 'string',
displayOptions: {
show: {
'/type': ['refine'],
},
},
default: REFINE_PROMPT_TEMPLATE,
hint: 'The prompt to refine the summary based on the next document (or chunk)',
typeOptions: {
rows: 6,
},
},
{
displayName: 'Initial Prompt',
name: 'refineQuestionPrompt',
type: 'string',
displayOptions: {
show: {
'/type': ['refine'],
},
},
default: DEFAULT_PROMPT_TEMPLATE,
hint: 'The prompt for the first document (or chunk)',
typeOptions: {
rows: 6,
},
},
],
},
],
};
async execute(this: IExecuteFunctions): Promise<INodeExecutionData[][]> {
this.logger.verbose('Executing Vector Store QA Chain');
const type = this.getNodeParameter('type', 0) as 'map_reduce' | 'stuff' | 'refine';
const model = (await this.getInputConnectionData(
NodeConnectionType.AiLanguageModel,
0,
)) as BaseLanguageModel;
const documentInput = (await this.getInputConnectionData(NodeConnectionType.AiDocument, 0)) as
| N8nJsonLoader
| Array<Document<Record<string, unknown>>>;
const options = this.getNodeParameter('options', 0, {}) as {
prompt?: string;
refineQuestionPrompt?: string;
refinePrompt?: string;
combineMapPrompt?: string;
defaultVersion: 2,
};
const chainArgs: SummarizationChainParams = {
type,
const nodeVersions: IVersionedNodeType['nodeVersions'] = {
1: new ChainSummarizationV1(baseDescription),
2: new ChainSummarizationV2(baseDescription),
};
// Map reduce prompt override
if (type === 'map_reduce') {
const mapReduceArgs = chainArgs as SummarizationChainParams & {
type: 'map_reduce';
};
if (options.combineMapPrompt) {
mapReduceArgs.combineMapPrompt = new PromptTemplate({
template: options.combineMapPrompt,
inputVariables: ['text'],
});
}
if (options.prompt) {
mapReduceArgs.combinePrompt = new PromptTemplate({
template: options.prompt,
inputVariables: ['text'],
});
}
}
// Stuff prompt override
if (type === 'stuff') {
const stuffArgs = chainArgs as SummarizationChainParams & {
type: 'stuff';
};
if (options.prompt) {
stuffArgs.prompt = new PromptTemplate({
template: options.prompt,
inputVariables: ['text'],
});
}
}
// Refine prompt override
if (type === 'refine') {
const refineArgs = chainArgs as SummarizationChainParams & {
type: 'refine';
};
if (options.refinePrompt) {
refineArgs.refinePrompt = new PromptTemplate({
template: options.refinePrompt,
inputVariables: ['existing_answer', 'text'],
});
}
if (options.refineQuestionPrompt) {
refineArgs.questionPrompt = new PromptTemplate({
template: options.refineQuestionPrompt,
inputVariables: ['text'],
});
}
}
const chain = loadSummarizationChain(model, chainArgs);
const items = this.getInputData();
const returnData: INodeExecutionData[] = [];
for (let itemIndex = 0; itemIndex < items.length; itemIndex++) {
let processedDocuments: Document[];
if (documentInput instanceof N8nJsonLoader || documentInput instanceof N8nBinaryLoader) {
processedDocuments = await documentInput.processItem(items[itemIndex], itemIndex);
} else {
processedDocuments = documentInput;
}
const response = await chain.call({
input_documents: processedDocuments,
});
returnData.push({ json: { response } });
}
return this.prepareOutputData(returnData);
super(nodeVersions, baseDescription);
}
}

View File

@@ -0,0 +1,263 @@
import {
NodeConnectionType,
type INodeTypeBaseDescription,
type IExecuteFunctions,
type INodeExecutionData,
type INodeType,
type INodeTypeDescription,
} from 'n8n-workflow';
import type { SummarizationChainParams } from 'langchain/chains';
import { loadSummarizationChain } from 'langchain/chains';
import type { BaseLanguageModel } from 'langchain/dist/base_language';
import type { Document } from 'langchain/document';
import { PromptTemplate } from 'langchain/prompts';
import { N8nJsonLoader } from '../../../../utils/N8nJsonLoader';
import { N8nBinaryLoader } from '../../../../utils/N8nBinaryLoader';
import { getTemplateNoticeField } from '../../../../utils/sharedFields';
import { REFINE_PROMPT_TEMPLATE, DEFAULT_PROMPT_TEMPLATE } from '../prompt';
export class ChainSummarizationV1 implements INodeType {
description: INodeTypeDescription;
constructor(baseDescription: INodeTypeBaseDescription) {
this.description = {
...baseDescription,
version: 1,
defaults: {
name: 'Summarization Chain',
color: '#909298',
},
// eslint-disable-next-line n8n-nodes-base/node-class-description-inputs-wrong-regular-node
inputs: [
NodeConnectionType.Main,
{
displayName: 'Model',
maxConnections: 1,
type: NodeConnectionType.AiLanguageModel,
required: true,
},
{
displayName: 'Document',
maxConnections: 1,
type: NodeConnectionType.AiDocument,
required: true,
},
],
outputs: [NodeConnectionType.Main],
credentials: [],
properties: [
getTemplateNoticeField(1951),
{
displayName: 'Type',
name: 'type',
type: 'options',
description: 'The type of summarization to run',
default: 'map_reduce',
options: [
{
name: 'Map Reduce (Recommended)',
value: 'map_reduce',
description:
'Summarize each document (or chunk) individually, then summarize those summaries',
},
{
name: 'Refine',
value: 'refine',
description:
'Summarize the first document (or chunk). Then update that summary based on the next document (or chunk), and repeat.',
},
{
name: 'Stuff',
value: 'stuff',
description: 'Pass all documents (or chunks) at once. Ideal for small datasets.',
},
],
},
{
displayName: 'Options',
name: 'options',
type: 'collection',
default: {},
placeholder: 'Add Option',
options: [
{
displayName: 'Final Prompt to Combine',
name: 'combineMapPrompt',
type: 'string',
hint: 'The prompt to combine individual summaries',
displayOptions: {
show: {
'/type': ['map_reduce'],
},
},
default: DEFAULT_PROMPT_TEMPLATE,
typeOptions: {
rows: 6,
},
},
{
displayName: 'Individual Summary Prompt',
name: 'prompt',
type: 'string',
default: DEFAULT_PROMPT_TEMPLATE,
hint: 'The prompt to summarize an individual document (or chunk)',
displayOptions: {
show: {
'/type': ['map_reduce'],
},
},
typeOptions: {
rows: 6,
},
},
{
displayName: 'Prompt',
name: 'prompt',
type: 'string',
default: DEFAULT_PROMPT_TEMPLATE,
displayOptions: {
show: {
'/type': ['stuff'],
},
},
typeOptions: {
rows: 6,
},
},
{
displayName: 'Subsequent (Refine) Prompt',
name: 'refinePrompt',
type: 'string',
displayOptions: {
show: {
'/type': ['refine'],
},
},
default: REFINE_PROMPT_TEMPLATE,
hint: 'The prompt to refine the summary based on the next document (or chunk)',
typeOptions: {
rows: 6,
},
},
{
displayName: 'Initial Prompt',
name: 'refineQuestionPrompt',
type: 'string',
displayOptions: {
show: {
'/type': ['refine'],
},
},
default: DEFAULT_PROMPT_TEMPLATE,
hint: 'The prompt for the first document (or chunk)',
typeOptions: {
rows: 6,
},
},
],
},
],
};
}
async execute(this: IExecuteFunctions): Promise<INodeExecutionData[][]> {
this.logger.verbose('Executing Vector Store QA Chain');
const type = this.getNodeParameter('type', 0) as 'map_reduce' | 'stuff' | 'refine';
const model = (await this.getInputConnectionData(
NodeConnectionType.AiLanguageModel,
0,
)) as BaseLanguageModel;
const documentInput = (await this.getInputConnectionData(NodeConnectionType.AiDocument, 0)) as
| N8nJsonLoader
| Array<Document<Record<string, unknown>>>;
const options = this.getNodeParameter('options', 0, {}) as {
prompt?: string;
refineQuestionPrompt?: string;
refinePrompt?: string;
combineMapPrompt?: string;
};
const chainArgs: SummarizationChainParams = {
type,
};
// Map reduce prompt override
if (type === 'map_reduce') {
const mapReduceArgs = chainArgs as SummarizationChainParams & {
type: 'map_reduce';
};
if (options.combineMapPrompt) {
mapReduceArgs.combineMapPrompt = new PromptTemplate({
template: options.combineMapPrompt,
inputVariables: ['text'],
});
}
if (options.prompt) {
mapReduceArgs.combinePrompt = new PromptTemplate({
template: options.prompt,
inputVariables: ['text'],
});
}
}
// Stuff prompt override
if (type === 'stuff') {
const stuffArgs = chainArgs as SummarizationChainParams & {
type: 'stuff';
};
if (options.prompt) {
stuffArgs.prompt = new PromptTemplate({
template: options.prompt,
inputVariables: ['text'],
});
}
}
// Refine prompt override
if (type === 'refine') {
const refineArgs = chainArgs as SummarizationChainParams & {
type: 'refine';
};
if (options.refinePrompt) {
refineArgs.refinePrompt = new PromptTemplate({
template: options.refinePrompt,
inputVariables: ['existing_answer', 'text'],
});
}
if (options.refineQuestionPrompt) {
refineArgs.questionPrompt = new PromptTemplate({
template: options.refineQuestionPrompt,
inputVariables: ['text'],
});
}
}
const chain = loadSummarizationChain(model, chainArgs);
const items = this.getInputData();
const returnData: INodeExecutionData[] = [];
for (let itemIndex = 0; itemIndex < items.length; itemIndex++) {
let processedDocuments: Document[];
if (documentInput instanceof N8nJsonLoader || documentInput instanceof N8nBinaryLoader) {
processedDocuments = await documentInput.processItem(items[itemIndex], itemIndex);
} else {
processedDocuments = documentInput;
}
const response = await chain.call({
input_documents: processedDocuments,
});
returnData.push({ json: { response } });
}
return this.prepareOutputData(returnData);
}
}

View File

@@ -0,0 +1,420 @@
import { NodeConnectionType } from 'n8n-workflow';
import type {
INodeTypeBaseDescription,
IExecuteFunctions,
INodeExecutionData,
INodeType,
INodeTypeDescription,
IDataObject,
} from 'n8n-workflow';
import { loadSummarizationChain } from 'langchain/chains';
import type { BaseLanguageModel } from 'langchain/dist/base_language';
import type { Document } from 'langchain/document';
import type { TextSplitter } from 'langchain/text_splitter';
import { RecursiveCharacterTextSplitter } from 'langchain/text_splitter';
import { N8nJsonLoader } from '../../../../utils/N8nJsonLoader';
import { N8nBinaryLoader } from '../../../../utils/N8nBinaryLoader';
import { getTemplateNoticeField } from '../../../../utils/sharedFields';
import { REFINE_PROMPT_TEMPLATE, DEFAULT_PROMPT_TEMPLATE } from '../prompt';
import { getChainPromptsArgs } from '../helpers';
function getInputs(parameters: IDataObject) {
const chunkingMode = parameters?.chunkingMode;
const operationMode = parameters?.operationMode;
const inputs = [
{ displayName: '', type: NodeConnectionType.Main },
{
displayName: 'Model',
maxConnections: 1,
type: NodeConnectionType.AiLanguageModel,
required: true,
},
];
if (operationMode === 'documentLoader') {
inputs.push({
displayName: 'Document',
type: NodeConnectionType.AiDocument,
required: true,
maxConnections: 1,
});
return inputs;
}
if (chunkingMode === 'advanced') {
inputs.push({
displayName: 'Text Splitter',
type: NodeConnectionType.AiTextSplitter,
required: false,
maxConnections: 1,
});
return inputs;
}
return inputs;
}
export class ChainSummarizationV2 implements INodeType {
description: INodeTypeDescription;
constructor(baseDescription: INodeTypeBaseDescription) {
this.description = {
...baseDescription,
version: [2],
defaults: {
name: 'Summarization Chain',
color: '#909298',
},
// eslint-disable-next-line n8n-nodes-base/node-class-description-inputs-wrong-regular-node
inputs: `={{ ((parameter) => { ${getInputs.toString()}; return getInputs(parameter) })($parameter) }}`,
outputs: [NodeConnectionType.Main],
credentials: [],
properties: [
getTemplateNoticeField(1951),
{
displayName: 'Data to Summarize',
name: 'operationMode',
noDataExpression: true,
type: 'options',
description: 'How to pass data into the summarization chain',
default: 'nodeInputJson',
options: [
{
name: 'Use Node Input (JSON)',
value: 'nodeInputJson',
description: 'Summarize the JSON data coming into this node from the previous one',
},
{
name: 'Use Node Input (Binary)',
value: 'nodeInputBinary',
description: 'Summarize the binary data coming into this node from the previous one',
},
{
name: 'Use Document Loader',
value: 'documentLoader',
description: 'Use a loader sub-node with more configuration options',
},
],
},
{
displayName: 'Chunking Strategy',
name: 'chunkingMode',
noDataExpression: true,
type: 'options',
description: 'Chunk splitting strategy',
default: 'simple',
options: [
{
name: 'Simple (Define Below)',
value: 'simple',
},
{
name: 'Advanced',
value: 'advanced',
description: 'Use a splitter sub-node with more configuration options',
},
],
displayOptions: {
show: {
'/operationMode': ['nodeInputJson', 'nodeInputBinary'],
},
},
},
{
displayName: 'Characters Per Chunk',
name: 'chunkSize',
description:
'Controls the max size (in terms of number of characters) of the final document chunk',
type: 'number',
default: 1000,
displayOptions: {
show: {
'/chunkingMode': ['simple'],
},
},
},
{
displayName: 'Chunk Overlap (Characters)',
name: 'chunkOverlap',
type: 'number',
description: 'Specifies how much characters overlap there should be between chunks',
default: 200,
displayOptions: {
show: {
'/chunkingMode': ['simple'],
},
},
},
{
displayName: 'Options',
name: 'options',
type: 'collection',
default: {},
placeholder: 'Add Option',
options: [
{
displayName: 'Input Data Field Name',
name: 'binaryDataKey',
type: 'string',
default: 'data',
description:
'The name of the field in the agent or chains input that contains the binary file to be processed',
displayOptions: {
show: {
'/operationMode': ['nodeInputBinary'],
},
},
},
{
displayName: 'Summarization Method and Prompts',
name: 'summarizationMethodAndPrompts',
type: 'fixedCollection',
default: {
values: {
summarizationMethod: 'map_reduce',
prompt: DEFAULT_PROMPT_TEMPLATE,
combineMapPrompt: DEFAULT_PROMPT_TEMPLATE,
},
},
placeholder: 'Add Option',
typeOptions: {},
options: [
{
name: 'values',
displayName: 'Values',
values: [
{
displayName: 'Summarization Method',
name: 'summarizationMethod',
type: 'options',
description: 'The type of summarization to run',
default: 'map_reduce',
options: [
{
name: 'Map Reduce (Recommended)',
value: 'map_reduce',
description:
'Summarize each document (or chunk) individually, then summarize those summaries',
},
{
name: 'Refine',
value: 'refine',
description:
'Summarize the first document (or chunk). Then update that summary based on the next document (or chunk), and repeat.',
},
{
name: 'Stuff',
value: 'stuff',
description:
'Pass all documents (or chunks) at once. Ideal for small datasets.',
},
],
},
{
displayName: 'Final Prompt to Combine',
name: 'combineMapPrompt',
type: 'string',
hint: 'The prompt to combine individual summaries',
displayOptions: {
hide: {
'/options.summarizationMethodAndPrompts.values.summarizationMethod': [
'stuff',
'refine',
],
},
},
default: DEFAULT_PROMPT_TEMPLATE,
typeOptions: {
rows: 9,
},
},
{
displayName: 'Individual Summary Prompt',
name: 'prompt',
type: 'string',
default: DEFAULT_PROMPT_TEMPLATE,
hint: 'The prompt to summarize an individual document (or chunk)',
displayOptions: {
hide: {
'/options.summarizationMethodAndPrompts.values.summarizationMethod': [
'stuff',
'refine',
],
},
},
typeOptions: {
rows: 9,
},
},
{
displayName: 'Prompt',
name: 'prompt',
type: 'string',
default: DEFAULT_PROMPT_TEMPLATE,
displayOptions: {
hide: {
'/options.summarizationMethodAndPrompts.values.summarizationMethod': [
'refine',
'map_reduce',
],
},
},
typeOptions: {
rows: 9,
},
},
{
displayName: 'Subsequent (Refine) Prompt',
name: 'refinePrompt',
type: 'string',
displayOptions: {
hide: {
'/options.summarizationMethodAndPrompts.values.summarizationMethod': [
'stuff',
'map_reduce',
],
},
},
default: REFINE_PROMPT_TEMPLATE,
hint: 'The prompt to refine the summary based on the next document (or chunk)',
typeOptions: {
rows: 9,
},
},
{
displayName: 'Initial Prompt',
name: 'refineQuestionPrompt',
type: 'string',
displayOptions: {
hide: {
'/options.summarizationMethodAndPrompts.values.summarizationMethod': [
'stuff',
'map_reduce',
],
},
},
default: DEFAULT_PROMPT_TEMPLATE,
hint: 'The prompt for the first document (or chunk)',
typeOptions: {
rows: 9,
},
},
],
},
],
},
],
},
],
};
}
async execute(this: IExecuteFunctions): Promise<INodeExecutionData[][]> {
this.logger.verbose('Executing Summarization Chain V2');
const operationMode = this.getNodeParameter('operationMode', 0, 'nodeInputJson') as
| 'nodeInputJson'
| 'nodeInputBinary'
| 'documentLoader';
const chunkingMode = this.getNodeParameter('chunkingMode', 0, 'simple') as
| 'simple'
| 'advanced';
const model = (await this.getInputConnectionData(
NodeConnectionType.AiLanguageModel,
0,
)) as BaseLanguageModel;
const items = this.getInputData();
const returnData: INodeExecutionData[] = [];
for (let itemIndex = 0; itemIndex < items.length; itemIndex++) {
const summarizationMethodAndPrompts = this.getNodeParameter(
'options.summarizationMethodAndPrompts.values',
itemIndex,
{},
) as {
prompt?: string;
refineQuestionPrompt?: string;
refinePrompt?: string;
summarizationMethod: 'map_reduce' | 'stuff' | 'refine';
combineMapPrompt?: string;
};
const chainArgs = getChainPromptsArgs(
summarizationMethodAndPrompts.summarizationMethod ?? 'map_reduce',
summarizationMethodAndPrompts,
);
const chain = loadSummarizationChain(model, chainArgs);
const item = items[itemIndex];
let processedDocuments: Document[];
// Use dedicated document loader input to load documents
if (operationMode === 'documentLoader') {
const documentInput = (await this.getInputConnectionData(
NodeConnectionType.AiDocument,
0,
)) as N8nJsonLoader | Array<Document<Record<string, unknown>>>;
const isN8nLoader =
documentInput instanceof N8nJsonLoader || documentInput instanceof N8nBinaryLoader;
processedDocuments = isN8nLoader
? await documentInput.processItem(item, itemIndex)
: documentInput;
const response = await chain.call({
input_documents: processedDocuments,
});
returnData.push({ json: { response } });
}
// Take the input and use binary or json loader
if (['nodeInputJson', 'nodeInputBinary'].includes(operationMode)) {
let textSplitter: TextSplitter | undefined;
switch (chunkingMode) {
// In simple mode we use recursive character splitter with default settings
case 'simple':
const chunkSize = this.getNodeParameter('chunkSize', itemIndex, 1000) as number;
const chunkOverlap = this.getNodeParameter('chunkOverlap', itemIndex, 200) as number;
textSplitter = new RecursiveCharacterTextSplitter({ chunkOverlap, chunkSize });
break;
// In advanced mode user can connect text splitter node so we just retrieve it
case 'advanced':
textSplitter = (await this.getInputConnectionData(
NodeConnectionType.AiTextSplitter,
0,
)) as TextSplitter | undefined;
break;
default:
break;
}
let processor: N8nJsonLoader | N8nBinaryLoader;
if (operationMode === 'nodeInputBinary') {
const binaryDataKey = this.getNodeParameter(
'options.binaryDataKey',
itemIndex,
'data',
) as string;
processor = new N8nBinaryLoader(this, 'options.', binaryDataKey, textSplitter);
} else {
processor = new N8nJsonLoader(this, 'options.', textSplitter);
}
const processedItem = await processor.processItem(item, itemIndex);
const response = await chain.call({
input_documents: processedItem,
});
returnData.push({ json: { response } });
}
}
return this.prepareOutputData(returnData);
}
}

View File

@@ -0,0 +1,72 @@
import type { SummarizationChainParams } from 'langchain/chains';
import { PromptTemplate } from 'langchain/prompts';
interface ChainTypeOptions {
combineMapPrompt?: string;
prompt?: string;
refinePrompt?: string;
refineQuestionPrompt?: string;
}
export function getChainPromptsArgs(
type: 'stuff' | 'map_reduce' | 'refine',
options: ChainTypeOptions,
) {
const chainArgs: SummarizationChainParams = {
type,
};
// Map reduce prompt override
if (type === 'map_reduce') {
const mapReduceArgs = chainArgs as SummarizationChainParams & {
type: 'map_reduce';
};
if (options.combineMapPrompt) {
mapReduceArgs.combineMapPrompt = new PromptTemplate({
template: options.combineMapPrompt,
inputVariables: ['text'],
});
}
if (options.prompt) {
mapReduceArgs.combinePrompt = new PromptTemplate({
template: options.prompt,
inputVariables: ['text'],
});
}
}
// Stuff prompt override
if (type === 'stuff') {
const stuffArgs = chainArgs as SummarizationChainParams & {
type: 'stuff';
};
if (options.prompt) {
stuffArgs.prompt = new PromptTemplate({
template: options.prompt,
inputVariables: ['text'],
});
}
}
// Refine prompt override
if (type === 'refine') {
const refineArgs = chainArgs as SummarizationChainParams & {
type: 'refine';
};
if (options.refinePrompt) {
refineArgs.refinePrompt = new PromptTemplate({
template: options.refinePrompt,
inputVariables: ['existing_answer', 'text'],
});
}
if (options.refineQuestionPrompt) {
refineArgs.questionPrompt = new PromptTemplate({
template: options.refineQuestionPrompt,
inputVariables: ['text'],
});
}
}
return chainArgs;
}

View File

@@ -17,6 +17,7 @@ import { getConnectionHintNoticeField, metadataFilterField } from '../../../util
import 'mammoth'; // for docx
import 'epub2'; // for epub
import 'pdf-parse'; // for pdf
import type { TextSplitter } from 'langchain/text_splitter';
export class DocumentBinaryInputLoader implements INodeType {
description: INodeTypeDescription = {
@@ -177,7 +178,13 @@ export class DocumentBinaryInputLoader implements INodeType {
async supplyData(this: IExecuteFunctions): Promise<SupplyData> {
this.logger.verbose('Supply Data for Binary Input Loader');
const processor = new N8nBinaryLoader(this);
const textSplitter = (await this.getInputConnectionData(
NodeConnectionType.AiTextSplitter,
0,
)) as TextSplitter | undefined;
const binaryDataKey = this.getNodeParameter('binaryDataKey', 0) as string;
const processor = new N8nBinaryLoader(this, undefined, binaryDataKey, textSplitter);
return {
response: logWrapper(processor, this),

View File

@@ -7,6 +7,7 @@ import {
type SupplyData,
} from 'n8n-workflow';
import type { TextSplitter } from 'langchain/text_splitter';
import { logWrapper } from '../../../utils/logWrapper';
import { N8nBinaryLoader } from '../../../utils/N8nBinaryLoader';
import { metadataFilterField } from '../../../utils/sharedFields';
@@ -257,11 +258,16 @@ export class DocumentDefaultDataLoader implements INodeType {
async supplyData(this: IExecuteFunctions, itemIndex: number): Promise<SupplyData> {
const dataType = this.getNodeParameter('dataType', itemIndex, 'json') as 'json' | 'binary';
const textSplitter = (await this.getInputConnectionData(
NodeConnectionType.AiTextSplitter,
0,
)) as TextSplitter | undefined;
const binaryDataKey = this.getNodeParameter('binaryDataKey', itemIndex, '') as string;
const processor =
dataType === 'binary'
? new N8nBinaryLoader(this, 'options.')
: new N8nJsonLoader(this, 'options.');
? new N8nBinaryLoader(this, 'options.', binaryDataKey, textSplitter)
: new N8nJsonLoader(this, 'options.', textSplitter);
return {
response: logWrapper(processor, this),

View File

@@ -7,6 +7,7 @@ import {
type SupplyData,
} from 'n8n-workflow';
import type { TextSplitter } from 'langchain/text_splitter';
import { logWrapper } from '../../../utils/logWrapper';
import { N8nJsonLoader } from '../../../utils/N8nJsonLoader';
import { getConnectionHintNoticeField, metadataFilterField } from '../../../utils/sharedFields';
@@ -80,7 +81,12 @@ export class DocumentJsonInputLoader implements INodeType {
async supplyData(this: IExecuteFunctions): Promise<SupplyData> {
this.logger.verbose('Supply Data for JSON Input Loader');
const processor = new N8nJsonLoader(this);
const textSplitter = (await this.getInputConnectionData(
NodeConnectionType.AiTextSplitter,
0,
)) as TextSplitter | undefined;
const processor = new N8nJsonLoader(this, undefined, textSplitter);
return {
response: logWrapper(processor, this),