|
| 1 | +import { get } from 'lodash' |
| 2 | +import { Document } from '@langchain/core/documents' |
| 3 | +import { VectorStore, VectorStoreRetriever, VectorStoreRetrieverInput } from '@langchain/core/vectorstores' |
| 4 | +import { INode, INodeData, INodeParams, INodeOutputsValue } from '../../../src/Interface' |
| 5 | +import { handleEscapeCharacters } from '../../../src' |
| 6 | + |
| 7 | +const defaultReturnFormat = '{{context}}\nSource: {{metadata.source}}' |
| 8 | + |
| 9 | +class CustomRetriever_Retrievers implements INode { |
| 10 | + label: string |
| 11 | + name: string |
| 12 | + version: number |
| 13 | + description: string |
| 14 | + type: string |
| 15 | + icon: string |
| 16 | + category: string |
| 17 | + baseClasses: string[] |
| 18 | + inputs: INodeParams[] |
| 19 | + outputs: INodeOutputsValue[] |
| 20 | + |
| 21 | + constructor() { |
| 22 | + this.label = 'Custom Retriever' |
| 23 | + this.name = 'customRetriever' |
| 24 | + this.version = 1.0 |
| 25 | + this.type = 'CustomRetriever' |
| 26 | + this.icon = 'customRetriever.svg' |
| 27 | + this.category = 'Retrievers' |
| 28 | + this.description = 'Return results based on predefined format' |
| 29 | + this.baseClasses = [this.type, 'BaseRetriever'] |
| 30 | + this.inputs = [ |
| 31 | + { |
| 32 | + label: 'Vector Store', |
| 33 | + name: 'vectorStore', |
| 34 | + type: 'VectorStore' |
| 35 | + }, |
| 36 | + { |
| 37 | + label: 'Query', |
| 38 | + name: 'query', |
| 39 | + type: 'string', |
| 40 | + description: 'Query to retrieve documents from retriever. If not specified, user question will be used', |
| 41 | + optional: true, |
| 42 | + acceptVariable: true |
| 43 | + }, |
| 44 | + { |
| 45 | + label: 'Result Format', |
| 46 | + name: 'resultFormat', |
| 47 | + type: 'string', |
| 48 | + rows: 4, |
| 49 | + description: |
| 50 | + 'Format to return the results in. Use {{context}} to insert the pageContent of the document and {{metadata.key}} to insert metadata values.', |
| 51 | + default: defaultReturnFormat |
| 52 | + }, |
| 53 | + { |
| 54 | + label: 'Top K', |
| 55 | + name: 'topK', |
| 56 | + description: 'Number of top results to fetch. Default to vector store topK', |
| 57 | + placeholder: '4', |
| 58 | + type: 'number', |
| 59 | + additionalParams: true, |
| 60 | + optional: true |
| 61 | + } |
| 62 | + ] |
| 63 | + this.outputs = [ |
| 64 | + { |
| 65 | + label: 'Custom Retriever', |
| 66 | + name: 'retriever', |
| 67 | + baseClasses: this.baseClasses |
| 68 | + }, |
| 69 | + { |
| 70 | + label: 'Document', |
| 71 | + name: 'document', |
| 72 | + description: 'Array of document objects containing metadata and pageContent', |
| 73 | + baseClasses: ['Document', 'json'] |
| 74 | + }, |
| 75 | + { |
| 76 | + label: 'Text', |
| 77 | + name: 'text', |
| 78 | + description: 'Concatenated string from pageContent of documents', |
| 79 | + baseClasses: ['string', 'json'] |
| 80 | + } |
| 81 | + ] |
| 82 | + } |
| 83 | + |
| 84 | + async init(nodeData: INodeData, input: string): Promise<any> { |
| 85 | + const vectorStore = nodeData.inputs?.vectorStore as VectorStore |
| 86 | + const query = nodeData.inputs?.query as string |
| 87 | + const topK = nodeData.inputs?.topK as string |
| 88 | + const resultFormat = nodeData.inputs?.resultFormat as string |
| 89 | + |
| 90 | + const output = nodeData.outputs?.output as string |
| 91 | + |
| 92 | + const retriever = CustomRetriever.fromVectorStore(vectorStore, { |
| 93 | + resultFormat, |
| 94 | + topK: topK ? parseInt(topK, 10) : (vectorStore as any)?.k ?? 4 |
| 95 | + }) |
| 96 | + |
| 97 | + if (output === 'retriever') return retriever |
| 98 | + else if (output === 'document') return await retriever.getRelevantDocuments(query ? query : input) |
| 99 | + else if (output === 'text') { |
| 100 | + let finaltext = '' |
| 101 | + |
| 102 | + const docs = await retriever.getRelevantDocuments(query ? query : input) |
| 103 | + |
| 104 | + for (const doc of docs) finaltext += `${doc.pageContent}\n` |
| 105 | + |
| 106 | + return handleEscapeCharacters(finaltext, false) |
| 107 | + } |
| 108 | + |
| 109 | + return retriever |
| 110 | + } |
| 111 | +} |
| 112 | + |
| 113 | +type RetrieverInput<V extends VectorStore> = Omit<VectorStoreRetrieverInput<V>, 'k'> & { |
| 114 | + topK?: number |
| 115 | + resultFormat?: string |
| 116 | +} |
| 117 | + |
| 118 | +class CustomRetriever<V extends VectorStore> extends VectorStoreRetriever<V> { |
| 119 | + resultFormat: string |
| 120 | + topK = 4 |
| 121 | + |
| 122 | + constructor(input: RetrieverInput<V>) { |
| 123 | + super(input) |
| 124 | + this.topK = input.topK ?? this.topK |
| 125 | + this.resultFormat = input.resultFormat ?? this.resultFormat |
| 126 | + } |
| 127 | + |
| 128 | + async getRelevantDocuments(query: string): Promise<Document[]> { |
| 129 | + const results = await this.vectorStore.similaritySearchWithScore(query, this.topK, this.filter) |
| 130 | + |
| 131 | + const finalDocs: Document[] = [] |
| 132 | + for (const result of results) { |
| 133 | + let res = this.resultFormat.replace(/{{context}}/g, result[0].pageContent) |
| 134 | + res = replaceMetadata(res, result[0].metadata) |
| 135 | + finalDocs.push( |
| 136 | + new Document({ |
| 137 | + pageContent: res, |
| 138 | + metadata: result[0].metadata |
| 139 | + }) |
| 140 | + ) |
| 141 | + } |
| 142 | + return finalDocs |
| 143 | + } |
| 144 | + |
| 145 | + static fromVectorStore<V extends VectorStore>(vectorStore: V, options: Omit<RetrieverInput<V>, 'vectorStore'>) { |
| 146 | + return new this<V>({ ...options, vectorStore }) |
| 147 | + } |
| 148 | +} |
| 149 | + |
| 150 | +function replaceMetadata(template: string, metadata: Record<string, any>): string { |
| 151 | + const metadataRegex = /{{metadata\.([\w.]+)}}/g |
| 152 | + |
| 153 | + return template.replace(metadataRegex, (match, path) => { |
| 154 | + const value = get(metadata, path) |
| 155 | + return value !== undefined ? String(value) : match |
| 156 | + }) |
| 157 | +} |
| 158 | + |
| 159 | +module.exports = { nodeClass: CustomRetriever_Retrievers } |
0 commit comments