mirror of
https://github.com/farcasclaudiu/Flowise.git
synced 2026-06-28 15:00:57 +03:00
Merge pull request #1462 from vectara/add_mmr_vectara
added support for MMR
This commit is contained in:
Henry Heng
@@ -69,22 +69,23 @@ class VectaraChain_Chains implements INode {
|
||||
options: [
|
||||
{
|
||||
label: 'vectara-summary-ext-v1.2.0 (gpt-3.5-turbo)',
|
||||
name: 'vectara-summary-ext-v1.2.0'
|
||||
name: 'vectara-summary-ext-v1.2.0',
|
||||
description: 'base summarizer, available to all Vectara users'
|
||||
},
|
||||
{
|
||||
label: 'vectara-experimental-summary-ext-2023-10-23-small (gpt-3.5-turbo)',
|
||||
name: 'vectara-experimental-summary-ext-2023-10-23-small',
|
||||
description: 'In beta, available to both Growth and Scale Vectara users'
|
||||
description: `In beta, available to both Growth and <a target="_blank" href="https://vectara.com/pricing/">Scale</a> Vectara users`
|
||||
},
|
||||
{
|
||||
label: 'vectara-summary-ext-v1.3.0 (gpt-4.0)',
|
||||
name: 'vectara-summary-ext-v1.3.0',
|
||||
description: 'Only available to paying Scale Vectara users'
|
||||
description: 'Only available to <a target="_blank" href="https://vectara.com/pricing/">Scale</a> Vectara users'
|
||||
},
|
||||
{
|
||||
label: 'vectara-experimental-summary-ext-2023-10-23-med (gpt-4.0)',
|
||||
name: 'vectara-experimental-summary-ext-2023-10-23-med',
|
||||
description: 'In beta, only available to paying Scale Vectara users'
|
||||
description: `In beta, only available to <a target="_blank" href="https://vectara.com/pricing/">Scale</a> Vectara users`
|
||||
}
|
||||
],
|
||||
default: 'vectara-summary-ext-v1.2.0'
|
||||
@@ -228,7 +229,7 @@ class VectaraChain_Chains implements INode {
|
||||
|
||||
async run(nodeData: INodeData, input: string): Promise<object> {
|
||||
const vectorStore = nodeData.inputs?.vectaraStore as VectaraStore
|
||||
const responseLang = (nodeData.inputs?.responseLang as string) ?? 'auto'
|
||||
const responseLang = (nodeData.inputs?.responseLang as string) ?? 'eng'
|
||||
const summarizerPromptName = nodeData.inputs?.summarizerPromptName as string
|
||||
const maxSummarizedResultsStr = nodeData.inputs?.maxSummarizedResults as string
|
||||
const maxSummarizedResults = maxSummarizedResultsStr ? parseInt(maxSummarizedResultsStr, 10) : 7
|
||||
@@ -247,17 +248,31 @@ class VectaraChain_Chains implements INode {
|
||||
lexicalInterpolationConfig: { lambda: vectaraFilter?.lambda ?? 0.025 }
|
||||
}))
|
||||
|
||||
// Vectara reranker ID for MMR (https://docs.vectara.com/docs/api-reference/search-apis/reranking#maximal-marginal-relevance-mmr-reranker)
|
||||
const mmrRerankerId = 272725718
|
||||
const mmrEnabled = vectaraFilter?.mmrConfig?.enabled
|
||||
|
||||
const data = {
|
||||
query: [
|
||||
{
|
||||
query: input,
|
||||
start: 0,
|
||||
numResults: topK,
|
||||
numResults: mmrEnabled ? vectaraFilter?.mmrTopK : topK,
|
||||
corpusKey: corpusKeys,
|
||||
contextConfig: {
|
||||
sentencesAfter: vectaraFilter?.contextConfig?.sentencesAfter ?? 2,
|
||||
sentencesBefore: vectaraFilter?.contextConfig?.sentencesBefore ?? 2
|
||||
},
|
||||
corpusKey: corpusKeys,
|
||||
...(mmrEnabled
|
||||
? {
|
||||
rerankingConfig: {
|
||||
rerankerId: mmrRerankerId,
|
||||
mmrConfig: {
|
||||
diversityBias: vectaraFilter?.mmrConfig.diversityBias
|
||||
}
|
||||
}
|
||||
}
|
||||
: {}),
|
||||
summary: [
|
||||
{
|
||||
summarizerPromptName,
|
||||
@@ -285,6 +300,14 @@ class VectaraChain_Chains implements INode {
|
||||
const documents = result.responseSet[0].document
|
||||
let rawSummarizedText = ''
|
||||
|
||||
// remove responses that are not in the topK (in case of MMR)
|
||||
// Note that this does not really matter functionally due to the reorder citations, but it is more efficient
|
||||
const maxResponses = mmrEnabled ? Math.min(responses.length, topK) : responses.length
|
||||
if (responses.length > maxResponses) {
|
||||
responses.splice(0, maxResponses)
|
||||
}
|
||||
|
||||
// Add metadata to each text response given its corresponding document metadata
|
||||
for (let i = 0; i < responses.length; i += 1) {
|
||||
const responseMetadata = responses[i].metadata
|
||||
const documentMetadata = documents[responses[i].documentIndex].metadata
|
||||
@@ -301,13 +324,13 @@ class VectaraChain_Chains implements INode {
|
||||
responses[i].metadata = combinedMetadata
|
||||
}
|
||||
|
||||
// Create the summarization response
|
||||
const summaryStatus = result.responseSet[0].summary[0].status
|
||||
if (summaryStatus.length > 0 && summaryStatus[0].code === 'BAD_REQUEST') {
|
||||
throw new Error(
|
||||
`BAD REQUEST: Too much text for the summarizer to summarize. Please try reducing the number of search results to summarize, or the context of each result by adjusting the 'summary_num_sentences', and 'summary_num_results' parameters respectively.`
|
||||
)
|
||||
}
|
||||
|
||||
if (
|
||||
summaryStatus.length > 0 &&
|
||||
summaryStatus[0].code === 'NOT_FOUND' &&
|
||||
@@ -316,8 +339,8 @@ class VectaraChain_Chains implements INode {
|
||||
throw new Error(`BAD REQUEST: summarizer ${summarizerPromptName} is invalid for this account.`)
|
||||
}
|
||||
|
||||
// Reorder citations in summary and create the list of returned source documents
|
||||
rawSummarizedText = result.responseSet[0].summary[0]?.text
|
||||
|
||||
let summarizedText = reorderCitations(rawSummarizedText)
|
||||
let summaryResponses = applyCitationOrder(responses, rawSummarizedText)
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
import { flatten } from 'lodash'
|
||||
import { VectaraStore, VectaraLibArgs, VectaraFilter, VectaraContextConfig, VectaraFile } from 'langchain/vectorstores/vectara'
|
||||
import { VectaraStore, VectaraLibArgs, VectaraFilter, VectaraContextConfig, VectaraFile, MMRConfig } from 'langchain/vectorstores/vectara'
|
||||
import { Document } from 'langchain/document'
|
||||
import { Embeddings } from 'langchain/embeddings/base'
|
||||
import { ICommonObject, INode, INodeData, INodeOutputsValue, INodeParams } from '../../../src/Interface'
|
||||
@@ -22,7 +22,7 @@ class Vectara_VectorStores implements INode {
|
||||
constructor() {
|
||||
this.label = 'Vectara'
|
||||
this.name = 'vectara'
|
||||
this.version = 1.0
|
||||
this.version = 2.0
|
||||
this.type = 'Vectara'
|
||||
this.icon = 'vectara.png'
|
||||
this.category = 'Vector Stores'
|
||||
@@ -82,7 +82,9 @@ class Vectara_VectorStores implements INode {
|
||||
label: 'Lambda',
|
||||
name: 'lambda',
|
||||
description:
|
||||
'Improves retrieval accuracy by adjusting the balance (from 0 to 1) between neural search and keyword-based search factors.',
|
||||
'Enable hybrid search to improve retrieval accuracy by adjusting the balance (from 0 to 1) between neural search and keyword-based search factors.' +
|
||||
'A value of 0.0 means that only neural search is used, while a value of 1.0 means that only keyword-based search is used. Defaults to 0.0 (neural only).',
|
||||
default: 0.0,
|
||||
type: 'number',
|
||||
additionalParams: true,
|
||||
optional: true
|
||||
@@ -90,8 +92,30 @@ class Vectara_VectorStores implements INode {
|
||||
{
|
||||
label: 'Top K',
|
||||
name: 'topK',
|
||||
description: 'Number of top results to fetch. Defaults to 4',
|
||||
placeholder: '4',
|
||||
description: 'Number of top results to fetch. Defaults to 5',
|
||||
placeholder: '5',
|
||||
type: 'number',
|
||||
additionalParams: true,
|
||||
optional: true
|
||||
},
|
||||
{
|
||||
label: 'MMR K',
|
||||
name: 'mmrK',
|
||||
description: 'Number of top results to fetch for MMR. Defaults to 50',
|
||||
placeholder: '50',
|
||||
type: 'number',
|
||||
additionalParams: true,
|
||||
optional: true
|
||||
},
|
||||
{
|
||||
label: 'MMR diversity bias',
|
||||
name: 'mmrDiversityBias',
|
||||
step: 0.1,
|
||||
description:
|
||||
'The diversity bias to use for MMR. This is a value between 0.0 and 1.0' +
|
||||
'Values closer to 1.0 optimize for the most diverse results.' +
|
||||
'Defaults to 0 (MMR disabled)',
|
||||
placeholder: '0.0',
|
||||
type: 'number',
|
||||
additionalParams: true,
|
||||
optional: true
|
||||
@@ -191,7 +215,9 @@ class Vectara_VectorStores implements INode {
|
||||
const lambda = nodeData.inputs?.lambda as number
|
||||
const output = nodeData.outputs?.output as string
|
||||
const topK = nodeData.inputs?.topK as string
|
||||
const k = topK ? parseFloat(topK) : 4
|
||||
const k = topK ? parseFloat(topK) : 5
|
||||
const mmrK = nodeData.inputs?.mmrK as number
|
||||
const mmrDiversityBias = nodeData.inputs?.mmrDiversityBias as number
|
||||
|
||||
const vectaraArgs: VectaraLibArgs = {
|
||||
apiKey: apiKey,
|
||||
@@ -208,6 +234,11 @@ class Vectara_VectorStores implements INode {
|
||||
if (sentencesBefore) vectaraContextConfig.sentencesBefore = sentencesBefore
|
||||
if (sentencesAfter) vectaraContextConfig.sentencesAfter = sentencesAfter
|
||||
vectaraFilter.contextConfig = vectaraContextConfig
|
||||
const mmrConfig: MMRConfig = {}
|
||||
mmrConfig.enabled = mmrDiversityBias > 0
|
||||
mmrConfig.mmrTopK = mmrK
|
||||
mmrConfig.diversityBias = mmrDiversityBias
|
||||
vectaraFilter.mmrConfig = mmrConfig
|
||||
|
||||
const vectorStore = new VectaraStore(vectaraArgs)
|
||||
|
||||
|
||||
@@ -49,7 +49,7 @@
|
||||
"express": "^4.17.3",
|
||||
"faiss-node": "^0.2.2",
|
||||
"form-data": "^4.0.0",
|
||||
"google-auth-library": "^9.0.0",
|
||||
"google-auth-library": "^9.4.0",
|
||||
"graphql": "^16.6.0",
|
||||
"html-to-text": "^9.0.5",
|
||||
"husky": "^8.0.3",
|
||||
|
||||
@@ -361,12 +361,33 @@
|
||||
{
|
||||
"label": "Top K",
|
||||
"name": "topK",
|
||||
"description": "Number of top results to fetch. Defaults to 4",
|
||||
"placeholder": "4",
|
||||
"description": "Number of top results to fetch. Defaults to 5",
|
||||
"placeholder": "5",
|
||||
"type": "number",
|
||||
"additionalParams": true,
|
||||
"optional": true,
|
||||
"id": "vectara_0-input-topK-number"
|
||||
},
|
||||
{
|
||||
"label": "MMR K",
|
||||
"name": "mmrK",
|
||||
"description": "The number of results to rerank if MMR is enabled.",
|
||||
"placeholder": "50",
|
||||
"type": "number",
|
||||
"additionalParams": true,
|
||||
"optional": true,
|
||||
"id": "vectara_0-input-mmrK-number"
|
||||
},
|
||||
{
|
||||
"label": "MMR Diversity Bias",
|
||||
"name": "mmrDiversityBias",
|
||||
"step": 0.1,
|
||||
"description": "Diversity Bias parameter for MMR, if enabled. 0.0 means no diversiry bias, 1.0 means maximum diversity bias. Defaults to 0.0 (MMR disabled).",
|
||||
"placeholder": "0.0",
|
||||
"type": "number",
|
||||
"additionalParams": true,
|
||||
"optional": true,
|
||||
"id": "vectara_0-input-mmrDiversityBias-number"
|
||||
}
|
||||
],
|
||||
"inputAnchors": [
|
||||
@@ -385,7 +406,9 @@
|
||||
"sentencesBefore": "",
|
||||
"sentencesAfter": "",
|
||||
"lambda": "",
|
||||
"topK": ""
|
||||
"topK": "",
|
||||
"mmrK": "",
|
||||
"mmrDiversityBias": ""
|
||||
},
|
||||
"outputAnchors": [
|
||||
{
|
||||
|
||||
Reference in New Issue
Block a user