[Prostgres Vector Store] Add PGVector Driver option + Fix null character issue w/ TypeORM Driver (#3367)

* Add PGVector Driver option + Fix null character issue w/ TypeORM Driver

* Handle file upload case with PGVector

* Cleanup

* Fix data filtering for chatflow uploaded files

* Add distanceStrategy parameter

* Fix query to improve chatflow uploaded files filtering

* Ensure PGVector release connections

* Await client connected

* Make Postgres credentials optionnal when set on env variables

* Document env variables in nodes directories

* Prevent reuse client

* Fix empty metadataFilter

* Update CONTRIBUTING.md

* Update Postgres.ts

---------

Co-authored-by: Henry Heng <henryheng@flowiseai.com>
This commit is contained in:
Jérémy JOURDIN
2024-11-01 19:13:45 +01:00
committed by GitHub
parent 39380a4bc7
commit 15d59a9052
10 changed files with 535 additions and 204 deletions
@@ -1,13 +1,16 @@
import { Pool } from 'pg'
import { flatten } from 'lodash'
import { DataSourceOptions } from 'typeorm'
import { Embeddings } from '@langchain/core/embeddings'
import { Document } from '@langchain/core/documents'
import { TypeORMVectorStore, TypeORMVectorStoreDocument } from '@langchain/community/vectorstores/typeorm'
import { ICommonObject, INode, INodeData, INodeOutputsValue, INodeParams, IndexingResult } from '../../../src/Interface'
import { FLOWISE_CHATID, getBaseClasses, getCredentialData, getCredentialParam } from '../../../src/utils'
import { FLOWISE_CHATID, getBaseClasses } from '../../../src/utils'
import { index } from '../../../src/indexing'
import { howToUseFileUpload } from '../VectorStoreUtils'
import { VectorStore } from '@langchain/core/vectorstores'
import { VectorStoreDriver } from './driver/Base'
import { TypeORMDriver } from './driver/TypeORM'
import { PGVectorDriver } from './driver/PGVector'
import { getContentColumnName, getDatabase, getHost, getPort, getTableName } from './utils'
const serverCredentialsExists = !!process.env.POSTGRES_VECTORSTORE_USER && !!process.env.POSTGRES_VECTORSTORE_PASSWORD
class Postgres_VectorStores implements INode {
label: string
@@ -26,7 +29,7 @@ class Postgres_VectorStores implements INode {
constructor() {
this.label = 'Postgres'
this.name = 'postgres'
this.version = 6.0
this.version = 7.0
this.type = 'Postgres'
this.icon = 'postgres.svg'
this.category = 'Vector Stores'
@@ -36,7 +39,8 @@ class Postgres_VectorStores implements INode {
label: 'Connect Credential',
name: 'credential',
type: 'credential',
credentialNames: ['PostgresApi']
credentialNames: ['PostgresApi'],
optional: serverCredentialsExists
}
this.inputs = [
{
@@ -61,28 +65,74 @@ class Postgres_VectorStores implements INode {
{
label: 'Host',
name: 'host',
type: 'string'
type: 'string',
placeholder: getHost(),
optional: !!getHost()
},
{
label: 'Database',
name: 'database',
type: 'string'
type: 'string',
placeholder: getDatabase(),
optional: !!getDatabase()
},
{
label: 'Port',
name: 'port',
type: 'number',
placeholder: '5432',
placeholder: getPort(),
optional: true
},
{
label: 'Table Name',
name: 'tableName',
type: 'string',
placeholder: 'documents',
placeholder: getTableName(),
additionalParams: true,
optional: true
},
{
label: 'Driver',
name: 'driver',
type: 'options',
default: 'typeorm',
description: 'Different option to connect to Postgres',
options: [
{
label: 'TypeORM',
name: 'typeorm'
},
{
label: 'PGVector',
name: 'pgvector'
}
],
optional: true,
additionalParams: true
},
{
label: 'Distance Strategy',
name: 'distanceStrategy',
description: 'Strategy for calculating distances between vectors',
type: 'options',
options: [
{
label: 'Cosine',
name: 'cosine'
},
{
label: 'Euclidean',
name: 'euclidean'
},
{
label: 'Inner Product',
name: 'innerProduct'
}
],
additionalParams: true,
default: 'cosine',
optional: true
},
{
label: 'File Upload',
name: 'fileUpload',
@@ -117,6 +167,15 @@ class Postgres_VectorStores implements INode {
type: 'json',
additionalParams: true,
optional: true
},
{
label: 'Content Column Name',
name: 'contentColumnName',
description: 'Column name to store the text content (PGVector Driver only, others use pageContent)',
type: 'string',
placeholder: getContentColumnName(),
additionalParams: true,
optional: true
}
]
this.outputs = [
@@ -128,7 +187,7 @@ class Postgres_VectorStores implements INode {
{
label: 'Postgres Vector Store',
name: 'vectorStore',
baseClasses: [this.type, ...getBaseClasses(TypeORMVectorStore)]
baseClasses: [this.type, ...getBaseClasses(VectorStore)]
}
]
}
@@ -136,43 +195,15 @@ class Postgres_VectorStores implements INode {
//@ts-ignore
vectorStoreMethods = {
async upsert(nodeData: INodeData, options: ICommonObject): Promise<Partial<IndexingResult>> {
const credentialData = await getCredentialData(nodeData.credential ?? '', options)
const user = getCredentialParam('user', credentialData, nodeData)
const password = getCredentialParam('password', credentialData, nodeData)
const _tableName = nodeData.inputs?.tableName as string
const tableName = _tableName ? _tableName : 'documents'
const tableName = getTableName(nodeData)
const docs = nodeData.inputs?.document as Document[]
const embeddings = nodeData.inputs?.embeddings as Embeddings
const additionalConfig = nodeData.inputs?.additionalConfig as string
const recordManager = nodeData.inputs?.recordManager
const isFileUploadEnabled = nodeData.inputs?.fileUpload as boolean
let additionalConfiguration = {}
if (additionalConfig) {
try {
additionalConfiguration = typeof additionalConfig === 'object' ? additionalConfig : JSON.parse(additionalConfig)
} catch (exception) {
throw new Error('Invalid JSON in the Additional Configuration: ' + exception)
}
}
const postgresConnectionOptions = {
...additionalConfiguration,
type: 'postgres',
host: nodeData.inputs?.host as string,
port: nodeData.inputs?.port as number,
username: user,
password: password,
database: nodeData.inputs?.database as string
}
const args = {
postgresConnectionOptions: postgresConnectionOptions as DataSourceOptions,
tableName: tableName
}
const vectorStoreDriver: VectorStoreDriver = Postgres_VectorStores.getDriverFromConfig(nodeData, options)
const flattenDocs = docs && docs.length ? flatten(docs) : []
const finalDocs = []
for (let i = 0; i < flattenDocs.length; i += 1) {
if (flattenDocs[i] && flattenDocs[i].pageContent) {
if (isFileUploadEnabled && options.chatId) {
@@ -184,24 +215,7 @@ class Postgres_VectorStores implements INode {
try {
if (recordManager) {
const vectorStore = await TypeORMVectorStore.fromDataSource(embeddings, args)
// Avoid Illegal invocation error
vectorStore.similaritySearchVectorWithScore = async (query: number[], k: number, filter?: any) => {
return await similaritySearchVectorWithScore(query, k, tableName, postgresConnectionOptions, filter)
}
vectorStore.delete = async (params: { ids: string[] }): Promise<void> => {
const { ids } = params
if (ids?.length) {
try {
vectorStore.appDataSource.getRepository(vectorStore.documentEntity).delete(ids)
} catch (e) {
console.error('Failed to delete')
}
}
}
const vectorStore = await vectorStoreDriver.instanciate()
await recordManager.createSchema()
@@ -218,12 +232,7 @@ class Postgres_VectorStores implements INode {
return res
} else {
const vectorStore = await TypeORMVectorStore.fromDocuments(finalDocs, embeddings, args)
// Avoid Illegal invocation error
vectorStore.similaritySearchVectorWithScore = async (query: number[], k: number, filter?: any) => {
return await similaritySearchVectorWithScore(query, k, tableName, postgresConnectionOptions, filter)
}
await vectorStoreDriver.fromDocuments(finalDocs)
return { numAdded: finalDocs.length, addedDocs: finalDocs }
}
@@ -232,40 +241,11 @@ class Postgres_VectorStores implements INode {
}
},
async delete(nodeData: INodeData, ids: string[], options: ICommonObject): Promise<void> {
const credentialData = await getCredentialData(nodeData.credential ?? '', options)
const user = getCredentialParam('user', credentialData, nodeData)
const password = getCredentialParam('password', credentialData, nodeData)
const _tableName = nodeData.inputs?.tableName as string
const tableName = _tableName ? _tableName : 'documents'
const embeddings = nodeData.inputs?.embeddings as Embeddings
const additionalConfig = nodeData.inputs?.additionalConfig as string
const vectorStoreDriver: VectorStoreDriver = Postgres_VectorStores.getDriverFromConfig(nodeData, options)
const tableName = getTableName(nodeData)
const recordManager = nodeData.inputs?.recordManager
let additionalConfiguration = {}
if (additionalConfig) {
try {
additionalConfiguration = typeof additionalConfig === 'object' ? additionalConfig : JSON.parse(additionalConfig)
} catch (exception) {
throw new Error('Invalid JSON in the Additional Configuration: ' + exception)
}
}
const postgresConnectionOptions = {
...additionalConfiguration,
type: 'postgres',
host: nodeData.inputs?.host as string,
port: nodeData.inputs?.port as number,
username: user,
password: password,
database: nodeData.inputs?.database as string
}
const args = {
postgresConnectionOptions: postgresConnectionOptions as DataSourceOptions,
tableName: tableName
}
const vectorStore = await TypeORMVectorStore.fromDataSource(embeddings, args)
const vectorStore = await vectorStoreDriver.instanciate()
try {
if (recordManager) {
@@ -286,13 +266,7 @@ class Postgres_VectorStores implements INode {
}
async init(nodeData: INodeData, _: string, options: ICommonObject): Promise<any> {
const credentialData = await getCredentialData(nodeData.credential ?? '', options)
const user = getCredentialParam('user', credentialData, nodeData)
const password = getCredentialParam('password', credentialData, nodeData)
const _tableName = nodeData.inputs?.tableName as string
const tableName = _tableName ? _tableName : 'documents'
const embeddings = nodeData.inputs?.embeddings as Embeddings
const additionalConfig = nodeData.inputs?.additionalConfig as string
const vectorStoreDriver: VectorStoreDriver = Postgres_VectorStores.getDriverFromConfig(nodeData, options)
const output = nodeData.outputs?.output as string
const topK = nodeData.inputs?.topK as string
const k = topK ? parseFloat(topK) : 4
@@ -304,50 +278,13 @@ class Postgres_VectorStores implements INode {
pgMetadataFilter = typeof _pgMetadataFilter === 'object' ? _pgMetadataFilter : JSON.parse(_pgMetadataFilter)
}
if (isFileUploadEnabled && options.chatId) {
pgMetadataFilter = pgMetadataFilter || {}
pgMetadataFilter = {
...pgMetadataFilter,
[FLOWISE_CHATID]: options.chatId,
$notexists: FLOWISE_CHATID // special filter to check if the field does not exist
...(pgMetadataFilter || {}),
[FLOWISE_CHATID]: options.chatId
}
}
let additionalConfiguration = {}
if (additionalConfig) {
try {
additionalConfiguration = typeof additionalConfig === 'object' ? additionalConfig : JSON.parse(additionalConfig)
} catch (exception) {
throw new Error('Invalid JSON in the Additional Configuration: ' + exception)
}
}
const postgresConnectionOptions = {
...additionalConfiguration,
type: 'postgres',
host: nodeData.inputs?.host as string,
port: nodeData.inputs?.port as number,
username: user, // Required by TypeORMVectorStore
user: user, // Required by Pool in similaritySearchVectorWithScore
password: password,
database: nodeData.inputs?.database as string
}
const args = {
postgresConnectionOptions: postgresConnectionOptions as DataSourceOptions,
tableName: tableName
}
const vectorStore = await TypeORMVectorStore.fromDataSource(embeddings, args)
// Rewrite the method to use pg pool connection instead of the default connection
/* Otherwise a connection error is displayed when the chain tries to execute the function
[chain/start] [1:chain:ConversationalRetrievalQAChain] Entering Chain run with input: { "question": "what the document is about", "chat_history": [] }
[retriever/start] [1:chain:ConversationalRetrievalQAChain > 2:retriever:VectorStoreRetriever] Entering Retriever run with input: { "query": "what the document is about" }
[ERROR]: uncaughtException: Illegal invocation TypeError: Illegal invocation at Socket.ref (node:net:1524:18) at Connection.ref (.../node_modules/pg/lib/connection.js:183:17) at Client.ref (.../node_modules/pg/lib/client.js:591:21) at BoundPool._pulseQueue (/node_modules/pg-pool/index.js:148:28) at .../node_modules/pg-pool/index.js:184:37 at process.processTicksAndRejections (node:internal/process/task_queues:77:11)
*/
vectorStore.similaritySearchVectorWithScore = async (query: number[], k: number, filter?: any) => {
return await similaritySearchVectorWithScore(query, k, tableName, postgresConnectionOptions, filter ?? pgMetadataFilter)
}
const vectorStore = await vectorStoreDriver.instanciate(pgMetadataFilter)
if (output === 'retriever') {
const retriever = vectorStore.asRetriever(k)
@@ -361,51 +298,17 @@ class Postgres_VectorStores implements INode {
}
return vectorStore
}
}
const similaritySearchVectorWithScore = async (
query: number[],
k: number,
tableName: string,
postgresConnectionOptions: ICommonObject,
filter?: any
) => {
const embeddingString = `[${query.join(',')}]`
let _filter = '{}'
let notExists = ''
if (filter && typeof filter === 'object') {
if (filter.$notexists) {
notExists = `OR NOT (metadata ? '${filter.$notexists}')`
delete filter.$notexists
}
_filter = JSON.stringify(filter)
}
const queryString = `
SELECT *, embedding <=> $1 as "_distance"
FROM ${tableName}
WHERE metadata @> $2
${notExists}
ORDER BY "_distance" ASC
LIMIT $3;`
const pool = new Pool(postgresConnectionOptions)
const conn = await pool.connect()
const documents = await conn.query(queryString, [embeddingString, _filter, k])
conn.release()
const results = [] as [TypeORMVectorStoreDocument, number][]
for (const doc of documents.rows) {
if (doc._distance != null && doc.pageContent != null) {
const document = new Document(doc) as TypeORMVectorStoreDocument
document.id = doc.id
results.push([document, doc._distance])
static getDriverFromConfig(nodeData: INodeData, options: ICommonObject): VectorStoreDriver {
switch (nodeData.inputs?.driver) {
case 'typeorm':
return new TypeORMDriver(nodeData, options)
case 'pgvector':
return new PGVectorDriver(nodeData, options)
default:
return new TypeORMDriver(nodeData, options)
}
}
return results
}
module.exports = { nodeClass: Postgres_VectorStores }