[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
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
10 changed files with 535 additions and 204 deletions

View File

@ -2,6 +2,10 @@ import { ICommonObject, INode, INodeData, INodeParams } from '../../../src/Inter
import { getBaseClasses, getCredentialData, getCredentialParam } from '../../../src/utils' import { getBaseClasses, getCredentialData, getCredentialParam } from '../../../src/utils'
import { ListKeyOptions, RecordManagerInterface, UpdateOptions } from '@langchain/community/indexes/base' import { ListKeyOptions, RecordManagerInterface, UpdateOptions } from '@langchain/community/indexes/base'
import { DataSource, QueryRunner } from 'typeorm' import { DataSource, QueryRunner } from 'typeorm'
import { getHost } from '../../vectorstores/Postgres/utils'
import { getDatabase, getPort, getTableName } from './utils'
const serverCredentialsExists = !!process.env.POSTGRES_RECORDMANAGER_USER && !!process.env.POSTGRES_RECORDMANAGER_PASSWORD
class PostgresRecordManager_RecordManager implements INode { class PostgresRecordManager_RecordManager implements INode {
label: string label: string
@ -29,18 +33,22 @@ class PostgresRecordManager_RecordManager implements INode {
{ {
label: 'Host', label: 'Host',
name: 'host', name: 'host',
type: 'string' type: 'string',
placeholder: getHost(),
optional: !!getHost()
}, },
{ {
label: 'Database', label: 'Database',
name: 'database', name: 'database',
type: 'string' type: 'string',
placeholder: getDatabase(),
optional: !!getDatabase()
}, },
{ {
label: 'Port', label: 'Port',
name: 'port', name: 'port',
type: 'number', type: 'number',
placeholder: '5432', placeholder: getPort(),
optional: true optional: true
}, },
{ {
@ -54,7 +62,7 @@ class PostgresRecordManager_RecordManager implements INode {
label: 'Table Name', label: 'Table Name',
name: 'tableName', name: 'tableName',
type: 'string', type: 'string',
placeholder: 'upsertion_records', placeholder: getTableName(),
additionalParams: true, additionalParams: true,
optional: true optional: true
}, },
@ -110,16 +118,16 @@ class PostgresRecordManager_RecordManager implements INode {
label: 'Connect Credential', label: 'Connect Credential',
name: 'credential', name: 'credential',
type: 'credential', type: 'credential',
credentialNames: ['PostgresApi'] credentialNames: ['PostgresApi'],
optional: serverCredentialsExists
} }
} }
async init(nodeData: INodeData, _: string, options: ICommonObject): Promise<any> { async init(nodeData: INodeData, _: string, options: ICommonObject): Promise<any> {
const credentialData = await getCredentialData(nodeData.credential ?? '', options) const credentialData = await getCredentialData(nodeData.credential ?? '', options)
const user = getCredentialParam('user', credentialData, nodeData) const user = getCredentialParam('user', credentialData, nodeData, process.env.POSTGRES_RECORDMANAGER_USER)
const password = getCredentialParam('password', credentialData, nodeData) const password = getCredentialParam('password', credentialData, nodeData, process.env.POSTGRES_RECORDMANAGER_PASSWORD)
const _tableName = nodeData.inputs?.tableName as string const tableName = getTableName(nodeData)
const tableName = _tableName ? _tableName : 'upsertion_records'
const additionalConfig = nodeData.inputs?.additionalConfig as string const additionalConfig = nodeData.inputs?.additionalConfig as string
const _namespace = nodeData.inputs?.namespace as string const _namespace = nodeData.inputs?.namespace as string
const namespace = _namespace ? _namespace : options.chatflowid const namespace = _namespace ? _namespace : options.chatflowid
@ -139,11 +147,11 @@ class PostgresRecordManager_RecordManager implements INode {
const postgresConnectionOptions = { const postgresConnectionOptions = {
...additionalConfiguration, ...additionalConfiguration,
type: 'postgres', type: 'postgres',
host: nodeData.inputs?.host as string, host: getHost(nodeData),
port: nodeData.inputs?.port as number, port: getPort(nodeData),
username: user, username: user,
password: password, password: password,
database: nodeData.inputs?.database as string database: getDatabase(nodeData)
} }
const args = { const args = {
@ -162,7 +170,7 @@ class PostgresRecordManager_RecordManager implements INode {
type PostgresRecordManagerOptions = { type PostgresRecordManagerOptions = {
postgresConnectionOptions: any postgresConnectionOptions: any
tableName?: string tableName: string
} }
class PostgresRecordManager implements RecordManagerInterface { class PostgresRecordManager implements RecordManagerInterface {
@ -180,7 +188,7 @@ class PostgresRecordManager implements RecordManagerInterface {
const { postgresConnectionOptions, tableName } = config const { postgresConnectionOptions, tableName } = config
this.namespace = namespace this.namespace = namespace
this.datasource = new DataSource(postgresConnectionOptions) this.datasource = new DataSource(postgresConnectionOptions)
this.tableName = tableName || 'upsertion_records' this.tableName = tableName
} }
async createSchema(): Promise<void> { async createSchema(): Promise<void> {

View File

@ -0,0 +1,18 @@
# Postgres Record Manager
Postgres Record Manager integration for Flowise
## 🌱 Env Variables
| Variable | Description | Type | Default |
| ---------------------------- | ----------------------------------------------------------------------------------------------- | ------------------------------------------------ | ----------------------------------- |
| POSTGRES_RECORDMANAGER_HOST | Default `host` for Postgres Record Manager | String | |
| POSTGRES_RECORDMANAGER_PORT | Default `port` for Postgres Record Manager | Number | 5432 |
| POSTGRES_RECORDMANAGER_USER | Default `user` for Postgres Record Manager | String | |
| POSTGRES_RECORDMANAGER_PASSWORD | Default `password` for Postgres Record Manager | String | |
| POSTGRES_RECORDMANAGER_DATABASE | Default `database` for Postgres Record Manager | String | |
| POSTGRES_RECORDMANAGER_TABLE_NAME | Default `tableName` for Postgres Record Manager | String | upsertion_records |
## License
Source code in this repository is made available under the [Apache License Version 2.0](https://github.com/FlowiseAI/Flowise/blob/master/LICENSE.md).

View File

@ -0,0 +1,17 @@
import { defaultChain, INodeData } from '../../../src'
export function getHost(nodeData?: INodeData) {
return defaultChain(nodeData?.inputs?.host, process.env.POSTGRES_RECORDMANAGER_HOST)
}
export function getDatabase(nodeData?: INodeData) {
return defaultChain(nodeData?.inputs?.database, process.env.POSTGRES_RECORDMANAGER_DATABASE)
}
export function getPort(nodeData?: INodeData) {
return defaultChain(nodeData?.inputs?.port, process.env.POSTGRES_RECORDMANAGER_PORT, '5432')
}
export function getTableName(nodeData?: INodeData) {
return defaultChain(nodeData?.inputs?.tableName, process.env.POSTGRES_RECORDMANAGER_TABLE_NAME, 'upsertion_records')
}

View File

@ -1,13 +1,16 @@
import { Pool } from 'pg'
import { flatten } from 'lodash' import { flatten } from 'lodash'
import { DataSourceOptions } from 'typeorm'
import { Embeddings } from '@langchain/core/embeddings'
import { Document } from '@langchain/core/documents' 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 { 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 { index } from '../../../src/indexing'
import { howToUseFileUpload } from '../VectorStoreUtils' 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 { class Postgres_VectorStores implements INode {
label: string label: string
@ -26,7 +29,7 @@ class Postgres_VectorStores implements INode {
constructor() { constructor() {
this.label = 'Postgres' this.label = 'Postgres'
this.name = 'postgres' this.name = 'postgres'
this.version = 6.0 this.version = 7.0
this.type = 'Postgres' this.type = 'Postgres'
this.icon = 'postgres.svg' this.icon = 'postgres.svg'
this.category = 'Vector Stores' this.category = 'Vector Stores'
@ -36,7 +39,8 @@ class Postgres_VectorStores implements INode {
label: 'Connect Credential', label: 'Connect Credential',
name: 'credential', name: 'credential',
type: 'credential', type: 'credential',
credentialNames: ['PostgresApi'] credentialNames: ['PostgresApi'],
optional: serverCredentialsExists
} }
this.inputs = [ this.inputs = [
{ {
@ -61,28 +65,74 @@ class Postgres_VectorStores implements INode {
{ {
label: 'Host', label: 'Host',
name: 'host', name: 'host',
type: 'string' type: 'string',
placeholder: getHost(),
optional: !!getHost()
}, },
{ {
label: 'Database', label: 'Database',
name: 'database', name: 'database',
type: 'string' type: 'string',
placeholder: getDatabase(),
optional: !!getDatabase()
}, },
{ {
label: 'Port', label: 'Port',
name: 'port', name: 'port',
type: 'number', type: 'number',
placeholder: '5432', placeholder: getPort(),
optional: true optional: true
}, },
{ {
label: 'Table Name', label: 'Table Name',
name: 'tableName', name: 'tableName',
type: 'string', type: 'string',
placeholder: 'documents', placeholder: getTableName(),
additionalParams: true, additionalParams: true,
optional: 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', label: 'File Upload',
name: 'fileUpload', name: 'fileUpload',
@ -117,6 +167,15 @@ class Postgres_VectorStores implements INode {
type: 'json', type: 'json',
additionalParams: true, additionalParams: true,
optional: 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 = [ this.outputs = [
@ -128,7 +187,7 @@ class Postgres_VectorStores implements INode {
{ {
label: 'Postgres Vector Store', label: 'Postgres Vector Store',
name: 'vectorStore', name: 'vectorStore',
baseClasses: [this.type, ...getBaseClasses(TypeORMVectorStore)] baseClasses: [this.type, ...getBaseClasses(VectorStore)]
} }
] ]
} }
@ -136,43 +195,15 @@ class Postgres_VectorStores implements INode {
//@ts-ignore //@ts-ignore
vectorStoreMethods = { vectorStoreMethods = {
async upsert(nodeData: INodeData, options: ICommonObject): Promise<Partial<IndexingResult>> { async upsert(nodeData: INodeData, options: ICommonObject): Promise<Partial<IndexingResult>> {
const credentialData = await getCredentialData(nodeData.credential ?? '', options) const tableName = getTableName(nodeData)
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 docs = nodeData.inputs?.document as Document[] 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 recordManager = nodeData.inputs?.recordManager
const isFileUploadEnabled = nodeData.inputs?.fileUpload as boolean const isFileUploadEnabled = nodeData.inputs?.fileUpload as boolean
const vectorStoreDriver: VectorStoreDriver = Postgres_VectorStores.getDriverFromConfig(nodeData, options)
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 flattenDocs = docs && docs.length ? flatten(docs) : [] const flattenDocs = docs && docs.length ? flatten(docs) : []
const finalDocs = [] const finalDocs = []
for (let i = 0; i < flattenDocs.length; i += 1) { for (let i = 0; i < flattenDocs.length; i += 1) {
if (flattenDocs[i] && flattenDocs[i].pageContent) { if (flattenDocs[i] && flattenDocs[i].pageContent) {
if (isFileUploadEnabled && options.chatId) { if (isFileUploadEnabled && options.chatId) {
@ -184,24 +215,7 @@ class Postgres_VectorStores implements INode {
try { try {
if (recordManager) { if (recordManager) {
const vectorStore = await TypeORMVectorStore.fromDataSource(embeddings, args) const vectorStore = await vectorStoreDriver.instanciate()
// 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')
}
}
}
await recordManager.createSchema() await recordManager.createSchema()
@ -218,12 +232,7 @@ class Postgres_VectorStores implements INode {
return res return res
} else { } else {
const vectorStore = await TypeORMVectorStore.fromDocuments(finalDocs, embeddings, args) await vectorStoreDriver.fromDocuments(finalDocs)
// Avoid Illegal invocation error
vectorStore.similaritySearchVectorWithScore = async (query: number[], k: number, filter?: any) => {
return await similaritySearchVectorWithScore(query, k, tableName, postgresConnectionOptions, filter)
}
return { numAdded: finalDocs.length, addedDocs: 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> { async delete(nodeData: INodeData, ids: string[], options: ICommonObject): Promise<void> {
const credentialData = await getCredentialData(nodeData.credential ?? '', options) const vectorStoreDriver: VectorStoreDriver = Postgres_VectorStores.getDriverFromConfig(nodeData, options)
const user = getCredentialParam('user', credentialData, nodeData) const tableName = getTableName(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 recordManager = nodeData.inputs?.recordManager const recordManager = nodeData.inputs?.recordManager
let additionalConfiguration = {} const vectorStore = await vectorStoreDriver.instanciate()
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)
try { try {
if (recordManager) { if (recordManager) {
@ -286,13 +266,7 @@ class Postgres_VectorStores implements INode {
} }
async init(nodeData: INodeData, _: string, options: ICommonObject): Promise<any> { async init(nodeData: INodeData, _: string, options: ICommonObject): Promise<any> {
const credentialData = await getCredentialData(nodeData.credential ?? '', options) const vectorStoreDriver: VectorStoreDriver = Postgres_VectorStores.getDriverFromConfig(nodeData, 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 output = nodeData.outputs?.output as string const output = nodeData.outputs?.output as string
const topK = nodeData.inputs?.topK as string const topK = nodeData.inputs?.topK as string
const k = topK ? parseFloat(topK) : 4 const k = topK ? parseFloat(topK) : 4
@ -304,50 +278,13 @@ class Postgres_VectorStores implements INode {
pgMetadataFilter = typeof _pgMetadataFilter === 'object' ? _pgMetadataFilter : JSON.parse(_pgMetadataFilter) pgMetadataFilter = typeof _pgMetadataFilter === 'object' ? _pgMetadataFilter : JSON.parse(_pgMetadataFilter)
} }
if (isFileUploadEnabled && options.chatId) { if (isFileUploadEnabled && options.chatId) {
pgMetadataFilter = pgMetadataFilter || {}
pgMetadataFilter = { pgMetadataFilter = {
...pgMetadataFilter, ...(pgMetadataFilter || {}),
[FLOWISE_CHATID]: options.chatId, [FLOWISE_CHATID]: options.chatId
$notexists: FLOWISE_CHATID // special filter to check if the field does not exist
} }
} }
let additionalConfiguration = {} const vectorStore = await vectorStoreDriver.instanciate(pgMetadataFilter)
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)
}
if (output === 'retriever') { if (output === 'retriever') {
const retriever = vectorStore.asRetriever(k) const retriever = vectorStore.asRetriever(k)
@ -361,51 +298,17 @@ class Postgres_VectorStores implements INode {
} }
return vectorStore return vectorStore
} }
}
const similaritySearchVectorWithScore = async ( static getDriverFromConfig(nodeData: INodeData, options: ICommonObject): VectorStoreDriver {
query: number[], switch (nodeData.inputs?.driver) {
k: number, case 'typeorm':
tableName: string, return new TypeORMDriver(nodeData, options)
postgresConnectionOptions: ICommonObject, case 'pgvector':
filter?: any return new PGVectorDriver(nodeData, options)
) => { default:
const embeddingString = `[${query.join(',')}]` return new TypeORMDriver(nodeData, options)
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])
} }
} }
return results
} }
module.exports = { nodeClass: Postgres_VectorStores } module.exports = { nodeClass: Postgres_VectorStores }

View File

@ -0,0 +1,19 @@
# Postgres Vector Store
Postgres Vector Store integration for Flowise
## 🌱 Env Variables
| Variable | Description | Type | Default |
| ---------------------------- | ----------------------------------------------------------------------------------------------- | ------------------------------------------------ | ----------------------------------- |
| POSTGRES_VECTORSTORE_HOST | Default `host` for Postgres Vector Store | String | |
| POSTGRES_VECTORSTORE_PORT | Default `port` for Postgres Vector Store | Number | 5432 |
| POSTGRES_VECTORSTORE_USER | Default `user` for Postgres Vector Store | String | |
| POSTGRES_VECTORSTORE_PASSWORD | Default `password` for Postgres Vector Store | String | |
| POSTGRES_VECTORSTORE_DATABASE | Default `database` for Postgres Vector Store | String | |
| POSTGRES_VECTORSTORE_TABLE_NAME | Default `tableName` for Postgres Vector Store | String | documents |
| POSTGRES_VECTORSTORE_CONTENT_COLUMN_NAME | Default `contentColumnName` for Postgres Vector Store | String | pageContent |
## License
Source code in this repository is made available under the [Apache License Version 2.0](https://github.com/FlowiseAI/Flowise/blob/master/LICENSE.md).

View File

@ -0,0 +1,48 @@
import { VectorStore } from '@langchain/core/vectorstores'
import { getCredentialData, getCredentialParam, ICommonObject, INodeData } from '../../../../src'
import { Document } from '@langchain/core/documents'
import { Embeddings } from '@langchain/core/embeddings'
import { getDatabase, getHost, getPort, getTableName } from '../utils'
export abstract class VectorStoreDriver {
constructor(protected nodeData: INodeData, protected options: ICommonObject) {}
abstract instanciate(metaDataFilters?: any): Promise<VectorStore>
abstract fromDocuments(documents: Document[]): Promise<VectorStore>
protected async adaptInstance(instance: VectorStore, _metaDataFilters?: any): Promise<VectorStore> {
return instance
}
getHost() {
return getHost(this.nodeData) as string
}
getPort() {
return getPort(this.nodeData) as number
}
getDatabase() {
return getDatabase(this.nodeData) as string
}
getTableName() {
return getTableName(this.nodeData)
}
getEmbeddings() {
return this.nodeData.inputs?.embeddings as Embeddings
}
async getCredentials() {
const credentialData = await getCredentialData(this.nodeData.credential ?? '', this.options)
const user = getCredentialParam('user', credentialData, this.nodeData, process.env.POSTGRES_VECTORSTORE_USER)
const password = getCredentialParam('password', credentialData, this.nodeData, process.env.POSTGRES_VECTORSTORE_PASSWORD)
return {
user,
password
}
}
}

View File

@ -0,0 +1,117 @@
import { VectorStoreDriver } from './Base'
import { FLOWISE_CHATID } from '../../../../src'
import { DistanceStrategy, PGVectorStore, PGVectorStoreArgs } from '@langchain/community/vectorstores/pgvector'
import { Document } from '@langchain/core/documents'
import { PoolConfig } from 'pg'
import { getContentColumnName } from '../utils'
export class PGVectorDriver extends VectorStoreDriver {
static CONTENT_COLUMN_NAME_DEFAULT: string = 'pageContent'
protected _postgresConnectionOptions: PoolConfig
protected async getPostgresConnectionOptions() {
if (!this._postgresConnectionOptions) {
const { user, password } = await this.getCredentials()
const additionalConfig = this.nodeData.inputs?.additionalConfig as string
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)
}
}
this._postgresConnectionOptions = {
...additionalConfiguration,
host: this.getHost(),
port: this.getPort(),
user: user,
password: password,
database: this.getDatabase()
}
}
return this._postgresConnectionOptions
}
async getArgs(): Promise<PGVectorStoreArgs> {
return {
postgresConnectionOptions: await this.getPostgresConnectionOptions(),
tableName: this.getTableName(),
columns: {
contentColumnName: getContentColumnName(this.nodeData)
},
distanceStrategy: (this.nodeData.inputs?.distanceStrategy || 'cosine') as DistanceStrategy
}
}
async instanciate(metadataFilters?: any) {
return this.adaptInstance(await PGVectorStore.initialize(this.getEmbeddings(), await this.getArgs()), metadataFilters)
}
async fromDocuments(documents: Document[]) {
const instance = await this.instanciate()
await instance.addDocuments(documents)
return this.adaptInstance(instance)
}
protected async adaptInstance(instance: PGVectorStore, metadataFilters?: any): Promise<PGVectorStore> {
const { [FLOWISE_CHATID]: chatId, ...pgMetadataFilter } = metadataFilters || {}
const baseSimilaritySearchVectorWithScoreFn = instance.similaritySearchVectorWithScore.bind(instance)
instance.similaritySearchVectorWithScore = async (query, k, filter) => {
return await baseSimilaritySearchVectorWithScoreFn(query, k, filter ?? pgMetadataFilter)
}
const basePoolQueryFn = instance.pool.query.bind(instance.pool)
// @ts-ignore
instance.pool.query = async (queryString: string, parameters: any[]) => {
if (!instance.client) {
instance.client = await instance.pool.connect()
}
const whereClauseRegex = /WHERE ([^\n]+)/
let chatflowOr = ''
// Match chatflow uploaded file and keep filtering on other files:
// https://github.com/FlowiseAI/Flowise/pull/3367#discussion_r1804229295
if (chatId) {
parameters.push({ [FLOWISE_CHATID]: chatId })
chatflowOr = `OR metadata @> $${parameters.length}`
}
if (queryString.match(whereClauseRegex)) {
queryString = queryString.replace(whereClauseRegex, `WHERE (($1) AND NOT (metadata ? '${FLOWISE_CHATID}')) ${chatflowOr}`)
} else {
const orderByClauseRegex = /ORDER BY (.*)/
// Insert WHERE clause before ORDER BY
queryString = queryString.replace(
orderByClauseRegex,
`WHERE (metadata @> '{}' AND NOT (metadata ? '${FLOWISE_CHATID}')) ${chatflowOr}
ORDER BY $1
`
)
}
// Run base function
const queryResult = await basePoolQueryFn(queryString, parameters)
// ensure connection is released
instance.client.release()
instance.client = undefined
return queryResult
}
return instance
}
}

View File

@ -0,0 +1,169 @@
import { DataSourceOptions } from 'typeorm'
import { VectorStoreDriver } from './Base'
import { FLOWISE_CHATID, ICommonObject } from '../../../../src'
import { TypeORMVectorStore, TypeORMVectorStoreArgs, TypeORMVectorStoreDocument } from '@langchain/community/vectorstores/typeorm'
import { VectorStore } from '@langchain/core/vectorstores'
import { Document } from '@langchain/core/documents'
import { Pool } from 'pg'
export class TypeORMDriver extends VectorStoreDriver {
protected _postgresConnectionOptions: DataSourceOptions
protected async getPostgresConnectionOptions() {
if (!this._postgresConnectionOptions) {
const { user, password } = await this.getCredentials()
const additionalConfig = this.nodeData.inputs?.additionalConfig as string
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)
}
}
this._postgresConnectionOptions = {
...additionalConfiguration,
type: 'postgres',
host: this.getHost(),
port: this.getPort(),
username: user, // Required by TypeORMVectorStore
user: user, // Required by Pool in similaritySearchVectorWithScore
password: password,
database: this.getDatabase()
} as DataSourceOptions
}
return this._postgresConnectionOptions
}
async getArgs(): Promise<TypeORMVectorStoreArgs> {
return {
postgresConnectionOptions: await this.getPostgresConnectionOptions(),
tableName: this.getTableName()
}
}
async instanciate(metadataFilters?: any) {
return this.adaptInstance(await TypeORMVectorStore.fromDataSource(this.getEmbeddings(), await this.getArgs()), metadataFilters)
}
async fromDocuments(documents: Document[]) {
return this.adaptInstance(await TypeORMVectorStore.fromDocuments(documents, this.getEmbeddings(), await this.getArgs()))
}
sanitizeDocuments(documents: Document[]) {
// Remove NULL characters which triggers error on PG
for (var i in documents) {
documents[i].pageContent = documents[i].pageContent.replace(/\0/g, '')
}
return documents
}
protected async adaptInstance(instance: TypeORMVectorStore, metadataFilters?: any): Promise<VectorStore> {
const tableName = this.getTableName()
// 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)
*/
instance.similaritySearchVectorWithScore = async (query: number[], k: number, filter?: any) => {
return await TypeORMDriver.similaritySearchVectorWithScore(
query,
k,
tableName,
await this.getPostgresConnectionOptions(),
filter ?? metadataFilters,
this.computedOperatorString
)
}
instance.delete = async (params: { ids: string[] }): Promise<void> => {
const { ids } = params
if (ids?.length) {
try {
instance.appDataSource.getRepository(instance.documentEntity).delete(ids)
} catch (e) {
console.error('Failed to delete')
}
}
}
const baseAddVectorsFn = instance.addVectors.bind(instance)
instance.addVectors = async (vectors, documents) => {
return baseAddVectorsFn(vectors, this.sanitizeDocuments(documents))
}
return instance
}
get computedOperatorString() {
const { distanceStrategy = 'cosine' } = this.nodeData.inputs || {}
switch (distanceStrategy) {
case 'cosine':
return '<=>'
case 'innerProduct':
return '<#>'
case 'euclidean':
return '<->'
default:
throw new Error(`Unknown distance strategy: ${distanceStrategy}`)
}
}
static similaritySearchVectorWithScore = async (
query: number[],
k: number,
tableName: string,
postgresConnectionOptions: ICommonObject,
filter?: any,
distanceOperator: string = '<=>'
) => {
const embeddingString = `[${query.join(',')}]`
let chatflowOr = ''
const { [FLOWISE_CHATID]: chatId, ...restFilters } = filter || {}
const _filter = JSON.stringify(restFilters || {})
const parameters: any[] = [embeddingString, _filter, k]
// Match chatflow uploaded file and keep filtering on other files:
// https://github.com/FlowiseAI/Flowise/pull/3367#discussion_r1804229295
if (chatId) {
parameters.push({ [FLOWISE_CHATID]: chatId })
chatflowOr = `OR metadata @> $${parameters.length}`
}
const queryString = `
SELECT *, embedding ${distanceOperator} $1 as "_distance"
FROM ${tableName}
WHERE ((metadata @> $2) AND NOT (metadata ? '${FLOWISE_CHATID}')) ${chatflowOr}
ORDER BY "_distance" ASC
LIMIT $3;`
const pool = new Pool(postgresConnectionOptions)
const conn = await pool.connect()
const documents = await conn.query(queryString, parameters)
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])
}
}
return results
}
}

View File

@ -0,0 +1,21 @@
import { defaultChain, INodeData } from '../../../src'
export function getHost(nodeData?: INodeData) {
return defaultChain(nodeData?.inputs?.host, process.env.POSTGRES_VECTORSTORE_HOST)
}
export function getDatabase(nodeData?: INodeData) {
return defaultChain(nodeData?.inputs?.database, process.env.POSTGRES_VECTORSTORE_DATABASE)
}
export function getPort(nodeData?: INodeData) {
return defaultChain(nodeData?.inputs?.port, process.env.POSTGRES_VECTORSTORE_PORT, '5432')
}
export function getTableName(nodeData?: INodeData) {
return defaultChain(nodeData?.inputs?.tableName, process.env.POSTGRES_VECTORSTORE_TABLE_NAME, 'documents')
}
export function getContentColumnName(nodeData?: INodeData) {
return defaultChain(nodeData?.inputs?.contentColumnName, process.env.POSTGRES_VECTORSTORE_CONTENT_COLUMN_NAME, 'pageContent')
}

View File

@ -542,8 +542,19 @@ export const getCredentialData = async (selectedCredentialId: string, options: I
} }
} }
export const getCredentialParam = (paramName: string, credentialData: ICommonObject, nodeData: INodeData): any => { /**
return (nodeData.inputs as ICommonObject)[paramName] ?? credentialData[paramName] ?? undefined * Get first non falsy value
*
* @param {...any} values
*
* @returns {any|undefined}
*/
export const defaultChain = (...values: any[]): any | undefined => {
return values.filter(Boolean)[0]
}
export const getCredentialParam = (paramName: string, credentialData: ICommonObject, nodeData: INodeData, defaultValue?: any): any => {
return (nodeData.inputs as ICommonObject)[paramName] ?? credentialData[paramName] ?? defaultValue ?? undefined
} }
// reference https://www.freeformatter.com/json-escape.html // reference https://www.freeformatter.com/json-escape.html