Adding support for LLM Caching.

This commit is contained in:
vinodkiran
2023-09-29 07:54:16 +05:30
parent 113d90e5dc
commit d588ac0480
6 changed files with 87 additions and 5 deletions
@@ -0,0 +1,44 @@
import { INode, INodeData, INodeOutputsValue, INodeParams } from '../../../src/Interface'
import { InMemoryCache } from 'langchain/cache'
import { getBaseClasses } from '../../../src'
class LocalMemoryCache implements INode {
label: string
name: string
version: number
description: string
type: string
icon: string
category: string
baseClasses: string[]
inputs: INodeParams[]
outputs: INodeOutputsValue[]
inMemoryCache: any
constructor() {
this.label = 'Local (Builtin) Cache'
this.name = 'localCache'
this.version = 1.0
this.type = 'LLMCache'
this.icon = 'memorycache.png'
this.category = 'LLM Cache'
this.baseClasses = [this.type, 'LLMCacheBase']
this.inputs = []
this.outputs = [
{
label: 'LLM Cache',
name: 'cache',
baseClasses: [this.type, ...getBaseClasses(InMemoryCache)]
}
]
}
async init(nodeData: INodeData): Promise<any> {
if (!this.inMemoryCache) {
this.inMemoryCache = InMemoryCache.global()
}
return this.inMemoryCache
}
}
module.exports = { nodeClass: LocalMemoryCache }
Binary file not shown.

After

Width:  |  Height:  |  Size: 2.7 KiB

@@ -1,6 +1,8 @@
import { ICommonObject, INode, INodeData, INodeParams } from '../../../src/Interface'
import { getBaseClasses, getCredentialData, getCredentialParam } from '../../../src/utils'
import { OpenAI, OpenAIInput } from 'langchain/llms/openai'
import { BaseLLMParams } from 'langchain/dist/llms/base'
import { BaseCache } from 'langchain/schema'
class OpenAI_LLMs implements INode {
label: string
@@ -17,7 +19,7 @@ class OpenAI_LLMs implements INode {
constructor() {
this.label = 'OpenAI'
this.name = 'openAI'
this.version = 2.0
this.version = 3.0
this.type = 'OpenAI'
this.icon = 'openai.png'
this.category = 'LLMs'
@@ -30,6 +32,12 @@ class OpenAI_LLMs implements INode {
credentialNames: ['openAIApi']
}
this.inputs = [
{
label: 'Cache',
name: 'llmCache',
type: 'LLMCache',
optional: true
},
{
label: 'Model Name',
name: 'modelName',
@@ -149,7 +157,9 @@ class OpenAI_LLMs implements INode {
const credentialData = await getCredentialData(nodeData.credential ?? '', options)
const openAIApiKey = getCredentialParam('openAIApiKey', credentialData, nodeData)
const obj: Partial<OpenAIInput> & { openAIApiKey?: string } = {
const llmCache = nodeData.inputs?.llmCache as BaseCache
const obj: Partial<OpenAIInput> & BaseLLMParams & { openAIApiKey?: string } = {
temperature: parseFloat(temperature),
modelName,
openAIApiKey,
@@ -164,8 +174,9 @@ class OpenAI_LLMs implements INode {
if (batchSize) obj.batchSize = parseInt(batchSize, 10)
if (bestOf) obj.bestOf = parseInt(bestOf, 10)
let parsedBaseOptions: any | undefined = undefined
if (llmCache) obj.cache = llmCache
let parsedBaseOptions: any | undefined = undefined
if (baseOptions) {
try {
parsedBaseOptions = typeof baseOptions === 'object' ? baseOptions : JSON.parse(baseOptions)
+1 -1
View File
@@ -42,7 +42,7 @@
"google-auth-library": "^9.0.0",
"graphql": "^16.6.0",
"html-to-text": "^9.0.5",
"langchain": "^0.0.152",
"langchain": "^0.0.154",
"langfuse-langchain": "^1.0.14-alpha.0",
"langsmith": "^0.0.32",
"linkifyjs": "^4.1.1",
+7
View File
@@ -1,6 +1,7 @@
/**
* Types
*/
import { BaseCache } from 'langchain/schema'
export type NodeParamsType =
| 'asyncOptions'
@@ -176,3 +177,9 @@ export class VectorStoreRetriever {
this.vectorStore = fields.vectorStore
}
}
export interface LLMCacheBase {
name: string
description: string
baseCache: BaseCache
}
+21 -1
View File
@@ -151,6 +151,7 @@ export class CustomChainHandler extends BaseCallbackHandler {
socketIOClientId = ''
skipK = 0 // Skip streaming for first K numbers of handleLLMStart
returnSourceDocuments = false
cachedResponse = true
constructor(socketIO: Server, socketIOClientId: string, skipK?: number, returnSourceDocuments?: boolean) {
super()
@@ -161,6 +162,7 @@ export class CustomChainHandler extends BaseCallbackHandler {
}
handleLLMStart() {
this.cachedResponse = false
if (this.skipK > 0) this.skipK -= 1
}
@@ -175,13 +177,31 @@ export class CustomChainHandler extends BaseCallbackHandler {
}
handleLLMEnd() {
this.socketIO.to(this.socketIOClientId).emit('end')
/* send the end event from handleChainEnd */
// this.socketIO.to(this.socketIOClientId).emit('end')
}
handleChainEnd(outputs: ChainValues): void | Promise<void> {
if (this.returnSourceDocuments) {
this.socketIO.to(this.socketIOClientId).emit('sourceDocuments', outputs?.sourceDocuments)
}
/*
Langchain does not call handleLLMStart, handleLLMEnd, handleLLMNewToken when the chain is cached.
Callback Order is "Chain Start -> LLM Start --> LLM Token --> LLM End -> Chain End" for normal responses.
Callback Order is "Chain Start -> Chain End" for cached responses.
*/
if (this.cachedResponse) {
const cachedValue = outputs.text as string
//split at whitespace, and keep the whitespace. This is to preserve the original formatting.
const result = cachedValue.split(/(\s+)/)
result.forEach((token: string, index: number) => {
if (index === 0) {
this.socketIO.to(this.socketIOClientId).emit('start', token)
}
this.socketIO.to(this.socketIOClientId).emit('token', token)
})
}
this.socketIO.to(this.socketIOClientId).emit('end')
}
}