mincomk/risuai
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Refactor HypaProcesser instantiation to remove hardcoded model name and add models and others

Browse Source
This commit is contained in:
kwaroran
2024-12-07 00:58:09 +09:00
parent 998e8e6c65
commit e0f6c58540
15 changed files with 443 additions and 92 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
src/ts/process/embedding/addinfo.ts
View File

@@ -3,7 +3,7 @@ import { HypaProcesser } from '../memory/hypamemory'
import { getUserName } from "src/ts/util";
export async function additionalInformations(char: character,chats:Chat,){
const processer = new HypaProcesser('MiniLM')
const processer = new HypaProcesser()
const db = getDatabase()
const info = char.additionalText

6
src/ts/process/files/multisend.ts
View File

@@ -124,7 +124,7 @@ async function sendPDFFile(arg:sendFileArg) {
}
}
console.log(texts)
const hypa = new HypaProcesser('MiniLM')
const hypa = new HypaProcesser()
hypa.addText(texts)
const result = await hypa.similaritySearch(arg.query)
let message = ''
@@ -142,7 +142,7 @@ async function sendTxtFile(arg:sendFileArg) {
const lines = arg.file.split('\n').filter((a) => {
return a !== ''
})
const hypa = new HypaProcesser('MiniLM')
const hypa = new HypaProcesser()
hypa.addText(lines)
const result = await hypa.similaritySearch(arg.query)
let message = ''
@@ -157,7 +157,7 @@ async function sendTxtFile(arg:sendFileArg) {
}
async function sendXMLFile(arg:sendFileArg) {
const hypa = new HypaProcesser('MiniLM')
const hypa = new HypaProcesser()
let nodeTexts:string[] = []
const parser = new DOMParser();
const xmlDoc = parser.parseFromString(arg.file, "text/xml");

2
src/ts/process/index.svelte.ts
View File

@@ -1394,7 +1394,7 @@ export async function sendChat(chatProcessIndex = -1,arg:{
}
if(DBState.db.emotionProcesser === 'embedding'){
const hypaProcesser = new HypaProcesser('MiniLM')
const hypaProcesser = new HypaProcesser()
await hypaProcesser.addText(emotionList.map((v) => 'emotion:' + v))
let searched = (await hypaProcesser.similaritySearchScored(result)).map((v) => {
v[0] = v[0].replace("emotion:",'')

2
src/ts/process/lua.ts
View File

@@ -213,7 +213,7 @@ export async function runLua(code:string, arg:{
if(!LuaLowLevelIds.has(id)){
return
}
const processer = new HypaProcesser('MiniLM')
const processer = new HypaProcesser()
await processer.addText(value)
return await processer.similaritySearch(source)
})

2
src/ts/process/memory/hanuraiMemory.ts
View File

@@ -13,7 +13,7 @@ export async function hanuraiMemory(chats:OpenAIChat[],arg:{
}){
const db = getDatabase()
const tokenizer = arg.tokenizer
const processer = new HypaProcesser('MiniLM')
const processer = new HypaProcesser()
let addTexts:string[] = []
const queryStartIndex=chats.length-maxRecentChatQuery
console.log(chats.length,maxRecentChatQuery,queryStartIndex)

51
src/ts/process/memory/hypamemory.ts
View File

@@ -3,21 +3,47 @@ import { globalFetch } from "src/ts/globalApi.svelte";
import { runEmbedding } from "../transformers";
import { alertError } from "src/ts/alert";
import { appendLastPath } from "src/ts/util";
import { getDatabase } from "src/ts/storage/database.svelte";
export type HypaModel = 'ada'|'MiniLM'|'nomic'|'custom'|'nomicGPU'|'bgeSmallEn'|'bgeSmallEnGPU'|'bgem3'|'bgem3GPU'|'openai3small'|'openai3large'
const localModels = {
models: {
'MiniLM':'Xenova/all-MiniLM-L6-v2',
'nomic':'nomic-ai/nomic-embed-text-v1.5',
'nomicGPU':'nomic-ai/nomic-embed-text-v1.5',
'bgeSmallEn': 'BAAI/bge-small-en-v1.5',
'bgeSmallEnGPU': 'BAAI/bge-small-en-v1.5',
'bgem3': 'BAAI/bge-m3',
'bgem3GPU': 'BAAI/bge-m3',
},
gpuModels:[
'nomicGPU',
'bgeSmallEnGPU',
'bgem3GPU'
]
}
export class HypaProcesser{
oaikey:string
vectors:memoryVector[]
forage:LocalForage
model:'ada'|'MiniLM'|'nomic'|'custom'
model:HypaModel
customEmbeddingUrl:string
constructor(model:'ada'|'MiniLM'|'nomic'|'custom',customEmbeddingUrl?:string){
constructor(model:HypaModel|'auto' = 'auto',customEmbeddingUrl?:string){
this.forage = localforage.createInstance({
name: "hypaVector"
})
this.vectors = []
this.model = model
if(model === 'auto'){
const db = getDatabase()
this.model = db.hypaModel || 'MiniLM'
}
else{
this.model = model
}
this.customEmbeddingUrl = customEmbeddingUrl
}
@@ -39,9 +65,9 @@ export class HypaProcesser{
async getEmbeds(input:string[]|string):Promise<VectorArray[]> {
if(this.model === 'MiniLM' || this.model === 'nomic'){
if(Object.keys(localModels.models).includes(this.model)){
const inputs:string[] = Array.isArray(input) ? input : [input]
let results:Float32Array[] = await runEmbedding(inputs, this.model === 'nomic' ? 'nomic-ai/nomic-embed-text-v1.5' : 'Xenova/all-MiniLM-L6-v2')
let results:Float32Array[] = await runEmbedding(inputs, localModels.models[this.model], localModels.gpuModels.includes(this.model) ? 'webgpu' : 'wasm')
return results
}
let gf = null;
@@ -58,14 +84,21 @@ export class HypaProcesser{
},
})
}
if(this.model === 'ada'){
if(this.model === 'ada' || this.model === 'openai3small' || this.model === 'openai3large'){
const db = getDatabase()
const models = {
'ada':'text-embedding-ada-002',
'openai3small':'text-embedding-3-small',
'openai3large':'text-embedding-3-large'
}
gf = await globalFetch("https://api.openai.com/v1/embeddings", {
headers: {
"Authorization": "Bearer " + this.oaikey
"Authorization": "Bearer " + db.supaMemoryKey || this.oaikey
},
body: {
"input": input,
"model": "text-embedding-ada-002"
"input": input,
"model": models[this.model]
}
})
}

2
src/ts/process/scripts.ts
View File

@@ -321,7 +321,7 @@ export async function processScriptFull(char:character|groupChat|simpleCharacter
}
}
const processer = new HypaProcesser('MiniLM')
const processer = new HypaProcesser()
await processer.addText(assetNames)
const matches = data.matchAll(assetRegex)

26
src/ts/process/transformers.ts
View File

@@ -1,4 +1,4 @@
import {env, AutoTokenizer, pipeline, type SummarizationOutput, type TextGenerationConfig, type TextGenerationOutput, FeatureExtractionPipeline, TextToAudioPipeline, type ImageToTextOutput } from '@xenova/transformers';
import {env, AutoTokenizer, pipeline, type SummarizationOutput, type TextGenerationConfig, type TextGenerationOutput, FeatureExtractionPipeline, TextToAudioPipeline, type ImageToTextOutput } from '@huggingface/transformers';
import { unzip } from 'fflate';
import { globalFetch, loadAsset, saveAsset } from 'src/ts/globalApi.svelte';
import { selectSingleFile } from 'src/ts/util';
@@ -15,6 +15,7 @@ async function initTransformers(){
env.useBrowserCache = false
env.useFSCache = false
env.useCustomCache = true
env.allowLocalModels = true
env.customCache = {
put: async (url:URL|string, response:Response) => {
await tfCache.put(url, response)
@@ -33,10 +34,12 @@ async function initTransformers(){
console.log('transformers loaded')
}
export const runTransformers = async (baseText:string, model:string,config:TextGenerationConfig = {}) => {
export const runTransformers = async (baseText:string, model:string,config:TextGenerationConfig, device:'webgpu'|'wasm' = 'wasm') => {
await initTransformers()
let text = baseText
let generator = await pipeline('text-generation', model);
let generator = await pipeline('text-generation', model, {
device
});
let output = await generator(text, config) as TextGenerationOutput
const outputOne = output[0]
return outputOne
@@ -50,16 +53,25 @@ export const runSummarizer = async (text: string) => {
}
let extractor:FeatureExtractionPipeline = null
let lastEmbeddingModelQuery:string = ''
type EmbeddingModel = 'Xenova/all-MiniLM-L6-v2'|'nomic-ai/nomic-embed-text-v1.5'
export const runEmbedding = async (texts: string[], model:EmbeddingModel = 'Xenova/all-MiniLM-L6-v2'):Promise<Float32Array[]> => {
export const runEmbedding = async (texts: string[], model:EmbeddingModel = 'Xenova/all-MiniLM-L6-v2', device:'webgpu'|'wasm'):Promise<Float32Array[]> => {
await initTransformers()
if(!extractor){
extractor = await pipeline('feature-extraction', model);
console.log('running embedding')
let embeddingModelQuery = model + device
if(!extractor || embeddingModelQuery !== lastEmbeddingModelQuery){
extractor = await pipeline('feature-extraction', model, {
device: device,
progress_callback: (progress) => {
console.log(progress)
}
});
console.log('extractor loaded')
}
let result = await extractor(texts, { pooling: 'mean', normalize: true });
console.log(texts, result)
const data = result.data as Float32Array
console.log(data)
const lenPerText = data.length / texts.length
let res:Float32Array[] = []
for(let i = 0; i < texts.length; i++){

2
src/ts/process/triggers.ts
View File

@@ -459,7 +459,7 @@ export async function runTrigger(char:character,mode:triggerMode, arg:{
break
}
const processer = new HypaProcesser('MiniLM')
const processer = new HypaProcesser()
const effectValue = risuChatParser(effect.value,{chara:char})
const source = risuChatParser(effect.source,{chara:char})
await processer.addText(effectValue.split('§'))