From bc018a97cf41bbbfc3bde714a64e5fda9825e98b Mon Sep 17 00:00:00 2001
From: kwaroran <kwaworan@gmail.com>
Date: Sun, 6 Aug 2023 12:38:18 +0900
Subject: [PATCH] [feat] improved supa/hypa memory

---
 src/ts/process/memory/hypamemory.ts | 24 +++++++++++++++++++++++-
 src/ts/process/memory/supaMemory.ts | 27 ++++++++++-----------------
 2 files changed, 33 insertions(+), 18 deletions(-)

diff --git a/src/ts/process/memory/hypamemory.ts b/src/ts/process/memory/hypamemory.ts
index 5d6b8078..7b43d39a 100644
--- a/src/ts/process/memory/hypamemory.ts
+++ b/src/ts/process/memory/hypamemory.ts
@@ -1,18 +1,21 @@
 import localforage from "localforage";
 import { similarity } from "ml-distance";
 import { globalFetch } from "src/ts/storage/globalApi";
+import { runEmbedding } from "../embedding/transformers";
 
 
 export class HypaProcesser{
     oaikey:string
     vectors:memoryVector[]
     forage:LocalForage
+    model:'ada'|'MiniLM'
 
-    constructor(){
+    constructor(model:'ada'|'MiniLM'){
         this.forage = localforage.createInstance({
             name: "hypaVector"
         })
         this.vectors = []
+        this.model = model
     }
 
     async embedDocuments(texts: string[]): Promise<number[][]> {
@@ -33,6 +36,25 @@ export class HypaProcesser{
     
     
     async getEmbeds(input:string[]|string) {
+        if(this.model === 'MiniLM'){
+            const inputs:string[] = Array.isArray(input) ? input : [input]
+            let results:Float32Array[] = []
+            for(let i=0;i<inputs.length;i++){
+                const res = await runEmbedding(inputs[i])
+                results.push(res)
+            }
+            //convert to number[][]
+            const result:number[][] = []
+            for(let i=0;i<results.length;i++){
+                const res = results[i]
+                const arr:number[] = []
+                for(let j=0;j<res.length;j++){
+                    arr.push(res[j])
+                }
+                result.push(arr)
+            }
+            return result
+        }
         const gf = await globalFetch("https://api.openai.com/v1/embeddings", {
             headers: {
             "Authorization": "Bearer " + this.oaikey
diff --git a/src/ts/process/memory/supaMemory.ts b/src/ts/process/memory/supaMemory.ts
index 951ff988..59996dbf 100644
--- a/src/ts/process/memory/supaMemory.ts
+++ b/src/ts/process/memory/supaMemory.ts
@@ -202,7 +202,7 @@ export async function supaMemory(
         let hypaResult = ""
 
         if(arg.asHyper){
-            const hypa = new HypaProcesser()
+            const hypa = new HypaProcesser('MiniLM')
             hypa.oaikey = db.supaMemoryKey
             hypa.vectors = []
             hypaChunks = hypaChunks.filter((value) => value.length > 1)
@@ -216,6 +216,7 @@ export async function supaMemory(
                 role: "assistant",
                 content: hypaResult
             })
+            currentTokens += 10
         }
 
         while(currentTokens > maxContextTokens){
@@ -264,10 +265,11 @@ export async function supaMemory(
                 const tokens = await tokenizer.tokenizeChat(cont)
                 if((chunkSize + tokens) > maxChunkSize){
                     if(stringlizedChat === ''){
-                        
-
                         if(cont.role !== 'function' && cont.role !== 'system'){
                             stringlizedChat += `${cont.role === 'assistant' ? char.type === 'group' ? '' : char.name : db.username}: ${cont.content}\n\n`
+                            spiceLen += 1
+                            currentTokens -= tokens
+                            chunkSize += tokens
                         }
                     }
                     lastId = cont.memo
@@ -288,30 +290,21 @@ export async function supaMemory(
                 }
     
                 const tokenz = await tokenize(result + '\n\n')
-                currentTokens += tokenz
-                supaMemory += result.replace(/\n+/g,'\n') + '\n\n'
+                hypaChunks.push(result.replace(/\n+/g,'\n'))
 
-                let SupaMemoryList = supaMemory.split('\n\n')
+                let SupaMemoryList = supaMemory.split('\n\n').filter((value) => value.length > 1)
                 if(SupaMemoryList.length >= (arg.asHyper ? 3 : 4)){
                     const oldSupaMemory = supaMemory
-                    let modifies:string[] = []
-                    for(let i=0;i<3;i++){
-                        modifies.push(SupaMemoryList.shift())
-                    }
-                    hypaChunks.push(...modifies)
-
                     const result = await summarize(supaMemory)
                     if(typeof(result) !== 'string'){
                         return result
                     }
-
-                    modifies.unshift(result.replace(/\n+/g,'\n'))
-                    supaMemory = modifies.join('\n\n') + '\n\n'
-
+                    supaMemory = result
                     currentTokens -= await tokenize(oldSupaMemory)
                     currentTokens += await tokenize(supaMemory)
                 }
-                console.log(supaMemory)
+                currentTokens += tokenz
+                supaMemory += result.replace(/\n+/g,'\n')
             }
         }