Refactor inlay image parsing and add image embedding functionality

src/ts/parser.ts

@@ -154,16 +154,13 @@ async function parseAdditionalAssets(data:string, char:simpleCharacterArgument|c
 }
 
 async function parseInlayImages(data:string){
-    const db = get(DataBase)
-    if(db.inlayImage){
-        const inlayMatch = data.match(/{{inlay::(.+?)}}/g)
-        if(inlayMatch){
-            for(const inlay of inlayMatch){
-                const id = inlay.substring(9, inlay.length - 2)
-                const img = await getInlayImage(id)
-                if(img){
-                    data = data.replace(inlay, `<img src="${img.data}"/>`)
-                }
+    const inlayMatch = data.match(/{{inlay::(.+?)}}/g)
+    if(inlayMatch){
+        for(const inlay of inlayMatch){
+            const id = inlay.substring(9, inlay.length - 2)
+            const img = await getInlayImage(id)
+            if(img){
+                data = data.replace(inlay, `<img src="${img.data}"/>`)
             }
         }
     }

src/ts/process/index.ts

@@ -25,6 +25,7 @@ import { sendPeerChar } from "../sync/multiuser";
 import { runInlayScreen } from "./inlayScreen";
 import { runCharacterJS } from "../plugins/embedscript";
 import { addRerolls } from "./prereroll";
+import { runImageEmbedding } from "./transformers";
 
 export interface OpenAIChat{
     role: 'system'|'user'|'assistant'|'function'
@@ -562,16 +563,14 @@ export async function sendChat(chatProcessIndex = -1,arg:{chatAdditonalTokens?:n
             msg.chatId = v4()
         }
         let inlays:string[] = []
-        if(db.inlayImage){
-            if(msg.role === 'char'){
-                formatedChat = formatedChat.replace(/{{inlay::(.+?)}}/g, '')
-            }
-            else{
-                const inlayMatch = formatedChat.match(/{{inlay::(.+?)}}/g)
-                if(inlayMatch){
-                    for(const inlay of inlayMatch){
-                        inlays.push(inlay)
-                    }
+        if(msg.role === 'char'){
+            formatedChat = formatedChat.replace(/{{inlay::(.+?)}}/g, '')
+        }
+        else{
+            const inlayMatch = formatedChat.match(/{{inlay::(.+?)}}/g)
+            if(inlayMatch){
+                for(const inlay of inlayMatch){
+                    inlays.push(inlay)
                 }
             }
         }
@@ -590,6 +589,10 @@ export async function sendChat(chatProcessIndex = -1,arg:{chatAdditonalTokens?:n
                             height: inlayData.height
                         })
                     }
+                    else{
+                        const captionResult = await runImageEmbedding(inlayData.data) 
+                        formatedChat += `[${captionResult[0].generated_text}]`
+                    }
                 }
                 formatedChat = formatedChat.replace(inlay, '')
             }

src/ts/process/request.ts

@@ -172,35 +172,30 @@ export async function requestChatDataMain(arg:requestDataArgument, model:'model'
         case 'mistral-large-latest':
         case 'reverse_proxy':{
             let formatedChat:OpenAIChatExtra[] = []
-            if(db.inlayImage){
-                for(let i=0;i<formated.length;i++){
-                    const m = formated[i]
-                    if(m.multimodals && m.multimodals.length > 0 && m.role === 'user'){
-                        let v:OpenAIChatExtra = cloneDeep(m)
-                        let contents:OpenAIContents[] = []
-                        for(let j=0;j<m.multimodals.length;j++){
-                            contents.push({
-                                "type": "image",
-                                "image_url": {
-                                    "url": m.multimodals[j].base64,
-                                    "detail": db.gptVisionQuality
-                                }
-                            })
-                        }
+            for(let i=0;i<formated.length;i++){
+                const m = formated[i]
+                if(m.multimodals && m.multimodals.length > 0 && m.role === 'user'){
+                    let v:OpenAIChatExtra = cloneDeep(m)
+                    let contents:OpenAIContents[] = []
+                    for(let j=0;j<m.multimodals.length;j++){
                         contents.push({
-                            "type": "text",
-                            "text": m.content
+                            "type": "image",
+                            "image_url": {
+                                "url": m.multimodals[j].base64,
+                                "detail": db.gptVisionQuality
+                            }
                         })
-                        v.content = contents
-                        formatedChat.push(v)
-                    }
-                    else{
-                        formatedChat.push(m)
                     }
+                    contents.push({
+                        "type": "text",
+                        "text": m.content
+                    })
+                    v.content = contents
+                    formatedChat.push(v)
+                }
+                else{
+                    formatedChat.push(m)
                 }
-            }
-            else{
-                formatedChat = formated
             }
             
             let oobaSystemPrompts:string[] = []

src/ts/process/transformers.ts

@@ -1,4 +1,4 @@
-import {env, AutoTokenizer, pipeline, type SummarizationOutput, type TextGenerationConfig, type TextGenerationOutput, FeatureExtractionPipeline, TextToAudioPipeline } from '@xenova/transformers';
+import {env, AutoTokenizer, pipeline, type SummarizationOutput, type TextGenerationConfig, type TextGenerationOutput, FeatureExtractionPipeline, TextToAudioPipeline, type ImageToTextOutput } from '@xenova/transformers';
 import { unzip } from 'fflate';
 import { globalFetch, loadAsset, saveAsset } from 'src/ts/storage/globalApi';
 import { selectSingleFile } from 'src/ts/util';
@@ -93,6 +93,13 @@ export const runEmbedding = async (text: string):Promise<Float32Array> => {
     return (result?.data as Float32Array) ?? null;
 }
 
+export const runImageEmbedding = async (dataurl:string) => {
+    await initTransformers()
+    const captioner = await pipeline('image-to-text', 'Xenova/vit-gpt2-image-captioning');
+    const output = await captioner(dataurl)
+    return output as ImageToTextOutput
+}
+
 let synthesizer:TextToAudioPipeline = null
 let lastSynth:string = null