fix: improve HypaV3 Modal (#851)

# PR Checklist - [ ] Have you checked if it works normally in all models? *Ignore this if it doesn't use models.* - [ ] Have you checked if it works normally in all web, local, and node hosted versions? If it doesn't, have you blocked it in those versions? - [ ] Have you added type definitions? # Description This PR introduces following: - fix: pass message index when processing regex script in HypaV3 Modal - feat: add BGE-m3-ko embedding
2025-05-18 17:15:48 +09:00
parent 0a792db7c2 ba150a0913
commit a33f588fbd
6 changed files with 86 additions and 70 deletions
--- a/src/lib/Others/HypaV3Modal.svelte
+++ b/src/lib/Others/HypaV3Modal.svelte
@@ -413,11 +413,17 @@
  }

  function isOrphan(summaryIndex: number): boolean {
+    const char = DBState.db.characters[$selectedCharID];
+    const chat = char.chats[DBState.db.characters[$selectedCharID].chatPage];
    const summary = hypaV3DataState.summaries[summaryIndex];

    for (const chatMemo of summary.chatMemos) {
-      if (!getMessageFromChatMemo(chatMemo)) {
-        return true;
+      if (chatMemo == null) {
+        // Check first message exists
+        if (!getFirstMessage()) return true;
+      } else {
+        if (chat.message.findIndex((m) => m.chatId === chatMemo) === -1)
+          return true;
      }
    }

@@ -437,8 +443,7 @@
      const summary = hypaV3DataState.summaries[summaryIndex];
      const toSummarize: OpenAIChat[] = await Promise.all(
        summary.chatMemos.map(async (chatMemo) => {
-          // Processed message
-          const message = await getProcessedMessageFromChatMemo(chatMemo);
+          const message = await getMessageFromChatMemo(chatMemo);

          return {
            role: (message.role === "char"
@@ -495,31 +500,28 @@
    summaryUIState.isRerolledTranslating = false;
  }

-  async function getProcessedMessageFromChatMemo(
+  async function getMessageFromChatMemo(
    chatMemo: string | null
  ): Promise<Message | null> {
-    const unprocessed = getMessageFromChatMemo(chatMemo);
-
-    if (!unprocessed) {
-      return null;
-    }
-
-    return getCurrentHypaV3Preset().settings.processRegexScript
-      ? await processRegexScript(unprocessed)
-      : unprocessed;
-  }
-
-  function getMessageFromChatMemo(chatMemo: string | null): Message | null {
    const char = DBState.db.characters[$selectedCharID];
    const chat = char.chats[DBState.db.characters[$selectedCharID].chatPage];
+    const shouldProcess = getCurrentHypaV3Preset().settings.processRegexScript;
+
+    let msg = null;
+    let msgIndex = -1;

    if (chatMemo == null) {
      const firstMessage = getFirstMessage();

-      return firstMessage ? { role: "char", data: firstMessage } : null;
+      if (!firstMessage) return null;
+      msg = { role: "char", data: firstMessage };
+    } else {
+      msgIndex = chat.message.findIndex((m) => m.chatId === chatMemo);
+      if (msgIndex === -1) return null;
+      msg = chat.message[msgIndex];
    }

-    return chat.message.find((m) => m.chatId === chatMemo) || null;
+    return shouldProcess ? await processRegexScript(msg, msgIndex) : msg;
  }

  function getFirstMessage(): string | null {
@@ -533,15 +535,17 @@
        : null;
  }

-  async function processRegexScript(msg: Message): Promise<Message> {
+  async function processRegexScript(
+    msg: Message,
+    msgIndex: number = -1
+  ): Promise<Message> {
    const char = DBState.db.characters[$selectedCharID];
-    const chat = char.chats[DBState.db.characters[$selectedCharID].chatPage];
    const newData: string = (
      await processScriptFull(
        char,
        risuChatParser(msg.data, { chara: char, role: msg.role }),
        "editprocess",
-        -1,
+        msgIndex,
        {
          chatRole: msg.role,
        }
@@ -593,8 +597,7 @@
      return;
    }

-    // Processed message
-    const message = await getProcessedMessageFromChatMemo(
+    const message = await getMessageFromChatMemo(
      expandedMessageUIState.selectedChatMemo
    );

@@ -632,52 +635,41 @@
    }
  }

-  async function getProcessedNextSummarizationTarget(): Promise<Message | null> {
-    const unprocessed = getNextSummarizationTarget();
-
-    if (!unprocessed) {
-      return null;
-    }
-
-    return getCurrentHypaV3Preset().settings.processRegexScript
-      ? await processRegexScript(unprocessed)
-      : unprocessed;
-  }
-
-  function getNextSummarizationTarget(): Message | null {
+  async function getNextSummarizationTarget(): Promise<Message | null> {
    const char = DBState.db.characters[$selectedCharID];
    const chat = char.chats[DBState.db.characters[$selectedCharID].chatPage];
+    const shouldProcess = getCurrentHypaV3Preset().settings.processRegexScript;

    // Summaries exist
    if (hypaV3DataState.summaries.length > 0) {
      const lastSummary = hypaV3DataState.summaries.at(-1);
      const lastMessageIndex = chat.message.findIndex(
-        (msg) => msg.chatId === lastSummary.chatMemos.at(-1)
+        (m) => m.chatId === lastSummary.chatMemos.at(-1)
      );

      if (lastMessageIndex !== -1) {
-        const nextMessage = chat.message[lastMessageIndex + 1];
+        const next = chat.message[lastMessageIndex + 1] ?? null;

-        if (nextMessage) {
-          return nextMessage;
-        }
+        return next && shouldProcess
+          ? await processRegexScript(next, lastMessageIndex + 1)
+          : next;
      }
    }

+    // When no summaries exist OR couldn't find last connected message,
+    // check if first message is available
    const firstMessage = getFirstMessage();

-    // When no summaries exist OR couldn't find last connected message
-    // Check if first message is available
-    if (!firstMessage || firstMessage.trim() === "") {
-      if (chat.message.length > 0) {
-        return chat.message[0];
-      }
+    if (!firstMessage) {
+      const next = chat.message[0] ?? null;

-      return null;
+      return next && shouldProcess ? await processRegexScript(next, 0) : next;
    }

-    // will summarize first message
-    return { role: "char", chatId: "first", data: firstMessage };
+    // Will summarize first message
+    const next: Message = { role: "char", chatId: "first", data: firstMessage };
+
+    return shouldProcess ? await processRegexScript(next) : next;
  }

  function isHypaV2ConversionPossible(): boolean {
@@ -1278,8 +1270,7 @@
                {#if expandedMessageUIState?.summaryIndex === i}
                  <!-- Expanded Message -->
                  <div class="mt-2 sm:mt-4">
-                    <!-- Processed Message -->
-                    {#await getProcessedMessageFromChatMemo(expandedMessageUIState.selectedChatMemo) then expandedMessage}
+                    {#await getMessageFromChatMemo(expandedMessageUIState.selectedChatMemo) then expandedMessage}
                      {#if expandedMessage}
                        <!-- Role -->
                        <div class="mb-2 sm:mb-4 text-sm text-zinc-400">
@@ -1336,7 +1327,7 @@

        <!-- Next Summarization Target -->
        <div class="mt-2 sm:mt-4">
-          {#await getProcessedNextSummarizationTarget() then nextMessage}
+          {#await getNextSummarizationTarget() then nextMessage}
            {#if nextMessage}
              {@const chatId =
                nextMessage.chatId === "first"
--- a/src/lib/Playground/PlaygroundEmbedding.svelte
+++ b/src/lib/Playground/PlaygroundEmbedding.svelte
@@ -5,6 +5,7 @@
    import SelectInput from "../UI/GUI/SelectInput.svelte";
    import Button from "../UI/GUI/Button.svelte";
    import { HypaProcesser } from "src/ts/process/memory/hypamemory";
+    import { DBState } from "src/ts/stores.svelte"

    let query = $state("");
    let model = $state("MiniLM");
@@ -27,24 +28,43 @@
 <h2 class="text-4xl text-textcolor my-6 font-black relative">{language.embedding}</h2>

 <span class="text-textcolor text-lg">Model</span>
-<SelectInput bind:value={model}>
+<SelectInput bind:value={model} className="mb-4">
+    {#if 'gpu' in navigator}
+        <OptionInput value="MiniLMGPU">MiniLM L6 v2 (GPU)</OptionInput>
+        <OptionInput value="nomicGPU">Nomic Embed Text v1.5 (GPU)</OptionInput>
+        <OptionInput value="bgeSmallEnGPU">BGE Small English (GPU)</OptionInput>
+        <OptionInput value="bgem3GPU">BGE Medium 3 (GPU)</OptionInput>
+        <OptionInput value="multiMiniLMGPU">Multilingual MiniLM L12 v2 (GPU)</OptionInput>
+        <OptionInput value="bgeM3KoGPU">BGE Medium 3 Korean (GPU)</OptionInput>
+    {/if}
    <OptionInput value="MiniLM">MiniLM L6 v2 (CPU)</OptionInput>
    <OptionInput value="nomic">Nomic Embed Text v1.5 (CPU)</OptionInput>
-    <OptionInput value="nomicGPU">Nomic Embed Text v1.5 (GPU)</OptionInput>
    <OptionInput value="bgeSmallEn">BGE Small English (CPU)</OptionInput>
-    <OptionInput value="bgeSmallEnGPU">BGE Small English (GPU)</OptionInput>
    <OptionInput value="bgem3">BGE Medium 3 (CPU)</OptionInput>
-    <OptionInput value="bgem3GPU">BGE Medium 3 (GPU)</OptionInput>
+    <OptionInput value="multiMiniLM">Multilingual MiniLM L12 v2 (CPU)</OptionInput>
+    <OptionInput value="bgeM3Ko">BGE Medium 3 Korean (CPU)</OptionInput>
    <OptionInput value="openai3small">OpenAI text-embedding-3-small</OptionInput>
    <OptionInput value="openai3large">OpenAI text-embedding-3-large</OptionInput>
+    <OptionInput value="ada">OpenAI Ada</OptionInput>
    <OptionInput value="custom">Custom (OpenAI-compatible)</OptionInput>
 </SelectInput>

-{#if model === "custom"}
-    <span class="text-textcolor text-lg">Custom Server URL</span>
-    <TextInput bind:value={customEmbeddingUrl} size="lg" fullwidth />
+{#if model === 'openai3small' || model === 'openai3large' || model === 'ada'}
+    <span class="text-textcolor text-lg">OpenAI API Key</span>
+    <TextInput size="sm" marginBottom bind:value={DBState.db.supaMemoryKey}/>
 {/if}

+{#if model === "custom"}
+    <span class="text-textcolor text-lg">URL</span>
+    <TextInput size="sm" marginBottom bind:value={DBState.db.hypaCustomSettings.url}/>
+    <span class="text-textcolor text-lg">Key/Password</span>
+    <TextInput size="sm" marginBottom bind:value={DBState.db.hypaCustomSettings.key}/>
+    <span class="text-textcolor text-lg">Request Model</span>
+    <TextInput size="sm" marginBottom bind:value={DBState.db.hypaCustomSettings.model}/>
+{/if}
+
+<div class="mb-4"></div>
+
 <span class="text-textcolor text-lg">Query</span>
 <TextInput bind:value={query} size="lg" fullwidth />

--- a/src/lib/Setting/Pages/OtherBotSettings.svelte
+++ b/src/lib/Setting/Pages/OtherBotSettings.svelte
@@ -917,19 +917,21 @@
        {/if}

        <span class="text-textcolor">{language.embedding}</span>
-        <SelectInput className="mt-2 mb-2" bind:value={DBState.db.hypaModel}>
+        <SelectInput className="mb-4" bind:value={DBState.db.hypaModel}>
            {#if 'gpu' in navigator}
                <OptionInput value="MiniLMGPU">MiniLM L6 v2 (GPU)</OptionInput>
                <OptionInput value="nomicGPU">Nomic Embed Text v1.5 (GPU)</OptionInput>
                <OptionInput value="bgeSmallEnGPU">BGE Small English (GPU)</OptionInput>
                <OptionInput value="bgem3GPU">BGE Medium 3 (GPU)</OptionInput>
                <OptionInput value="multiMiniLMGPU">Multilingual MiniLM L12 v2 (GPU)</OptionInput>
+                <OptionInput value="bgeM3KoGPU">BGE Medium 3 Korean (GPU)</OptionInput>
            {/if}
            <OptionInput value="MiniLM">MiniLM L6 v2 (CPU)</OptionInput>
            <OptionInput value="nomic">Nomic Embed Text v1.5 (CPU)</OptionInput>
            <OptionInput value="bgeSmallEn">BGE Small English (CPU)</OptionInput>
            <OptionInput value="bgem3">BGE Medium 3 (CPU)</OptionInput>
            <OptionInput value="multiMiniLM">Multilingual MiniLM L12 v2 (CPU)</OptionInput>
+            <OptionInput value="bgeM3Ko">BGE Medium 3 Korean (CPU)</OptionInput>
            <OptionInput value="openai3small">OpenAI text-embedding-3-small</OptionInput>
            <OptionInput value="openai3large">OpenAI text-embedding-3-large</OptionInput>
            <OptionInput value="ada">OpenAI Ada</OptionInput>
@@ -942,12 +944,12 @@
        {/if}

        {#if DBState.db.hypaModel === 'custom'}
-        <span class="text-textcolor">URL</span>
-        <TextInput size="sm" marginBottom bind:value={DBState.db.hypaCustomSettings.url}/>
-        <span class="text-textcolor">Key/Password</span>
-        <TextInput size="sm" marginBottom bind:value={DBState.db.hypaCustomSettings.key}/>
-        <span class="text-textcolor">Request Model</span>
-        <TextInput size="sm" marginBottom bind:value={DBState.db.hypaCustomSettings.model}/>
+            <span class="text-textcolor">URL</span>
+            <TextInput size="sm" marginBottom bind:value={DBState.db.hypaCustomSettings.url}/>
+            <span class="text-textcolor">Key/Password</span>
+            <TextInput size="sm" marginBottom bind:value={DBState.db.hypaCustomSettings.key}/>
+            <span class="text-textcolor">Request Model</span>
+            <TextInput size="sm" marginBottom bind:value={DBState.db.hypaCustomSettings.model}/>
        {/if}

    </Arcodion>
--- a/src/ts/process/memory/hypamemory.ts
+++ b/src/ts/process/memory/hypamemory.ts
@@ -21,6 +21,8 @@ export const localModels = {
        'bgem3GPU': 'Xenova/bge-m3',
        'multiMiniLM': 'Xenova/paraphrase-multilingual-MiniLM-L12-v2',
        'multiMiniLMGPU': 'Xenova/paraphrase-multilingual-MiniLM-L12-v2',
+        'bgeM3Ko': 'HyperBlaze/BGE-m3-ko',
+        'bgeM3KoGPU': 'HyperBlaze/BGE-m3-ko',
    },
    gpuModels:[
        'MiniLMGPU',
@@ -28,6 +30,7 @@ export const localModels = {
        'bgeSmallEnGPU',
        'bgem3GPU',
        'multiMiniLMGPU',
+        'bgeM3KoGPU',
    ]
 }

--- a/src/ts/process/memory/hypav3.ts
+++ b/src/ts/process/memory/hypav3.ts
@@ -128,7 +128,7 @@ export async function hypaMemoryV3(
  } finally {
    if (settings.summarizationModel !== "subModel") {
      try {
-        unloadEngine();
+        await unloadEngine();
      } catch {}
    }
  }
--- a/src/ts/process/transformers.ts
+++ b/src/ts/process/transformers.ts
@@ -66,7 +66,7 @@ export const runEmbedding = async (texts: string[], model:EmbeddingModel = 'Xeno
        }
        extractor = await pipeline('feature-extraction', model, {
            // Default dtype for webgpu is fp32, so we can use q8, which is the default dtype in wasm.
-            ...(device === 'webgpu' ? { dtype: "q8" } : {}),
+            dtype: "q8",
            device: device,
            progress_callback: (progress) => {
                console.log(progress)