CFG: Include the entire prompt with negative prompt

CFG with LLMs works differently than stable diffusion. The main principle
is prompt mixing and utilizing the differences between the two prompts
rather than a full "negative prompt" of what the user doesn't want.

SillyTavern its own way of formatting a prompt sent to an LLM backend.
Therefore, take that prompt and add negatives to it.

Signed-off-by: kingbri <bdashore3@proton.me>

public/script.js

@@ -164,6 +164,7 @@ import { deviceInfo } from "./scripts/RossAscends-mods.js";
 import { registerPromptManagerMigration } from "./scripts/PromptManager.js";
 import { getRegexedString, regex_placement } from "./scripts/extensions/regex/engine.js";
 import { FILTER_TYPES, FilterHelper } from "./scripts/filters.js";
+import { getCfg, getNegativePrompt } from "./scripts/extensions/cfg/util.js";
 
 //exporting functions and vars for mods
 export {
@@ -2905,6 +2906,8 @@ async function Generate(type, { automatic_trigger, force_name2, resolve, reject,
             let this_amount_gen = parseInt(amount_gen); // how many tokens the AI will be requested to generate
             let this_settings = koboldai_settings[koboldai_setting_names[preset_settings]];
 
+            const cfgValues = getCfg(finalPromt);
+
             if (isMultigenEnabled() && type !== 'quiet') {
                 // if nothing has been generated yet..
                 this_amount_gen = getMultigenAmount();
@@ -2912,6 +2915,7 @@ async function Generate(type, { automatic_trigger, force_name2, resolve, reject,
 
             let thisPromptBits = [];
 
+            // TODO: Make this a switch
             if (main_api == 'koboldhorde' && horde_settings.auto_adjust_response_length) {
                 this_amount_gen = Math.min(this_amount_gen, adjustedParams.maxLength);
                 this_amount_gen = Math.max(this_amount_gen, MIN_AMOUNT_GEN); // prevent validation errors
@@ -2934,7 +2938,7 @@ async function Generate(type, { automatic_trigger, force_name2, resolve, reject,
                 }
             }
             else if (main_api == 'textgenerationwebui') {
-                generate_data = getTextGenGenerationData(finalPromt, this_amount_gen, isImpersonate);
+                generate_data = getTextGenGenerationData(finalPromt, this_amount_gen, isImpersonate, cfgValues);
                 generate_data.use_mancer = api_use_mancer_webui;
             }
             else if (main_api == 'novel') {

public/scripts/extensions/cfg/index.js

@@ -273,6 +273,7 @@ jQuery(async () => {
         saveSettingsDebounced();
     });
 
+    // TODO: Add negative insertion depth
     windowHtml.find('#global_cfg_negative_prompt').on('input', function() {
         extension_settings.cfg.global.negative_prompt = $(this).val();
         saveSettingsDebounced();

public/scripts/extensions/cfg/util.js

@@ -12,18 +12,23 @@ export const metadataKeys = {
     guidance_scale: "cfg_guidance_scale",
     negative_prompt: "cfg_negative_prompt",
     negative_combine: "cfg_negative_combine",
-    groupchat_individual_chars: "cfg_groupchat_individual_chars"
+    groupchat_individual_chars: "cfg_groupchat_individual_chars",
+    negative_insertion_depth: "cfg_negative_insertion_depth"
 }
 
 // Gets the CFG value from hierarchy of chat -> character -> global
 // Returns undefined values which should be handled in the respective backend APIs
-export function getCfg() {
+// TODO: Include a custom negative separator
+// TODO: Maybe use existing prompt building/substitution?
+export function getCfg(prompt) {
+    const splitPrompt = prompt?.split("\n") ?? [];
     let splitNegativePrompt = [];
     const charaCfg = extension_settings.cfg.chara?.find((e) => e.name === getCharaFilename(this_chid));
     const guidanceScale = getGuidanceScale(charaCfg);
     const chatNegativeCombine = chat_metadata[metadataKeys.negative_combine] ?? [];
 
     // If there's a guidance scale, continue. Otherwise assume undefined
+    // TODO: Run substitute params
     if (guidanceScale?.value && guidanceScale?.value !== 1) {
         if (guidanceScale.type === cfgType.chat || chatNegativeCombine.includes(cfgType.chat)) {
             splitNegativePrompt.push(chat_metadata[metadataKeys.negative_prompt]?.trim());
@@ -37,12 +42,15 @@ export function getCfg() {
             splitNegativePrompt.push(extension_settings.cfg.global.negative_prompt?.trim());
         }
 
-        const combinedNegatives = splitNegativePrompt.filter((e) => e.length > 0).join(", ");
+        // TODO: use a custom separator for join
-        console.debug(`Setting CFG with guidance scale: ${guidanceScale.value}, negatives: ${combinedNegatives}`)
+        const combinedNegatives = splitNegativePrompt.filter((e) => e.length > 0).join("\n");
+        const insertionDepth = chat_metadata[metadataKeys.negative_insertion_depth] ?? 1;
+        splitPrompt.splice(splitPrompt.length - insertionDepth, 0, combinedNegatives);
+        console.log(`Setting CFG with guidance scale: ${guidanceScale.value}, negatives: ${combinedNegatives}`);
 
         return {
             guidanceScale: guidanceScale.value,
-            negativePrompt: combinedNegatives
+            negativePrompt: splitPrompt.join("\n")
         }
     }
 }
@@ -70,3 +78,12 @@ function getGuidanceScale(charaCfg) {
         value: extension_settings.cfg.global.guidance_scale
     };
 }
+
+export function getNegativePrompt(prompt) {
+    const splitPrompt = prompt.split("\n");
+    const insertionDepth = chat_metadata[metadataKeys.negative_insertion_depth] ?? 1;
+    splitPrompt.splice(splitPrompt.length - insertionDepth, 0, "Test negative list");
+    console.log(splitPrompt);
+    const negativePrompt = splitPrompt.join("\n");
+    //console.log(negativePrompt);
+}

public/scripts/textgen-settings.js

@@ -235,9 +235,7 @@ async function generateTextGenWithStreaming(generate_data, signal) {
     }
 }
 
-export function getTextGenGenerationData(finalPromt, this_amount_gen, isImpersonate) {
+export function getTextGenGenerationData(finalPromt, this_amount_gen, isImpersonate, cfgValues) {
-    const cfgValues = getCfg();
-
     return {
         'prompt': finalPromt,
         'max_new_tokens': this_amount_gen,