Fix preset setting bugs (#1038)

* add nai min_p
- fix model format change
- move min p to general
- fix missing third party settings
- hide advanced prompt correctly in simple mode
* fix menu always opening in chat

common/template-parser.ts

@@ -6,6 +6,8 @@ import peggy from 'peggy'
 import { elapsedSince } from './util'
 import { v4 } from 'uuid'
 
+type Section = 'system' | 'def' | 'history' | 'post'
+
 export type TemplateOpts = {
   continue?: boolean
   parts?: Partial<PromptParts>
@@ -33,6 +35,11 @@ export type TemplateOpts = {
     output?: Record<string, { src: string; lines: string[] }>
   }
 
+  sections?: {
+    flags: { [key in Section]?: boolean }
+    sections: { [key in Section]: string[] }
+  }
+
   /**
    * Only allow repeatable placeholders. Excludes iterators, conditions, and prompt parts.
    */
@@ -168,11 +175,19 @@ export async function parseTemplate(
   length?: number
   linesAddedCount: number
   history?: string[]
+  sections: NonNullable<TemplateOpts['sections']>
 }> {
   if (opts.limit) {
     opts.limit.output = {}
   }
 
+  const sections: TemplateOpts['sections'] = {
+    flags: {},
+    sections: { system: [], def: [], history: [], post: [] },
+  }
+
+  opts.sections = sections
+
   const parts = opts.parts || {}
 
   if (parts.systemPrompt) {
@@ -200,6 +215,7 @@ export async function parseTemplate(
   // }
 
   /** Replace iterators */
+  let history: string[] = []
   if (opts.limit && opts.limit.output) {
     for (const [id, { lines, src }] of Object.entries(opts.limit.output)) {
       src
@@ -215,6 +231,7 @@ export async function parseTemplate(
       const trimmed = filled.adding.slice().reverse()
       output = output.replace(id, trimmed.join('\n'))
       linesAddedCount += filled.linesAddedCount
+      history = trimmed
     }
 
     // Adding the low priority blocks if we still have the budget for them,
@@ -236,11 +253,15 @@ export async function parseTemplate(
     .replace(/\r\n/g, '\n')
     .replace(/\n\n+/g, '\n\n')
     .trim()
+
+  sections.sections.history = history
+
   return {
     parsed: result,
     inserts: opts.inserts ?? new Map(),
     length: await opts.limit?.encoder?.(result),
     linesAddedCount,
+    sections,
   }
 }
 
@@ -324,22 +345,57 @@ function renderNodes(nodes: PNode[], opts: TemplateOpts) {
 
 function renderNode(node: PNode, opts: TemplateOpts, conditionText?: string) {
   if (typeof node === 'string') {
+    fillSection(opts.sections, node)
     return node
   }
 
   switch (node.kind) {
     case 'placeholder': {
-      return getPlaceholder(node, opts, conditionText)
+      const result = getPlaceholder(node, opts, conditionText)
+      // We need to end the definition section prior to filling the history section
+      // History is the only deterministic marker for the end of the definitions
+      if (node.value === 'history') {
+        flagSection(opts.sections, 'def')
+      }
+      fillSection(opts.sections, result)
+      if (node.value === 'history') {
+        flagSection(opts.sections, 'history')
+      }
+      if (node.value === 'system_prompt') {
+        flagSection(opts.sections, 'system')
+      }
+      return result
+    }
+
+    case 'each': {
+      const result = renderIterator(node.value, node.children, opts)
+      // See 'placeholder' comment above for why we do this
+      if (node.value === 'history') {
+        flagSection(opts.sections, 'def')
+      }
+      fillSection(opts.sections, result)
+      if (node.value === 'history') {
+        flagSection(opts.sections, 'history')
+      }
+
+      return result
     }
 
-    case 'each':
-      return renderIterator(node.value, node.children, opts)
+    case 'if': {
+      const result = renderCondition(node, node.children, opts)
+      fillSection(opts.sections, result)
+
+      if (node.value === 'system_prompt') {
+        flagSection(opts.sections, 'system')
+      }
 
-    case 'if':
-      return renderCondition(node, node.children, opts)
+      return result
+    }
 
-    case 'lowpriority':
-      return renderLowPriority(node, opts)
+    case 'lowpriority': {
+      const result = renderLowPriority(node, opts)
+      return result
+    }
   }
 }
 
@@ -715,3 +771,30 @@ function handleDice(node: DiceExpr) {
   const rand = usable.reduce((p, c) => p + c, 0) + adjust
   return rand
 }
+
+function fillSection(state: TemplateOpts['sections'], result: string | undefined) {
+  if (!state || !result) return
+  if (result === HISTORY_MARKER) return
+
+  if (!state.flags.system) {
+    state.sections.system.push(result)
+    return
+  }
+
+  if (!state.flags.def) {
+    state.sections.def.push(result)
+    return
+  }
+
+  if (!state.flags.history) {
+    state.sections.history.push(result)
+  }
+
+  state.sections.post.push(result)
+}
+
+function flagSection(state: TemplateOpts['sections'], section: Section) {
+  if (!state) return
+
+  state.flags[section] = true
+}

srv/adapter/gemini.ts

@@ -0,0 +1,5 @@
+import { ModelAdapter } from './type'
+
+export const handleGemini: ModelAdapter = async function* (opts) {}
+
+// function toPayload(prompt: string) {}

srv/adapter/novel.ts

@@ -183,6 +183,7 @@ function getModernParams(gen: Partial<AppSchema.GenSettings>) {
     top_k: gen.topK,
     top_p: gen.topP,
     top_a: gen.topA,
+    min_p: gen.minP ?? 0,
     typical_p: gen.typicalP,
     tail_free_sampling: gen.tailFreeSampling,
     repetition_penalty: gen.repetitionPenalty,

srv/db/presets.ts

@@ -162,7 +162,7 @@ function mergeModelFormats(gen: AppSchema.UserGenPreset | null) {
   if (!gen) return gen
 
   if (gen.useAdvancedPrompt) {
-    gen.modelFormat = 'ChatML'
+    gen.modelFormat = gen.modelFormat || 'ChatML'
   } else if (gen.promptOrderFormat && gen.promptOrderFormat in BUILTIN_FORMATS) {
     gen.modelFormat = gen.promptOrderFormat as any
   } else {

tests/__snapshots__/placeholder.spec.js.snap

@@ -8,5 +8,23 @@ Object {
   "parsed": "OtherBot replies a lot
 
 MainChar's personality: MainCharacter talks a lot",
+  "sections": Object {
+    "flags": Object {},
+    "sections": Object {
+      "def": Array [],
+      "history": Array [],
+      "post": Array [],
+      "system": Array [
+        "OtherBot replies a lot",
+        "
+
+",
+        "MainChar's personality: MainCharacter talks a lot",
+        "OtherBot replies a lot
+
+MainChar's personality: MainCharacter talks a lot",
+      ],
+    },
+  },
 }
 `;

tests/prompt.spec.ts

@@ -170,6 +170,7 @@ This is how {{char}} should talk: {{example_dialogue}}`,
         ultimeJailbreak: '!!UJB_PROMPT!!',
       },
     })
+    delete (actual.template as any).sections
     expect(actual).to.matchSnapshot()
   })
 

web/pages/Chat/ChatDetail.tsx

@@ -37,7 +37,7 @@ import { TitleCard } from '/web/shared/Card'
 import { ChatGraphModal } from './components/GraphModal'
 import { EVENTS, events } from '/web/emitter'
 import { AppSchema } from '/common/types'
-import { startTour } from '/web/tours'
+import { canStartTour, startTour } from '/web/tours'
 
 export { ChatDetail as default }
 
@@ -258,7 +258,7 @@ const ChatDetail: Component = () => {
       return nav(`/chat/${chats.lastId}`)
     }
 
-    if (charName) {
+    if (charName && canStartTour('chat')) {
       settingStore.menu(true)
       setTimeout(() => {
         startTour('chat')

web/pages/Settings/AISettings.tsx

@@ -248,11 +248,14 @@ const AISettings: Component<{
 export default AISettings
 
 const ApiAccessHelp = neat`
-  The subscriber API endpoint uses the *OpenAI Text Completion* format.
+  The subscriber API endpoint uses the *OpenAI Text Completion* or *OpenAI Chat Completion* format.
 
+  The full URL for these endpoints are:
+  - https://api.agnai.chat/v1/completions
+  - https://api.agnai.chat/v1/chat/completions
+
+  **Instructions**:
   1. Select your \`API Access Preset\`: This preset will be used for your API calls. Change this on your settings page.
   2. Generate your API Key.
   3. Use the API URL \`https://api.agnai.chat\` and your generated API key.
-
-  *SillyTavern*: Use _Ooba_ as the backend
 `

web/shared/PresetSettings/General.tsx

@@ -20,7 +20,7 @@ import { PhraseBias, StoppingStrings } from '../PhraseBias'
 import { BUILTIN_FORMATS } from '/common/presets/templates'
 import { getSubscriptionModelLimits } from '/common/util'
 import { forms } from '/web/emitter'
-import { Field, ContextSize, ModelFormat, ResponseLength, Temperature } from './Fields'
+import { Field, ContextSize, ModelFormat, ResponseLength, Temperature, ThirdParty } from './Fields'
 
 export const MODEL_FORMATS = Object.keys(BUILTIN_FORMATS).map((label) => ({ label, value: label }))
 
@@ -149,6 +149,8 @@ export const GeneralSettings: Field = (props) => {
       </Show>
 
       <Card hide={!serviceHasSetting(props.service, props.format, 'thirdPartyUrl')}>
+        <ThirdParty {...props} />
+
         <div class="flex flex-wrap items-start gap-2">
           <Toggle
             fieldName="thirdPartyUrlNoSuffix"
@@ -308,6 +310,19 @@ export const GeneralSettings: Field = (props) => {
 
         <Temperature {...props} />
 
+        <RangeInput
+          fieldName="minP"
+          label="Min P"
+          helperText="Used to discard tokens with the probability under a threshold (min_p) in the sampling process. Higher values will make text more predictable. (Put this value on 0 to disable its effect)"
+          min={0}
+          max={1}
+          step={0.01}
+          value={props.inherit?.minP ?? 0}
+          disabled={props.disabled}
+          aiSetting={'minP'}
+          recommended={props.sub?.preset.minP}
+        />
+
         <Toggle
           fieldName="streamResponse"
           label="Stream Response"

web/shared/PresetSettings/Prompt.tsx

@@ -120,7 +120,7 @@ export const PromptSettings: Field = (props) => {
             disabled={props.disabled}
             showHelp
             inherit={props.inherit}
-            hide={useAdvanced() === 'basic'}
+            hide={useAdvanced() === 'basic' || props.mode === 'simple'}
             showTemplates
           />
 

web/shared/PresetSettings/Sliders.tsx

@@ -89,19 +89,6 @@ export const SliderSettings: Component<
           aiSetting={'cfgScale'}
         />
 
-        <RangeInput
-          fieldName="minP"
-          label="Min P"
-          helperText="Used to discard tokens with the probability under a threshold (min_p) in the sampling process. Higher values will make text more predictable. (Put this value on 0 to disable its effect)"
-          min={0}
-          max={1}
-          step={0.01}
-          value={props.inherit?.minP ?? 0}
-          disabled={props.disabled}
-          aiSetting={'minP'}
-          recommended={props.sub?.preset.minP}
-        />
-
         <RangeInput
           fieldName="topP"
           label="Top P"

web/shared/PresetSettings/settings.ts

@@ -67,6 +67,7 @@ export const ADAPTER_SETTINGS: {
   tailFreeSampling: ['horde', 'novel', 'kobold', 'ooba', 'agnaistic', 'aphrodite', 'tabby'],
   minP: [
     'llamacpp',
+    'novel',
     'kobold',
     'koboldcpp',
     'exllamav2',