Spaces:
Sleeping
Sleeping
File size: 12,329 Bytes
c78c312 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 | /**
* AutoTune Feedback Loop Engine
*
* Collects quality signals (user ratings + automated heuristics) after each response,
* stores them alongside the parameters that produced them, and uses Exponential Moving
* Average to learn optimal parameter adjustments per context type over time.
*
* The learned adjustments blend into AutoTune's parameter selection β more feedback data
* means more influence, capped at 50% weight so base profiles remain the foundation.
*/
import type { AutoTuneParams, ContextType } from './autotune'
// ββ Types ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
export interface ResponseHeuristics {
responseLength: number
repetitionScore: number // 0.0 = no repetition, 1.0 = very repetitive
averageSentenceLength: number
vocabularyDiversity: number // unique words / total words ratio
}
export interface FeedbackRecord {
messageId: string
timestamp: number
contextType: ContextType
model: string
persona: string
params: AutoTuneParams
rating: 1 | -1 // thumbs up / thumbs down
heuristics: ResponseHeuristics
}
export interface LearnedProfile {
contextType: ContextType
sampleCount: number
positiveCount: number
negativeCount: number
positiveParams: AutoTuneParams // EMA of params from upvoted responses
negativeParams: AutoTuneParams // EMA of params from downvoted responses
adjustments: Partial<AutoTuneParams> // computed delta to apply
lastUpdated: number
}
export interface FeedbackState {
history: FeedbackRecord[]
learnedProfiles: Record<ContextType, LearnedProfile>
}
// ββ Constants ββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
const EMA_ALPHA = 0.3 // Weight for new observations (higher = faster learning)
const MAX_HISTORY = 500 // Cap feedback history to prevent unbounded growth
const MIN_SAMPLES_TO_APPLY = 3 // Minimum feedback samples before learned adjustments kick in
const MAX_LEARNED_WEIGHT = 0.5 // Maximum influence of learned adjustments (50%)
const SAMPLES_FOR_MAX_WEIGHT = 20 // Samples needed to reach maximum weight
// Neutral starting params (middle of each range, used to initialize EMA)
const NEUTRAL_PARAMS: AutoTuneParams = {
temperature: 0.7,
top_p: 0.9,
top_k: 50,
frequency_penalty: 0.2,
presence_penalty: 0.2,
repetition_penalty: 1.1
}
// ββ Heuristics Engine ββββββββββββββββββββββββββββββββββββββββββββββββ
/**
* Compute automated quality heuristics for a response.
* These supplement user ratings to provide signal even without explicit feedback.
*/
export function computeHeuristics(response: string): ResponseHeuristics {
const responseLength = response.length
// Repetition score: check for repeated n-grams
const repetitionScore = computeRepetitionScore(response)
// Average sentence length
const sentences = response.split(/[.!?]+/).filter(s => s.trim().length > 0)
const averageSentenceLength = sentences.length > 0
? sentences.reduce((sum, s) => sum + s.trim().split(/\s+/).length, 0) / sentences.length
: 0
// Vocabulary diversity: unique words / total words
const words = response.toLowerCase().split(/\s+/).filter(w => w.length > 0)
const uniqueWords = new Set(words)
const vocabularyDiversity = words.length > 0 ? uniqueWords.size / words.length : 1
return {
responseLength,
repetitionScore,
averageSentenceLength,
vocabularyDiversity
}
}
/**
* Detect repetition by checking for repeated 3-gram sequences.
* Returns 0.0 (no repetition) to 1.0 (extremely repetitive).
*/
function computeRepetitionScore(text: string): number {
const words = text.toLowerCase().split(/\s+/).filter(w => w.length > 0)
if (words.length < 6) return 0
const trigrams = new Map<string, number>()
let totalTrigrams = 0
for (let i = 0; i <= words.length - 3; i++) {
const trigram = `${words[i]} ${words[i + 1]} ${words[i + 2]}`
trigrams.set(trigram, (trigrams.get(trigram) || 0) + 1)
totalTrigrams++
}
if (totalTrigrams === 0) return 0
// Count trigrams that appear more than once
let repeatedCount = 0
trigrams.forEach((count) => {
if (count > 1) {
repeatedCount += count - 1
}
})
return Math.min(repeatedCount / totalTrigrams, 1.0)
}
// ββ Learning Engine ββββββββββββββββββββββββββββββββββββββββββββββββββ
/**
* Create initial empty feedback state.
*/
export function createInitialFeedbackState(): FeedbackState {
const contexts: ContextType[] = ['code', 'creative', 'analytical', 'conversational', 'chaotic']
const learnedProfiles: Record<string, LearnedProfile> = {}
for (const ctx of contexts) {
learnedProfiles[ctx] = {
contextType: ctx,
sampleCount: 0,
positiveCount: 0,
negativeCount: 0,
positiveParams: { ...NEUTRAL_PARAMS },
negativeParams: { ...NEUTRAL_PARAMS },
adjustments: {},
lastUpdated: 0
}
}
return {
history: [],
learnedProfiles: learnedProfiles as Record<ContextType, LearnedProfile>
}
}
/**
* Process a new feedback record and update learned profiles using EMA.
* Returns the updated feedback state.
*/
export function processFeedback(
state: FeedbackState,
record: FeedbackRecord
): FeedbackState {
// Add to history (capped)
const newHistory = [...state.history, record]
if (newHistory.length > MAX_HISTORY) {
newHistory.splice(0, newHistory.length - MAX_HISTORY)
}
// Update the learned profile for this context type
const profile = { ...state.learnedProfiles[record.contextType] }
profile.sampleCount++
profile.lastUpdated = Date.now()
if (record.rating === 1) {
// Positive feedback: update positive EMA
profile.positiveCount++
profile.positiveParams = emaUpdate(profile.positiveParams, record.params, EMA_ALPHA)
} else {
// Negative feedback: update negative EMA
profile.negativeCount++
profile.negativeParams = emaUpdate(profile.negativeParams, record.params, EMA_ALPHA)
}
// Recompute adjustments: direction from negative toward positive
profile.adjustments = computeAdjustments(profile)
const newProfiles = {
...state.learnedProfiles,
[record.contextType]: profile
}
return {
history: newHistory,
learnedProfiles: newProfiles
}
}
/**
* EMA update: new_value = (1 - alpha) * old_value + alpha * observation
*/
function emaUpdate(
current: AutoTuneParams,
observation: AutoTuneParams,
alpha: number
): AutoTuneParams {
const inv = 1 - alpha
return {
temperature: current.temperature * inv + observation.temperature * alpha,
top_p: current.top_p * inv + observation.top_p * alpha,
top_k: Math.round(current.top_k * inv + observation.top_k * alpha),
frequency_penalty: current.frequency_penalty * inv + observation.frequency_penalty * alpha,
presence_penalty: current.presence_penalty * inv + observation.presence_penalty * alpha,
repetition_penalty: current.repetition_penalty * inv + observation.repetition_penalty * alpha
}
}
/**
* Compute parameter adjustments based on the difference between
* positively-rated and negatively-rated parameter EMAs.
*
* The idea: if upvoted responses used temp=0.3 and downvoted used temp=0.8,
* the adjustment nudges temperature downward.
*/
function computeAdjustments(profile: LearnedProfile): Partial<AutoTuneParams> {
// Need both positive and negative samples to compute a meaningful delta
if (profile.positiveCount < 1 || profile.negativeCount < 1) {
// With only positive data, use the delta from neutral as a mild nudge
if (profile.positiveCount >= MIN_SAMPLES_TO_APPLY) {
return computeDeltaFromNeutral(profile.positiveParams, 0.5)
}
return {}
}
const adj: Partial<AutoTuneParams> = {}
const keys: (keyof AutoTuneParams)[] = [
'temperature', 'top_p', 'top_k',
'frequency_penalty', 'presence_penalty', 'repetition_penalty'
]
for (const key of keys) {
const posDelta = profile.positiveParams[key] - NEUTRAL_PARAMS[key]
const negDelta = profile.negativeParams[key] - NEUTRAL_PARAMS[key]
// Push toward positive, away from negative
const adjustment = (posDelta - negDelta) * 0.5
// Only include non-trivial adjustments
if (Math.abs(adjustment) > 0.01) {
adj[key] = adjustment
}
}
return adj
}
/**
* Compute a mild adjustment from neutral toward the positively-rated average.
*/
function computeDeltaFromNeutral(
positiveParams: AutoTuneParams,
scale: number
): Partial<AutoTuneParams> {
const adj: Partial<AutoTuneParams> = {}
const keys: (keyof AutoTuneParams)[] = [
'temperature', 'top_p', 'top_k',
'frequency_penalty', 'presence_penalty', 'repetition_penalty'
]
for (const key of keys) {
const delta = (positiveParams[key] - NEUTRAL_PARAMS[key]) * scale
if (Math.abs(delta) > 0.01) {
adj[key] = delta
}
}
return adj
}
// ββ Integration with AutoTune ββββββββββββββββββββββββββββββββββββββββ
/**
* Apply learned adjustments to a base parameter set.
* Weight is determined by how much feedback data exists for this context.
*
* Returns the adjusted params and a note about what was applied.
*/
export function applyLearnedAdjustments(
baseParams: AutoTuneParams,
contextType: ContextType,
learnedProfiles: Record<ContextType, LearnedProfile>
): { params: AutoTuneParams; applied: boolean; note: string } {
const profile = learnedProfiles[contextType]
if (!profile || profile.sampleCount < MIN_SAMPLES_TO_APPLY || Object.keys(profile.adjustments).length === 0) {
return { params: baseParams, applied: false, note: '' }
}
// Weight scales from 0 to MAX_LEARNED_WEIGHT based on sample count
const weight = Math.min(
(profile.sampleCount / SAMPLES_FOR_MAX_WEIGHT) * MAX_LEARNED_WEIGHT,
MAX_LEARNED_WEIGHT
)
const adjusted = { ...baseParams }
const appliedKeys: string[] = []
for (const [key, delta] of Object.entries(profile.adjustments)) {
const k = key as keyof AutoTuneParams
if (delta !== undefined) {
adjusted[k] = (adjusted[k] as number) + (delta as number) * weight
appliedKeys.push(key)
}
}
const note = `Learned: ${appliedKeys.length} params adjusted (${profile.sampleCount} samples, ${Math.round(weight * 100)}% weight)`
return { params: adjusted, applied: true, note }
}
// ββ Stats / Display Helpers ββββββββββββββββββββββββββββββββββββββββββ
/**
* Get summary stats for the feedback learning system.
*/
export function getFeedbackStats(state: FeedbackState): {
totalFeedback: number
positiveRate: number
contextBreakdown: Record<ContextType, { total: number; positive: number; negative: number; hasLearned: boolean }>
oldestRecord: number | null
newestRecord: number | null
} {
const contexts: ContextType[] = ['code', 'creative', 'analytical', 'conversational', 'chaotic']
const totalFeedback = state.history.length
const positiveCount = state.history.filter(r => r.rating === 1).length
const contextBreakdown = {} as Record<ContextType, { total: number; positive: number; negative: number; hasLearned: boolean }>
for (const ctx of contexts) {
const profile = state.learnedProfiles[ctx]
contextBreakdown[ctx] = {
total: profile.sampleCount,
positive: profile.positiveCount,
negative: profile.negativeCount,
hasLearned: profile.sampleCount >= MIN_SAMPLES_TO_APPLY && Object.keys(profile.adjustments).length > 0
}
}
return {
totalFeedback,
positiveRate: totalFeedback > 0 ? positiveCount / totalFeedback : 0,
contextBreakdown,
oldestRecord: state.history.length > 0 ? state.history[0].timestamp : null,
newestRecord: state.history.length > 0 ? state.history[state.history.length - 1].timestamp : null
}
}
|