Files
Siteforge/scripts/evaluate-chatbot.mjs

138 lines
4.2 KiB
JavaScript

import fs from 'node:fs/promises'
import path from 'node:path'
const root = process.cwd()
const dataPath = path.join(root, 'data', 'chatbot-content.json')
const evalPath = path.join(root, 'data', 'chatbot-eval.json')
const STOP_WORDS = new Set([
'a','an','the','is','are','was','were','be','been','being','have','has','had','do','does','did',
'will','would','could','should','may','might','shall','can','i','you','he','she','it','we','they',
'me','him','her','us','them','my','your','his','its','our','their','this','that','these','those',
'and','but','or','nor','so','yet','for','of','in','on','at','to','from','with','by','about',
'what','how','why','when','where','who','which','if','then','than','as','just','not',
])
const TOKEN_ALIASES = {
bible: ['translation', 'version', 'scripture', 'bsb', 'berean'],
translation: ['version', 'bsb', 'berean', 'bible'],
version: ['translation', 'bsb', 'berean', 'bible'],
elders: ['elder', 'leadership', 'leaders', 'overseer', 'pastor'],
leadership: ['elders', 'elder', 'overseer', 'leaders'],
grace: ['salvation', 'saved', 'godliness', 'mercy'],
salvation: ['saved', 'grace', 'mercy', 'gospel'],
saved: ['salvation', 'grace', 'mercy', 'gospel'],
hope: ['blessed', 'appearing', 'return', 'coming'],
politics: ['public', 'government', 'authorities'],
}
function tokenize(text) {
return text
.toLowerCase()
.replace(/[^a-z0-9\s]/g, ' ')
.split(/\s+/)
.filter(token => token.length > 2 && !STOP_WORDS.has(token))
}
function expandTokens(tokens) {
const expanded = new Set(tokens)
for (const token of tokens) {
const aliases = TOKEN_ALIASES[token] ?? []
for (const alias of aliases) {
for (const aliasToken of tokenize(alias)) expanded.add(aliasToken)
}
}
return [...expanded]
}
function literalTerms(text) {
return [...new Set(
text
.toLowerCase()
.replace(/[^a-z0-9:\-\s']/g, ' ')
.split(/\s+/)
.map(term => term.trim())
.filter(term => term.length >= 2 && !STOP_WORDS.has(term)),
)]
}
function scoreEntry(entry, query) {
const indexText = `${entry.title} ${entry.content} ${entry.keywords.join(' ')}`.toLowerCase()
const queryTokens = expandTokens(tokenize(query))
const terms = literalTerms(query)
let score = 0
for (const token of queryTokens) {
if (entry.title.toLowerCase().includes(token)) score += 4
else if (indexText.includes(token)) score += 2
}
for (const term of terms) {
if (entry.title.toLowerCase().includes(term)) score += 2
else if (indexText.includes(term)) score += 1
}
const phrase = query.toLowerCase().replace(/[^a-z0-9:\-\s']/g, ' ').replace(/\s+/g, ' ').trim()
if (phrase.length > 6) {
if (entry.title.toLowerCase().includes(phrase)) score += 10
else if (indexText.includes(phrase)) score += 6
}
return score
}
async function main() {
const [contentRaw, evalRaw] = await Promise.all([
fs.readFile(dataPath, 'utf8'),
fs.readFile(evalPath, 'utf8'),
])
const entries = JSON.parse(contentRaw)
const tests = JSON.parse(evalRaw)
let pass = 0
const failures = []
for (const test of tests) {
const ranked = entries
.map(entry => ({ entry, score: scoreEntry(entry, test.query) }))
.sort((a, b) => b.score - a.score)
const topK = ranked.slice(0, test.expectedTopK)
const hit = topK.some(item => item.entry.id === test.expectedEntryId)
if (hit) {
pass += 1
continue
}
failures.push({
query: test.query,
expectedEntryId: test.expectedEntryId,
expectedTopK: test.expectedTopK,
actualTop: topK.map(item => ({ id: item.entry.id, title: item.entry.title, score: item.score })),
})
}
const total = tests.length
const pct = ((pass / total) * 100).toFixed(1)
console.log(`Chatbot eval: ${pass}/${total} (${pct}%) passed`)
if (failures.length > 0) {
console.log('\nFailures:')
for (const failure of failures) {
console.log(`- Query: ${failure.query}`)
console.log(` Expected: ${failure.expectedEntryId} in top ${failure.expectedTopK}`)
console.log(` Actual: ${failure.actualTop.map(item => `${item.id} (${item.score})`).join(', ')}`)
}
process.exitCode = 1
}
}
main().catch(error => {
console.error(error)
process.exitCode = 1
})