feat(memory): bm25 ranking and the rerank prompt

This commit is contained in:
bchanot committed 2026-11-05 16:18:39 +01:00
1 parent 6adb05ce1e
commit a75efabbf1
2 files changed
+200 -1

No files matched your search

+97 -1
View File
@@ -1,10 +1,12 @@
import { test, expect, mock } from 'claude-code/testing'
import type { Engine, MockClock } from 'claude-code/testing'
import type { On } from 'claude-code'
import type { Entry, Kind } from './register'
import {
cap, tail, redact, normalize, failureLine, sessionOf, addFailure,
entriesOf, parseIndex, isLive, lessonOf, assembleEntries, grams,
buildIndex, matchOf, deErrno,
buildIndex, matchOf, deErrno, tokensOf, bm25Rank, rerankPrompt,
parseIds, recallLine, recallBlock, failureKey,
segments, suiteKeyOf, commitLabel, finishLabel,
} from './register'
@@ -1832,3 +1834,97 @@ test('TIMEerrno: deErrno on 64 KiB of a phrase prefix, under 50 ms', () => {
deErrno('no such file or '.repeat(4096))
expect(now() - t0).toBeLessThan(50)
})
const mk = (id: string, title: string, lesson: string, body = ''): Entry =>
({ id, kind: id.slice(0, 3) as Kind, title, lesson, body })
const RB_ENTRIES = [
mk('LRN-001', 'css grid gaps', 'Use gap.', 'layout of cards'),
mk('BLK-002', 'heredoc quoting breaks', 'Quote the delimiter.',
'shell heredoc quoting eats the dollar sign in shell scripts'),
mk('BDR-003', 'store keys', 'Prefix them.', 'kv store naming'),
]
test('RB1: bm25Rank puts the heredoc entry first, drops score 0', () => {
const ids = bm25Rank(RB_ENTRIES, 'quoting heredoc shell', 5).map(e => e.id)
expect(ids).toEqual(['BLK-002'])
expect(bm25Rank(RB_ENTRIES, 'the and for', 5)).toEqual([])
expect(tokensOf('The Heredoc, a b ÉTÉ 42')).toEqual(['heredoc', 'été', '42'])
})
test('RB1b: ties go by numeric ID order, k bounds the result', () => {
const twin = [mk('LRN-10', 'zed', ''), mk('LRN-9', 'zed', '')]
expect(bm25Rank(twin, 'zed', 5).map(e => e.id)).toEqual(['LRN-9', 'LRN-10'])
expect(bm25Rank(twin, 'zed', 1).map(e => e.id)).toEqual(['LRN-9'])
})
test('RB2: parseIds keeps candidate IDs once, in order, up to max', () => {
const cands = [mk('BLK-041', 't', ''), mk('LRN-258', 't', '')]
expect(parseIds('BLK-041, BDR-999, LRN-258', cands, 3))
.toEqual(['BLK-041', 'LRN-258'])
expect(parseIds('LRN-258 BLK-041 LRN-258', cands, 1)).toEqual(['LRN-258'])
expect(parseIds('none', cands, 3)).toEqual([])
})
test('RB3: long Decisions give ellipsis lessons and a block under cap', () => {
const field = 'Use alpha ' + 'x'.repeat(1990)
const es = ['BDR-001', 'BDR-002', 'BDR-003'].map(id =>
assembleEntries([`## ${id} — t\n- **Decision**: ${field}\n`],
new Map([[id, { status: 'accepted', title: 'T' + id }]]))[0]!)
for (const e of es) expect(recallLine(e).endsWith('…')).toBe(true)
const block = recallBlock(es.map(recallLine), 1500)
expect(block.length).toBeLessThanOrEqual(1500)
expect(block.split('\n')[0]).toBe('memory recall:')
expect(block.split('\n')[1]).toBe(
'BDR-001 TBDR-001 — Use alpha ' + 'x'.repeat(280 - 11) + '…')
})
test('RB3b: recallLine and recallBlock edges', () => {
expect(recallLine(mk('LRN-1', 'T', 'a;'))).toBe('LRN-1 T — a;')
expect(recallLine(mk('LRN-1', 'T', 'Use X.'))).toBe('LRN-1 T — Use X.')
expect(recallLine(mk('LRN-1', 'T', ''))).toBe('LRN-1 T')
expect(recallBlock(['a'.repeat(50)], 20)).toBe('')
expect(recallBlock(['a'.repeat(50), 'bb'], 20)).toBe('memory recall:\nbb')
expect(recallBlock([], 100)).toBe('')
})
test('RB4: rerankPrompt is the literal prompt', () => {
const cands = [mk('LRN-1', 'T one', 'Do a.'), mk('BLK-2', 'T two', '')]
expect(rerankPrompt('fix the heredoc', cands)).toBe(
'Prompt: fix the heredoc\nCandidates:\nLRN-1 | T one | Do a.\n' +
'BLK-2 | T two | \nAnswer with at most 3 candidate IDs that would ' +
'help with this prompt, comma separated, or none.')
})
test('RB5: 500 entries and a 64 KiB query rank under 300 ms', () => {
const es = Array.from({ length: 500 }, (_, i) =>
mk(`LRN-${i + 1}`, `t${i}`, 'l', `w${i} word `.repeat(160)))
const head = 'w7 '.padEnd(2000, 'z ')
const t0 = now()
const a = bm25Rank(es, head + 'w9 '.repeat(20000), 5).map(e => e.id)
expect(now() - t0).toBeLessThan(300)
const b = bm25Rank(es, head + 'w300 '.repeat(20000), 5).map(e => e.id)
expect(a).toEqual(b)
expect(a[0]).toBe('LRN-8')
})
test('RB6: a secret-shaped token is redacted in the rerank prompt', () => {
const secret = 'sk_live_' + 'a'.repeat(24)
const p = rerankPrompt(`use ${secret} now`, [])
expect(p).toContain('<redacted:stripe>')
expect(p).not.toContain(secret)
})
test('RB7: failureKey is the normalized last line of the redacted tail', () => {
expect(failureKey('Exit code 1\nrm: /tmp/x/y: Permission denied', 'rm x'))
.toBe('rm: <path>: permission denied')
expect(failureKey('', 'ls')).toBeUndefined()
})
test('TIMErank: tokensOf and parseIds on 64 KiB, under 50 ms', () => {
const t0 = now()
tokensOf('ab '.repeat(21000))
tokensOf('x'.repeat(65536))
parseIds('BDR-1 '.repeat(10000) + 'LRN-'.repeat(15000), [], 3)
expect(now() - t0).toBeLessThan(50)
})
+103
View File
@@ -622,6 +622,109 @@ async function statusMap($: Api, root: string): Promise<StatusMap | undefined> {
}
}
const STOP_WORDS = new Set(
('the and for with that this from are was were not but you your have has ' +
'had will can all any its into than then them they their there what ' +
'when which who how why les des une pour dans avec est pas que qui sur ' +
'par aux ces ses son mais ont sont plus comme').split(' '),
)
const TOKEN = /[\p{L}\p{N}]{2,40}/gu
const MAX_QUERY = 2000
const MAX_DOC = 65536
const BM25_K1 = 1.2
const BM25_B = 0.75
const MAX_PICKS = 3
/** Lowercase word runs (2-40 letters or digits) minus function words. */
export function tokensOf(text: string): string[] {
const words = cap(text, MAX_DOC).toLowerCase().match(TOKEN) ?? []
return words.filter(w => !STOP_WORDS.has(w))
}
type Doc = { entry: Entry; tf: Map<string, number>; len: number }
function docOf(entry: Entry): Doc {
const tokens = tokensOf(`${entry.title} ${entry.lesson} ${entry.body}`)
const tf = new Map<string, number>()
for (const t of tokens) tf.set(t, (tf.get(t) ?? 0) + 1)
return { entry, tf, len: tokens.length }
}
function scoreOf(
doc: Doc, terms: string[], df: Map<string, number>, n: number, avg: number,
): number {
let score = 0
for (const term of terms) {
const f = doc.tf.get(term) ?? 0
if (f === 0) continue
const d = df.get(term) ?? 0
const idf = Math.log(1 + (n - d + 0.5) / (d + 0.5))
const norm = 1 - BM25_B + (BM25_B * doc.len) / avg
score += (idf * f * (BM25_K1 + 1)) / (f + BM25_K1 * norm)
}
return score
}
/** The `k` best entries for a query (BM25), highest first, score > 0. */
export function bm25Rank(entries: Entry[], query: string, k: number): Entry[] {
const terms = [...new Set(tokensOf(cap(query, MAX_QUERY)))]
const docs = entries.map(docOf)
const df = new Map<string, number>()
for (const d of docs) for (const t of d.tf.keys()) {
df.set(t, (df.get(t) ?? 0) + 1)
}
const avg = docs.reduce((sum, d) => sum + d.len, 0) / (docs.length || 1)
return docs
.map(d => ({ e: d.entry, s: scoreOf(d, terms, df, docs.length, avg || 1) }))
.filter(x => x.s > 0)
.sort((a, b) => b.s - a.s ||
a.e.id.localeCompare(b.e.id, 'en', { numeric: true }))
.slice(0, k)
.map(x => x.e)
}
/** The haiku prompt: the redacted user prompt and one line per candidate. */
export function rerankPrompt(query: string, cands: Entry[]): string {
const lines = cands.map(c => `${c.id} | ${c.title} | ${c.lesson}`)
return (
`Prompt: ${redact(cap(query, MAX_QUERY))}\nCandidates:\n` +
`${lines.join('\n')}\nAnswer with at most ${MAX_PICKS} candidate IDs ` +
'that would help with this prompt, comma separated, or none.'
)
}
const ID_IN_ANSWER = /(?:BDR|LRN|BLK)-\d{1,5}/g
/** The candidate IDs named in a model answer, deduplicated, in order. */
export function parseIds(text: string, cands: Entry[], max: number): string[] {
const known = new Set(cands.map(c => c.id))
const found = cap(text, MAX_REGEX_INPUT).match(ID_IN_ANSWER) ?? []
return [...new Set(found)].filter(id => known.has(id)).slice(0, max)
}
/** One recall line; the lesson keeps its own terminator. */
export function recallLine(e: Entry): string {
const head = `${e.id} ${e.title}`
return e.lesson === '' ? head : `${head} — ${e.lesson}`
}
/** The injected block; a line that would pass `max` characters is dropped. */
export function recallBlock(lines: string[], max: number): string {
const header = 'memory recall:'
let block = header
for (const line of lines) {
if (block.length + 1 + line.length <= max) block += `\n${line}`
}
return block === header ? '' : block
}
/** The dedup key of a failure text, as `buildFailure` stores it. */
export function failureKey(text: string, command: string): string | undefined {
void command // the key comes from the output only; kept for the harness
const line = lastLine(redact(tail(text, MAX_REGEX_INPUT)))
return line === undefined ? undefined : normalize(line)
}
type Registries = {
entries: Entry[]
status: StatusMap | undefined