From 955af1513c5809a6137c3c0cb5745d0d5de8ceca Mon Sep 17 00:00:00 2001 From: Afloat Date: Wed, 18 Nov 2026 20:42:10 +0800 Subject: [PATCH] fix: normalize plan text once per source Reuse the stripped, lowercased plan string across token matches instead of normalizing it for every token. Preserve extraction and search results while avoiding repeated full-text scans on CJK and mixed-language plans. Add a reproducible benchmark using actual compressed blocks and the plan-aware search entry point. Fixes ranxianglei/billion-context#7160 --- scripts/bench-plan-aware-search.ts | 77 ++++++++++++++++++++++++++++++ src/decompress-shared.ts | 4 +- 2 files changed, 79 insertions(+), 2 deletions(-) create mode 100644 scripts/bench-plan-aware-search.ts diff --git a/scripts/bench-plan-aware-search.ts b/scripts/bench-plan-aware-search.ts new file mode 100644 index 0000000000..dcb0d86b27 --- /dev/null +++ b/scripts/bench-plan-aware-search.ts @@ -0,0 +1,77 @@ +// Run with --import ./tests/preload-test-env.mjs --import tsx. +// Optional argument: module URL relative to this script (for a saved base). +import assert from "node:assert/strict"; +import { createHash } from "node:crypto"; +import { performance } from "node:perf_hooks"; +import { createCore, defaultConfig, type Config, type CoreMessage } from "acp-kernel"; +import { applyCompressSettings } from "../src/compress-settings.ts"; +import { createSession } from "../src/session.ts"; +import { applyRanges } from "../src/stream.ts"; +import { parseCompressInput } from "../src/compress-tool.ts"; + +const { executeSearchContext, extractPlanState } = await import( + new URL(process.argv[2] ?? "../src/decompress-shared.ts", import.meta.url).href +) as typeof import("../src/decompress-shared.ts"); +const core = createCore(); +const config = applyCompressSettings(defaultConfig(200_000), 200_000, { + absorb: { enabled: true, minToolTokens: 50 }, + ccr: { enabled: true, minToolTokens: 50 }, +}) as Config; +const session = createSession("plan-normalization-benchmark"); +const messages: CoreMessage[] = Array.from({ length: 40 }, (_, i) => ({ + id: `history_${i}`, role: i % 2 === 0 ? "user" : "assistant", contentType: "text", + text: `Historical detail ${i}. ${"retained context evidence. ".repeat(80)}`, + timestamp: 1_700_000_000_000 + i, +})); +const turn = core.processTurn({ messages, state: session.state, config, tokenCount: 9999, renderTags: "text-only" }); +session.state = turn.state; +const topics = ["auth", "storage", "validation", "refresh", "ratelimit 缓存路由会话", "metrics"]; +applyRanges(parseCompressInput({ content: topics.map((topic, i) => ({ + startId: `m${String(i * 5 + 1).padStart(5, "0")}`, + endId: `m${String(i * 5 + 5).padStart(5, "0")}`, + topic: `token ${topic}`, summary: `Historical token ${topic} implementation and verification.`, +})) }), { core, config, messages: turn.messages, session, log: () => {} }); +assert.equal(session.state.blocks.filter(b => b.active).length, 6); + +function measure(work: () => string) { + const expected = work(); + for (let i = 0; i < 3; i++) assert.equal(work(), expected); + const elapsed: number[] = []; + for (let i = 0; i < 15; i++) { + const start = performance.now(); + const actual = work(); + elapsed.push(performance.now() - start); + assert.equal(actual, expected); + } + elapsed.sort((a, b) => a - b); + return { + medianMs: elapsed[7], minMs: elapsed[0], maxMs: elapsed[14], + digest: createHash("sha256").update(expected).digest("hex"), + }; +} + +const languages = { + latin: "Implement Ratelimit Middleware And Preserve Session Identity With Tests. ", + cjk: "实现缓存路由并保留会话身份验证和回归测试。", + mixed: "实现Cache路由API和Session身份Verification测试。", +}; +for (const [language, sentence] of Object.entries(languages)) { + for (const characters of [1000, 2000, 4000]) { + const text = sentence.repeat(Math.ceil(characters / sentence.length)).slice(0, characters); + const plan: CoreMessage[] = [ + { id: "user", role: "user", contentType: "text", text }, + { id: "plan", role: "assistant", contentType: "tool-call", toolName: "TodoWrite", + text: JSON.stringify({ todos: [{ content: text, status: "in_progress", activeForm: "Implementing the plan" }] }) }, + ]; + const opts = { messages: plan, session, log: () => {}, protectedPatterns: config.protectedLatestTools }; + const search = () => executeSearchContext({ query: "token", limit: 5 }, core, session.state, undefined, opts).text; + assert.match(search(), /\[plan-aware\]/); + console.log(JSON.stringify({ language, characters, sources: plan.length, workload: "extract", ...measure( + () => JSON.stringify([...extractPlanState(plan)!.terms]), + ) })); + console.log(JSON.stringify({ language, characters, sources: plan.length, workload: "search", ...measure(search) })); + } +} +console.log(JSON.stringify({ workload: "plain-search", ...measure( + () => executeSearchContext({ query: "token", limit: 5 }, core, session.state).text, +) })); diff --git a/src/decompress-shared.ts b/src/decompress-shared.ts index ae17b50e0a..c02ab4eb96 100644 --- a/src/decompress-shared.ts +++ b/src/decompress-shared.ts @@ -556,10 +556,10 @@ type PlanTerms = Map; function planTokenize(text: string): string[] { const out: string[] = []; // Rendered ref-tag markup () is proxy metadata, not plan content. - const clean = text.replace(/\x3cacp\b[\s\S]*?\x3c\/acp\x3e/g, " "); + const clean = text.replace(/\x3cacp\b[\s\S]*?\x3c\/acp\x3e/g, " ").toLowerCase(); const re = /[a-z0-9]+|[\u4e00-\u9fff]/g; let m: RegExpExecArray | null; - while ((m = re.exec(clean.toLowerCase())) !== null) { + while ((m = re.exec(clean)) !== null) { const t = m[0]; const cjk = /[\u4e00-\u9fff]/.test(t); // Latin tokens need >= 3 chars to carry signal; CJK single chars count.