|
| 1 | +import assert from "node:assert/strict"; |
| 2 | +import { dirname, join } from "node:path"; |
| 3 | +import test from "node:test"; |
| 4 | +import { fileURLToPath } from "node:url"; |
| 5 | +import { buildQuestionContext } from "../src/ai/contextBuilder.js"; |
| 6 | +import { createProjectMap } from "../src/analyzers/projectMap.js"; |
| 7 | + |
| 8 | +const testDirectory = dirname(fileURLToPath(import.meta.url)); |
| 9 | +const fixtures = { |
| 10 | + nextjs: join(testDirectory, "fixtures", "nextjs-project"), |
| 11 | + express: join(testDirectory, "fixtures", "express-project") |
| 12 | +} as const; |
| 13 | + |
| 14 | +type ContextEvalCase = { |
| 15 | + name: string; |
| 16 | + fixture: keyof typeof fixtures; |
| 17 | + question: string; |
| 18 | + expectedFile: string; |
| 19 | +}; |
| 20 | + |
| 21 | +const evalCases: ContextEvalCase[] = [ |
| 22 | + { |
| 23 | + name: "authentication architecture", |
| 24 | + fixture: "nextjs", |
| 25 | + question: "How does authentication work?", |
| 26 | + expectedFile: "lib/auth.ts" |
| 27 | + }, |
| 28 | + { |
| 29 | + name: "Indonesian authentication", |
| 30 | + fixture: "nextjs", |
| 31 | + question: "Bagaimana autentikasi pengguna bekerja?", |
| 32 | + expectedFile: "lib/auth.ts" |
| 33 | + }, |
| 34 | + { |
| 35 | + name: "session implementation", |
| 36 | + fixture: "nextjs", |
| 37 | + question: "Where is the user session implemented?", |
| 38 | + expectedFile: "lib/auth.ts" |
| 39 | + }, |
| 40 | + { |
| 41 | + name: "login handling", |
| 42 | + fixture: "nextjs", |
| 43 | + question: "File mana yang menangani login?", |
| 44 | + expectedFile: "lib/auth.ts" |
| 45 | + }, |
| 46 | + { |
| 47 | + name: "NextAuth configuration", |
| 48 | + fixture: "nextjs", |
| 49 | + question: "Explain the NextAuth configuration", |
| 50 | + expectedFile: "lib/auth.ts" |
| 51 | + }, |
| 52 | + { |
| 53 | + name: "database client", |
| 54 | + fixture: "nextjs", |
| 55 | + question: "Where is the database client initialized?", |
| 56 | + expectedFile: "lib/db.ts" |
| 57 | + }, |
| 58 | + { |
| 59 | + name: "Indonesian database connection", |
| 60 | + fixture: "nextjs", |
| 61 | + question: "Di mana koneksi database dibuat?", |
| 62 | + expectedFile: "lib/db.ts" |
| 63 | + }, |
| 64 | + { |
| 65 | + name: "Prisma initialization", |
| 66 | + fixture: "nextjs", |
| 67 | + question: "Which file initializes Prisma?", |
| 68 | + expectedFile: "lib/db.ts" |
| 69 | + }, |
| 70 | + { |
| 71 | + name: "database access module", |
| 72 | + fixture: "nextjs", |
| 73 | + question: "Show the main db module", |
| 74 | + expectedFile: "lib/db.ts" |
| 75 | + }, |
| 76 | + { |
| 77 | + name: "session API endpoint", |
| 78 | + fixture: "nextjs", |
| 79 | + question: "Where is the session API endpoint?", |
| 80 | + expectedFile: "app/api/session/route.ts" |
| 81 | + }, |
| 82 | + { |
| 83 | + name: "Indonesian session route", |
| 84 | + fixture: "nextjs", |
| 85 | + question: "Route API mana yang mengembalikan sesi?", |
| 86 | + expectedFile: "app/api/session/route.ts" |
| 87 | + }, |
| 88 | + { |
| 89 | + name: "GET session handler", |
| 90 | + fixture: "nextjs", |
| 91 | + question: "Which GET endpoint handles session requests?", |
| 92 | + expectedFile: "app/api/session/route.ts" |
| 93 | + }, |
| 94 | + { |
| 95 | + name: "home page", |
| 96 | + fixture: "nextjs", |
| 97 | + question: "Where is the main home page?", |
| 98 | + expectedFile: "app/page.tsx" |
| 99 | + }, |
| 100 | + { |
| 101 | + name: "Indonesian main page", |
| 102 | + fixture: "nextjs", |
| 103 | + question: "File mana untuk page utama?", |
| 104 | + expectedFile: "app/page.tsx" |
| 105 | + }, |
| 106 | + { |
| 107 | + name: "root layout", |
| 108 | + fixture: "nextjs", |
| 109 | + question: "Where is the root layout component?", |
| 110 | + expectedFile: "app/layout.tsx" |
| 111 | + }, |
| 112 | + { |
| 113 | + name: "payment implementation", |
| 114 | + fixture: "express", |
| 115 | + question: "How are payments implemented?", |
| 116 | + expectedFile: "src/routes/payments.ts" |
| 117 | + }, |
| 118 | + { |
| 119 | + name: "Indonesian payment", |
| 120 | + fixture: "express", |
| 121 | + question: "Di mana pembayaran diproses?", |
| 122 | + expectedFile: "src/routes/payments.ts" |
| 123 | + }, |
| 124 | + { |
| 125 | + name: "Stripe setup", |
| 126 | + fixture: "express", |
| 127 | + question: "Which file configures Stripe?", |
| 128 | + expectedFile: "src/routes/payments.ts" |
| 129 | + }, |
| 130 | + { |
| 131 | + name: "Express entry point", |
| 132 | + fixture: "express", |
| 133 | + question: "Where does the Express server start?", |
| 134 | + expectedFile: "src/server.ts" |
| 135 | + }, |
| 136 | + { |
| 137 | + name: "Indonesian server entry", |
| 138 | + fixture: "express", |
| 139 | + question: "File mana yang menjalankan server utama?", |
| 140 | + expectedFile: "src/server.ts" |
| 141 | + } |
| 142 | +]; |
| 143 | + |
| 144 | +test("context builder benchmark keeps relevant files in the top results", async (testContext) => { |
| 145 | + const snapshots = { |
| 146 | + nextjs: await createProjectMap(fixtures.nextjs), |
| 147 | + express: await createProjectMap(fixtures.express) |
| 148 | + }; |
| 149 | + const failures: string[] = []; |
| 150 | + const topOneMisses: string[] = []; |
| 151 | + let topOneHits = 0; |
| 152 | + |
| 153 | + for (const evalCase of evalCases) { |
| 154 | + const context = await buildQuestionContext( |
| 155 | + fixtures[evalCase.fixture], |
| 156 | + snapshots[evalCase.fixture], |
| 157 | + evalCase.question |
| 158 | + ); |
| 159 | + const rankedPaths = context.files.map((file) => file.path); |
| 160 | + const expectedRank = rankedPaths.indexOf(evalCase.expectedFile); |
| 161 | + |
| 162 | + if (expectedRank === 0) { |
| 163 | + topOneHits += 1; |
| 164 | + } else { |
| 165 | + topOneMisses.push( |
| 166 | + `${evalCase.name}: expected ${evalCase.expectedFile}, got ${rankedPaths[0] ?? "none"}` |
| 167 | + ); |
| 168 | + } |
| 169 | + |
| 170 | + if (expectedRank < 0 || expectedRank >= 3) { |
| 171 | + failures.push( |
| 172 | + `${evalCase.name}: expected ${evalCase.expectedFile}, got ${rankedPaths.join(", ")}` |
| 173 | + ); |
| 174 | + } |
| 175 | + } |
| 176 | + |
| 177 | + assert.deepEqual(failures, []); |
| 178 | + assert.equal(topOneHits, evalCases.length); |
| 179 | + testContext.diagnostic(`Context Builder top-1 accuracy: ${topOneHits}/20`); |
| 180 | + testContext.diagnostic("Context Builder top-3 recall: 20/20"); |
| 181 | + for (const miss of topOneMisses) { |
| 182 | + testContext.diagnostic(`Top-1 miss: ${miss}`); |
| 183 | + } |
| 184 | +}); |
0 commit comments