MythEclipse/GMW
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

refactor(ai-analyzer): fix resource leaks, OOM risk, and strict structured outputs

Browse Source
This commit is contained in:
MythEclipse
2026-05-25 22:14:05 +07:00
parent c32b5274e9
commit cc2ee84c3b
6 changed files with 359 additions and 292 deletions
Download Patch File Download Diff File Expand all files Collapse all files
package.json
+1
View File
@@ -48,6 +48,7 @@
"opusscript": "^0.0.8",
"p-retry": "^8.0.0",
"pg": "^8.21.0",
"piscina": "^5.1.4",
"play-dl": "^1.9.7",
"prism-media": "2.0.0-alpha.0",
"prom-client": "^15.1.3",
pnpm-lock.yaml
Generated
+196
View File
@@ -74,6 +74,9 @@ importers:
pg:
specifier: ^8.21.0
version: 8.21.0
piscina:
specifier: ^5.1.4
version: 5.1.4
play-dl:
specifier: ^1.9.7
version: 1.9.7
@@ -1130,6 +1133,119 @@ packages:
'@minhducsun2002/leb128@1.0.0':
resolution: {integrity: sha512-eFrYUPDVHeuwWHluTG1kwNQUEUcFjVKYwPkU8z9DR1JH3AW7JtJsG9cRVGmwz809kKtGfwGJj58juCZxEvnI/g==}
'@napi-rs/nice-android-arm-eabi@1.1.1':
resolution: {integrity: sha512-kjirL3N6TnRPv5iuHw36wnucNqXAO46dzK9oPb0wj076R5Xm8PfUVA9nAFB5ZNMmfJQJVKACAPd/Z2KYMppthw==}
engines: {node: '>= 10'}
cpu: [arm]
os: [android]
'@napi-rs/nice-android-arm64@1.1.1':
resolution: {integrity: sha512-blG0i7dXgbInN5urONoUCNf+DUEAavRffrO7fZSeoRMJc5qD+BJeNcpr54msPF6qfDD6kzs9AQJogZvT2KD5nw==}
engines: {node: '>= 10'}
cpu: [arm64]
os: [android]
'@napi-rs/nice-darwin-arm64@1.1.1':
resolution: {integrity: sha512-s/E7w45NaLqTGuOjC2p96pct4jRfo61xb9bU1unM/MJ/RFkKlJyJDx7OJI/O0ll/hrfpqKopuAFDV8yo0hfT7A==}
engines: {node: '>= 10'}
cpu: [arm64]
os: [darwin]
'@napi-rs/nice-darwin-x64@1.1.1':
resolution: {integrity: sha512-dGoEBnVpsdcC+oHHmW1LRK5eiyzLwdgNQq3BmZIav+9/5WTZwBYX7r5ZkQC07Nxd3KHOCkgbHSh4wPkH1N1LiQ==}
engines: {node: '>= 10'}
cpu: [x64]
os: [darwin]
'@napi-rs/nice-freebsd-x64@1.1.1':
resolution: {integrity: sha512-kHv4kEHAylMYmlNwcQcDtXjklYp4FCf0b05E+0h6nDHsZ+F0bDe04U/tXNOqrx5CmIAth4vwfkjjUmp4c4JktQ==}
engines: {node: '>= 10'}
cpu: [x64]
os: [freebsd]
'@napi-rs/nice-linux-arm-gnueabihf@1.1.1':
resolution: {integrity: sha512-E1t7K0efyKXZDoZg1LzCOLxgolxV58HCkaEkEvIYQx12ht2pa8hoBo+4OB3qh7e+QiBlp1SRf+voWUZFxyhyqg==}
engines: {node: '>= 10'}
cpu: [arm]
os: [linux]
'@napi-rs/nice-linux-arm64-gnu@1.1.1':
resolution: {integrity: sha512-CIKLA12DTIZlmTaaKhQP88R3Xao+gyJxNWEn04wZwC2wmRapNnxCUZkVwggInMJvtVElA+D4ZzOU5sX4jV+SmQ==}
engines: {node: '>= 10'}
cpu: [arm64]
os: [linux]
libc: [glibc]
'@napi-rs/nice-linux-arm64-musl@1.1.1':
resolution: {integrity: sha512-+2Rzdb3nTIYZ0YJF43qf2twhqOCkiSrHx2Pg6DJaCPYhhaxbLcdlV8hCRMHghQ+EtZQWGNcS2xF4KxBhSGeutg==}
engines: {node: '>= 10'}
cpu: [arm64]
os: [linux]
libc: [musl]
'@napi-rs/nice-linux-ppc64-gnu@1.1.1':
resolution: {integrity: sha512-4FS8oc0GeHpwvv4tKciKkw3Y4jKsL7FRhaOeiPei0X9T4Jd619wHNe4xCLmN2EMgZoeGg+Q7GY7BsvwKpL22Tg==}
engines: {node: '>= 10'}
cpu: [ppc64]
os: [linux]
libc: [glibc]
'@napi-rs/nice-linux-riscv64-gnu@1.1.1':
resolution: {integrity: sha512-HU0nw9uD4FO/oGCCk409tCi5IzIZpH2agE6nN4fqpwVlCn5BOq0MS1dXGjXaG17JaAvrlpV5ZeyZwSon10XOXw==}
engines: {node: '>= 10'}
cpu: [riscv64]
os: [linux]
libc: [glibc]
'@napi-rs/nice-linux-s390x-gnu@1.1.1':
resolution: {integrity: sha512-2YqKJWWl24EwrX0DzCQgPLKQBxYDdBxOHot1KWEq7aY2uYeX+Uvtv4I8xFVVygJDgf6/92h9N3Y43WPx8+PAgQ==}
engines: {node: '>= 10'}
cpu: [s390x]
os: [linux]
libc: [glibc]
'@napi-rs/nice-linux-x64-gnu@1.1.1':
resolution: {integrity: sha512-/gaNz3R92t+dcrfCw/96pDopcmec7oCcAQ3l/M+Zxr82KT4DljD37CpgrnXV+pJC263JkW572pdbP3hP+KjcIg==}
engines: {node: '>= 10'}
cpu: [x64]
os: [linux]
libc: [glibc]
'@napi-rs/nice-linux-x64-musl@1.1.1':
resolution: {integrity: sha512-xScCGnyj/oppsNPMnevsBe3pvNaoK7FGvMjT35riz9YdhB2WtTG47ZlbxtOLpjeO9SqqQ2J2igCmz6IJOD5JYw==}
engines: {node: '>= 10'}
cpu: [x64]
os: [linux]
libc: [musl]
'@napi-rs/nice-openharmony-arm64@1.1.1':
resolution: {integrity: sha512-6uJPRVwVCLDeoOaNyeiW0gp2kFIM4r7PL2MczdZQHkFi9gVlgm+Vn+V6nTWRcu856mJ2WjYJiumEajfSm7arPQ==}
engines: {node: '>= 10'}
cpu: [arm64]
os: [openharmony]
'@napi-rs/nice-win32-arm64-msvc@1.1.1':
resolution: {integrity: sha512-uoTb4eAvM5B2aj/z8j+Nv8OttPf2m+HVx3UjA5jcFxASvNhQriyCQF1OB1lHL43ZhW+VwZlgvjmP5qF3+59atA==}
engines: {node: '>= 10'}
cpu: [arm64]
os: [win32]
'@napi-rs/nice-win32-ia32-msvc@1.1.1':
resolution: {integrity: sha512-CNQqlQT9MwuCsg1Vd/oKXiuH+TcsSPJmlAFc5frFyX/KkOh0UpBLEj7aoY656d5UKZQMQFP7vJNa1DNUNORvug==}
engines: {node: '>= 10'}
cpu: [ia32]
os: [win32]
'@napi-rs/nice-win32-x64-msvc@1.1.1':
resolution: {integrity: sha512-vB+4G/jBQCAh0jelMTY3+kgFy00Hlx2f2/1zjMoH821IbplbWZOkLiTYXQkygNTzQJTq5cvwBDgn2ppHD+bglQ==}
engines: {node: '>= 10'}
cpu: [x64]
os: [win32]
'@napi-rs/nice@1.1.1':
resolution: {integrity: sha512-xJIPs+bYuc9ASBl+cvGsKbGrJmS6fAKaSZCnT0lhahT5rhA2VVy9/EcIgd2JhtEuFOJNx7UHNn/qiTPTY4nrQw==}
engines: {node: '>= 10'}
'@napi-rs/wasm-runtime@1.1.4':
resolution: {integrity: sha512-3NQNNgA1YSlJb/kMH1ildASP9HW7/7kYnRI2szWJaofaS1hWmbGI4H+d3+22aGzXXN9IJ+n+GiFVcGipJP18ow==}
peerDependencies:
@@ -3713,6 +3829,10 @@ packages:
resolution: {integrity: sha512-QP88BAKvMam/3NxH6vj2o21R6MjxZUAd6nlwAS/pnGvN9IVLocLHxGYIzFhg6fUQ+5th6P4dv4eW9jX3DSIj7A==}
engines: {node: '>=12'}
piscina@5.1.4:
resolution: {integrity: sha512-7uU4ZnKeQq22t9AsmHGD2w4OYQGonwFnTypDypaWi7Qr2EvQIFVtG8J5D/3bE7W123Wdc9+v4CZDu5hJXVCtBg==}
engines: {node: '>=20.x'}
pkg-pr-new@0.0.62:
resolution: {integrity: sha512-K2jtf1PLCJJFDimQIpasPXDdnRTZSYPy36Ldz+QhMpLz2YN1Wi0ZQkomVt5Wi3NdBZcYuYGXekyFWfJ6fAHYjg==}
hasBin: true
@@ -5346,6 +5466,78 @@ snapshots:
'@minhducsun2002/leb128@1.0.0': {}
'@napi-rs/nice-android-arm-eabi@1.1.1':
optional: true
'@napi-rs/nice-android-arm64@1.1.1':
optional: true
'@napi-rs/nice-darwin-arm64@1.1.1':
optional: true
'@napi-rs/nice-darwin-x64@1.1.1':
optional: true
'@napi-rs/nice-freebsd-x64@1.1.1':
optional: true
'@napi-rs/nice-linux-arm-gnueabihf@1.1.1':
optional: true
'@napi-rs/nice-linux-arm64-gnu@1.1.1':
optional: true
'@napi-rs/nice-linux-arm64-musl@1.1.1':
optional: true
'@napi-rs/nice-linux-ppc64-gnu@1.1.1':
optional: true
'@napi-rs/nice-linux-riscv64-gnu@1.1.1':
optional: true
'@napi-rs/nice-linux-s390x-gnu@1.1.1':
optional: true
'@napi-rs/nice-linux-x64-gnu@1.1.1':
optional: true
'@napi-rs/nice-linux-x64-musl@1.1.1':
optional: true
'@napi-rs/nice-openharmony-arm64@1.1.1':
optional: true
'@napi-rs/nice-win32-arm64-msvc@1.1.1':
optional: true
'@napi-rs/nice-win32-ia32-msvc@1.1.1':
optional: true
'@napi-rs/nice-win32-x64-msvc@1.1.1':
optional: true
'@napi-rs/nice@1.1.1':
optionalDependencies:
'@napi-rs/nice-android-arm-eabi': 1.1.1
'@napi-rs/nice-android-arm64': 1.1.1
'@napi-rs/nice-darwin-arm64': 1.1.1
'@napi-rs/nice-darwin-x64': 1.1.1
'@napi-rs/nice-freebsd-x64': 1.1.1
'@napi-rs/nice-linux-arm-gnueabihf': 1.1.1
'@napi-rs/nice-linux-arm64-gnu': 1.1.1
'@napi-rs/nice-linux-arm64-musl': 1.1.1
'@napi-rs/nice-linux-ppc64-gnu': 1.1.1
'@napi-rs/nice-linux-riscv64-gnu': 1.1.1
'@napi-rs/nice-linux-s390x-gnu': 1.1.1
'@napi-rs/nice-linux-x64-gnu': 1.1.1
'@napi-rs/nice-linux-x64-musl': 1.1.1
'@napi-rs/nice-openharmony-arm64': 1.1.1
'@napi-rs/nice-win32-arm64-msvc': 1.1.1
'@napi-rs/nice-win32-ia32-msvc': 1.1.1
'@napi-rs/nice-win32-x64-msvc': 1.1.1
optional: true
'@napi-rs/wasm-runtime@1.1.4(@emnapi/core@1.10.0)(@emnapi/runtime@1.10.0)':
dependencies:
'@emnapi/core': 1.10.0
@@ -7851,6 +8043,10 @@ snapshots:
picomatch@4.0.4: {}
piscina@5.1.4:
optionalDependencies:
'@napi-rs/nice': 1.1.1
pkg-pr-new@0.0.62:
dependencies:
'@actions/core': 1.11.1
src/moderation/aiAnalysisWorker.ts
+22 -18
View File
@@ -1,4 +1,3 @@
import { parentPort } from "node:worker_threads";
import { config } from "../config.js";
import { initializeDatabase } from "../database/drizzle.js";
import { buildConversationPromptMessages } from "./conversationContext.js";
@@ -6,18 +5,29 @@ import { runModerationAnalysis } from "./llmModerationClient.js";
import {
getAttachmentsForMessages,
getConversationContextBefore,
updateMessageAIAnalysis,
updateMessagesAIAnalysisBulk,
} from "./messageStore.js";
import type { MessageRecord } from "./types.js";
let dbInitialized = false;
let dbInitPromise: Promise<any> | null = null;
interface AnalysisWorkerRequest {
async function ensureDb() {
if (dbInitialized) return;
if (!dbInitPromise) {
dbInitPromise = initializeDatabase().then(() => {
dbInitialized = true;
});
}
await dbInitPromise;
}
export interface AnalysisWorkerRequest {
conversationKey: string;
messages: MessageRecord[];
}
type AnalysisWorkerResponse =
export type AnalysisWorkerResponse =
| {
ok: true;
conversationKey: string;
@@ -30,16 +40,13 @@ type AnalysisWorkerResponse =
error: string;
};
async function processAnalysisRequest({
export default async function processAnalysisRequest({
conversationKey,
messages,
}: AnalysisWorkerRequest): Promise<AnalysisWorkerResponse> {
try {
try {
if (!dbInitialized) {
await initializeDatabase();
dbInitialized = true;
}
await ensureDb();
} catch (dbError) {
const msg = dbError instanceof Error ? dbError.message : String(dbError);
return {
@@ -77,9 +84,9 @@ async function processAnalysisRequest({
attachments,
});
const rows: MessageRecord[] = [];
for (const analysisResult of result.results) {
const row = await updateMessageAIAnalysis(analysisResult.messageId, {
const updates = result.results.map((analysisResult) => ({
messageId: analysisResult.messageId,
result: {
status: analysisResult.status,
flags: JSON.stringify(analysisResult.flags),
score: analysisResult.score,
@@ -87,9 +94,10 @@ async function processAnalysisRequest({
analysis: analysisResult.analysis,
analyzedAt: Date.now(),
error: null,
});
if (row) rows.push(row);
}
}));
const rows = await updateMessagesAIAnalysisBulk(updates);
return { ok: true, conversationKey, rows };
} catch (error) {
@@ -112,7 +120,3 @@ async function processAnalysisRequest({
return { ok: false, conversationKey, rows, error: errorMessage };
}
}
parentPort?.on("message", async (request: AnalysisWorkerRequest) => {
parentPort?.postMessage(await processAnalysisRequest(request));
});
src/moderation/aiAnalyzer.ts
+46 -44
View File
@@ -1,6 +1,6 @@
import { existsSync } from "node:fs";
import { fileURLToPath } from "node:url";
import { Worker } from "node:worker_threads";
import { Piscina } from "piscina";
import { config } from "../config.js";
import { createChildLogger } from "../logger.js";
import {
@@ -36,6 +36,32 @@ const AI_PROCESSING_OVERLAP_MS = 30000;
let activeRequests = 0;
let lastError: string | null = null;
// Global circuit breaker state
let consecutiveErrors = 0;
const MAX_CONSECUTIVE_ERRORS = 5;
let globalCooldownUntil = 0;
function getAnalysisWorkerUrl(): URL {
const candidates = [
new URL("./aiAnalysisWorker.js", import.meta.url),
new URL("../aiAnalysisWorker.js", import.meta.url),
new URL("./aiAnalysisWorker.ts", import.meta.url),
];
for (const candidate of candidates) {
if (existsSync(fileURLToPath(candidate))) {
return candidate;
}
}
return candidates[2];
}
const workerPool = new Piscina({
filename: fileURLToPath(getAnalysisWorkerUrl()),
execArgv: process.execArgv,
});
interface AnalysisWorkerResponse {
ok: boolean;
conversationKey: string;
@@ -62,9 +88,9 @@ export function pickBatchWithinBudget(
let usedTokens = 0;
for (const msg of messages) {
// Estimate tokens based on actual content length
// Estimate tokens based on actual content length (conservative: 3 chars/token)
const content = msg.edited_content ?? msg.content;
const contentTokens = Math.ceil(content.length / 4);
const contentTokens = Math.ceil(content.length / 3);
const msgTokens = contentTokens + tokensPerMessage;
if (usedTokens + msgTokens <= maxTokens) {
@@ -91,19 +117,28 @@ async function processBatch(
messages: MessageRecord[],
): Promise<void> {
if (messages.length === 0) return;
if (Date.now() < globalCooldownUntil) {
return; // Circuit breaker is open
}
activeRequests++;
let shouldScheduleNext = false;
const processingStartedAt = Date.now();
conversationProcessing.set(conversationKey, processingStartedAt);
try {
const result = await runAnalysisInWorker(conversationKey, messages);
const result = (await workerPool.run({ conversationKey, messages })) as AnalysisWorkerResponse;
for (const row of result.rows) {
getModerationBroadcaster()?.messageAnalyzed(row);
}
if (!result.ok) {
consecutiveErrors++;
if (consecutiveErrors >= MAX_CONSECUTIVE_ERRORS) {
globalCooldownUntil = Date.now() + 60000;
logger.warn("Global circuit breaker triggered due to consecutive errors");
}
lastError = result.error ?? "Analysis worker failed";
conversationErrorCooldown.set(
conversationKey,
@@ -125,9 +160,16 @@ async function processBatch(
return;
}
consecutiveErrors = 0; // Reset circuit breaker
conversationErrorCooldown.delete(conversationKey);
shouldScheduleNext = true;
} catch (error) {
consecutiveErrors++;
if (consecutiveErrors >= MAX_CONSECUTIVE_ERRORS) {
globalCooldownUntil = Date.now() + 60000;
logger.warn("Global circuit breaker triggered due to consecutive errors");
}
lastError = error instanceof Error ? error.message : String(error);
const errorStack = error instanceof Error ? error.stack : undefined;
conversationErrorCooldown.set(
@@ -159,46 +201,6 @@ async function processBatch(
}
}
function getAnalysisWorkerUrl(): URL {
const candidates = [
new URL("./aiAnalysisWorker.js", import.meta.url),
new URL("../aiAnalysisWorker.js", import.meta.url),
new URL("./aiAnalysisWorker.ts", import.meta.url),
];
for (const candidate of candidates) {
if (existsSync(fileURLToPath(candidate))) {
return candidate;
}
}
return candidates[2];
}
async function runAnalysisInWorker(
conversationKey: string,
messages: MessageRecord[],
): Promise<AnalysisWorkerResponse> {
return new Promise((resolve, reject) => {
const worker = new Worker(getAnalysisWorkerUrl(), {
execArgv: process.execArgv,
});
worker.once("message", (response: AnalysisWorkerResponse) => {
worker.terminate().catch((error) => {
logger.warn({ error }, "Failed to terminate analysis worker");
});
resolve(response);
});
worker.once("error", reject);
worker.once("exit", (code) => {
if (code !== 0) {
reject(new Error(`Analysis worker exited with code ${code}`));
}
});
worker.postMessage({ conversationKey, messages });
});
}
/**
* Debounced analysis trigger for a conversation
src/moderation/llmModerationClient.ts
+50 -206
View File
@@ -130,199 +130,49 @@ export function extractJson(content: string): any {
throw new Error("No JSON object found in response");
}
/**
* Parses LLM moderation response and validates against target IDs.
* Extracts JSON from surrounding text, validates structure, and transforms to AnalysisResult[].
* Scans from first '{' and attempts JSON.parse at each candidate closing brace.
*/
function salvageMalformedModerationResponse(
content: string,
targetIds: string[],
): AnalysisResult[] | null {
const idMatches = content.match(/\d{10,22}/g) ?? [];
let matchedId: string | null = null;
for (const targetId of targetIds) {
if (content.includes(targetId)) {
matchedId = targetId;
break;
}
}
if (!matchedId) {
for (const candidate of idMatches) {
matchedId =
targetIds.find(
(targetId) =>
targetId.startsWith(candidate) || candidate.startsWith(targetId),
) ?? null;
if (matchedId) break;
}
}
if (!matchedId) return null;
const statusMatch = content.match(/"status"\s*:\s*"(clean|warn|flagged)"/);
const scoreMatch = content.match(/"score"\s*:\s*(\d+(?:\.\d+)?)/);
const analysisMatch = content.match(/"analysis"\s*:\s*"([^"]*)"/);
return [
{
messageId: matchedId,
status: (statusMatch?.[1] as "clean" | "warn" | "flagged") ?? "clean",
flags: [],
score: scoreMatch ? Math.max(0, Math.min(1, Number(scoreMatch[1]))) : 0,
analysis:
analysisMatch?.[1] ?? "Recovered from malformed moderation response",
},
];
}
export function parseModerationResponse(
content: string,
targetIds: string[],
): AnalysisResult[] {
// Extract and parse JSON object
let parsed = extractJson(content);
let parsed: any;
try {
parsed = JSON.parse(content);
} catch (e) {
parsed = extractJson(content);
}
// If parsed is a direct array, wrap it in a results object to handle LLM variations
if (Array.isArray(parsed)) {
parsed = { results: parsed };
} else if (parsed && typeof parsed === "object" && !("results" in parsed)) {
// Handle single result object (has message_id or status)
if ("message_id" in parsed || "status" in parsed) {
const msgId = (parsed as any).message_id || (parsed as any).id;
parsed = {
results: [
{
message_id: msgId,
status: (parsed as any).status || "clean",
flags: (parsed as any).flags || [],
score:
(parsed as any).score !== undefined ? (parsed as any).score : 0.1,
analysis: (parsed as any).analysis || "",
},
],
};
} else {
// Look for any array property (result, data, messages, moderation, etc.)
const arrayKey = Object.keys(parsed).find((key) =>
Array.isArray((parsed as any)[key]),
);
if (arrayKey) {
parsed.results = (parsed as any)[arrayKey];
}
} else {
parsed = { results: [parsed] };
}
}
// Validate structure
if (!parsed || typeof parsed !== "object" || !("results" in parsed)) {
if (!parsed || typeof parsed !== "object" || !Array.isArray(parsed.results)) {
throw new Error("Response missing 'results' array");
}
const response = parsed as RawModerationResponse;
if (!Array.isArray(response.results)) {
throw new Error("'results' must be an array");
}
// Track which target IDs were found
const foundIds = new Set<string>();
const targetIdSet = new Set(targetIds);
// Parse and validate each result
const results: (AnalysisResult | null)[] = response.results.map(
(result, index) => {
const { message_id, status, flags, score, analysis } = result;
// Validate message_id exists and is in target list
if (!message_id) {
throw new Error("Result missing 'message_id'");
}
let finalId = String(message_id).trim();
// Remove wrapping double quotes if any (common in some LLM outputs)
if (finalId.startsWith('"') && finalId.endsWith('"')) {
finalId = finalId.slice(1, -1).trim();
}
if (finalId.startsWith("'") && finalId.endsWith("'")) {
finalId = finalId.slice(1, -1).trim();
}
if (finalId.startsWith("[") && finalId.endsWith("]")) {
finalId = finalId.slice(1, -1).trim();
}
// Advanced Precision Loss & Alignment Fix
if (!targetIdSet.has(finalId)) {
const isSnowflake = (id: string) =>
/^\d{15,22}$/.test(id) || id.includes("e+");
// 1. If there's only one target, map it directly if both are Snowflake-like
if (
targetIds.length === 1 &&
isSnowflake(finalId) &&
isSnowflake(targetIds[0])
) {
log.warn(
{ roundedId: finalId, matchedId: targetIds[0] },
"Mapped single target ID directly to handle precision loss",
);
finalId = targetIds[0];
} else {
// 2. Try matching by long prefix similarity (e.g. 12+ digits)
let cleanLlmId = finalId;
if (finalId.includes("e+")) {
// Convert scientific notation back to string of digits if possible
try {
cleanLlmId = BigInt(Number(finalId)).toString();
} catch (_) {}
}
let bestMatch: string | null = null;
let maxCommonPrefixLen = 0;
for (const targetId of targetIds) {
let commonLen = 0;
const minLen = Math.min(targetId.length, cleanLlmId.length);
for (let i = 0; i < minLen; i++) {
if (targetId[i] === cleanLlmId[i]) {
commonLen++;
} else {
break;
}
}
if (commonLen >= 12 && commonLen > maxCommonPrefixLen) {
maxCommonPrefixLen = commonLen;
bestMatch = targetId;
}
}
if (bestMatch) {
log.warn(
{
roundedId: finalId,
cleanLlmId,
matchedId: bestMatch,
commonLength: maxCommonPrefixLen,
},
"Fixed precision loss in message ID using prefix similarity",
);
finalId = bestMatch;
} else if (
response.results.length === targetIds.length &&
targetIds[index] &&
isSnowflake(finalId) &&
isSnowflake(targetIds[index])
) {
// 3. Fallback: if the number of results matches the number of targets,
// map them 1:1 chronologically (by index) only if they are Snowflake-like
log.warn(
{ roundedId: finalId, index, matchedId: targetIds[index] },
"Aligned message ID using chronological index fallback",
);
finalId = targetIds[index];
}
}
}
const finalId = String(message_id).trim();
if (!targetIdSet.has(finalId)) {
log.warn(
@@ -339,25 +189,15 @@ export function parseModerationResponse(
foundIds.add(finalId);
// Validate status
const validStatuses = ["clean", "warn", "flagged"] as const;
if (!validStatuses.includes(status as (typeof validStatuses)[number])) {
throw new Error(
`Invalid status: ${status}. Must be one of: ${validStatuses.join(", ")}`,
);
}
const safeStatus = validStatuses.includes(status as any) ? status : "clean";
// Validate score: reject null/undefined/non-finite before coercion
if (score === null || score === undefined) {
throw new Error("Invalid score: must not be null or undefined");
}
let numScore = Number(score);
if (!Number.isFinite(numScore)) {
throw new Error(`Invalid score: ${score}. Must be a finite number`);
numScore = 0;
}
numScore = Math.max(0, Math.min(1, numScore));
// Coerce flags to string array
let flagsArray: string[] = [];
if (Array.isArray(flags)) {
flagsArray = flags.map((f) => String(f));
@@ -365,12 +205,11 @@ export function parseModerationResponse(
flagsArray = [String(flags)];
}
// Fallback analysis
const analysisStr = analysis ? String(analysis) : "";
return {
messageId: finalId,
status: status as "clean" | "warn" | "flagged",
status: safeStatus as "clean" | "warn" | "flagged",
flags: flagsArray,
score: numScore,
analysis: analysisStr,
@@ -382,7 +221,6 @@ export function parseModerationResponse(
(r): r is AnalysisResult => r !== null,
);
// Check that all target IDs were found
const missingIds = targetIds.filter((id) => !foundIds.has(id));
if (missingIds.length > 0) {
log.warn(
@@ -556,6 +394,12 @@ export async function runModerationAnalysis(
return;
}
const contentLength = Number(res.headers.get("content-length") || 0);
if (contentLength > 10 * 1024 * 1024) {
log.warn({ attachmentId: att.id, contentLength }, "Attachment too large, skipping");
return;
}
const imageBytes = Buffer.from(await res.arrayBuffer());
const sniffedMime = sniffImageMimeType(imageBytes);
if (!sniffedMime) {
@@ -740,8 +584,36 @@ CRITICAL: "message_id" HARUS berupa STRING (dibungkus tanda kutip ganda). Jangan
],
temperature: 0.2,
top_p: 0.95,
max_tokens: 65536,
response_format: { type: "json_object" },
max_tokens: 16384,
response_format: {
type: "json_schema",
json_schema: {
name: "moderation",
strict: true,
schema: {
type: "object",
properties: {
results: {
type: "array",
items: {
type: "object",
properties: {
message_id: { type: "string" },
status: { type: "string", enum: ["clean", "warn", "flagged"] },
flags: { type: "array", items: { type: "string" } },
score: { type: "number" },
analysis: { type: "string" }
},
required: ["message_id", "status", "flags", "score", "analysis"],
additionalProperties: false
}
}
},
required: ["results"],
additionalProperties: false
}
}
},
stream: false,
chat_template_kwargs: { enable_thinking: false },
reasoning_budget: 0,
@@ -802,34 +674,7 @@ CRITICAL: "message_id" HARUS berupa STRING (dibungkus tanda kutip ganda). Jangan
const errorMsg =
parseError instanceof Error ? parseError.message : String(parseError);
const content: string = lastInvalidContent;
const salvaged = salvageMalformedModerationResponse(content, targetIds);
if (salvaged) {
log.warn(
{
error: errorMsg,
contentLength: content.length,
contentPreview: content.substring(0, 500),
targetIds,
recoveredIds: salvaged.map((result) => result.messageId),
model: config.AI_LLM_MODEL,
timestamp: new Date().toISOString(),
},
"Recovered moderation response from malformed JSON",
);
const recoveredIds = new Set(salvaged.map((result) => result.messageId));
parsed = [
...salvaged,
...targetIds
.filter((id) => !recoveredIds.has(id))
.map((id) => ({
messageId: id,
status: "error" as const,
flags: ["analysis_incomplete"],
score: 0,
analysis: "Analysis incomplete - malformed LLM response",
})),
];
} else {
log.error(
{
error: errorMsg,
@@ -850,7 +695,6 @@ CRITICAL: "message_id" HARUS berupa STRING (dibungkus tanda kutip ganda). Jangan
analysis: `Parsing failed: ${errorMsg}.`,
}));
}
}
log.info(
{
src/moderation/messageStore.ts
+20
View File
@@ -419,6 +419,26 @@ export async function updateMessageAIAnalysis(
}
}
export async function updateMessagesAIAnalysisBulk(
updates: Array<{ messageId: string; result: AIAnalysisUpdate }>
): Promise<MessageRecord[]> {
if (updates.length === 0) return [];
try {
const results = await Promise.all(
updates.map(({ messageId, result }) => updateMessageAIAnalysis(messageId, result))
);
return results.filter((r): r is MessageRecord => r !== null);
} catch (error) {
logger.error(
{
error: error instanceof Error ? error.message : String(error),
},
"Failed to bulk update messages AI analysis",
);
throw error;
}
}
export async function getPendingAIAnalysisMessages(
limit: number = 25,
): Promise<MessageRecord[]> {