Log the 2 types of loop detection (#4193)
This commit is contained in:
parent
734da8b9d2
commit
886faa2990
|
@ -101,7 +101,7 @@ export class GeminiClient {
|
||||||
*/
|
*/
|
||||||
private readonly COMPRESSION_PRESERVE_THRESHOLD = 0.3;
|
private readonly COMPRESSION_PRESERVE_THRESHOLD = 0.3;
|
||||||
|
|
||||||
private readonly loopDetector = new LoopDetectionService();
|
private readonly loopDetector: LoopDetectionService;
|
||||||
private lastPromptId?: string;
|
private lastPromptId?: string;
|
||||||
|
|
||||||
constructor(private config: Config) {
|
constructor(private config: Config) {
|
||||||
|
@ -110,6 +110,7 @@ export class GeminiClient {
|
||||||
}
|
}
|
||||||
|
|
||||||
this.embeddingModel = config.getEmbeddingModel();
|
this.embeddingModel = config.getEmbeddingModel();
|
||||||
|
this.loopDetector = new LoopDetectionService(config);
|
||||||
}
|
}
|
||||||
|
|
||||||
async initialize(contentGeneratorConfig: ContentGeneratorConfig) {
|
async initialize(contentGeneratorConfig: ContentGeneratorConfig) {
|
||||||
|
|
|
@ -4,7 +4,7 @@
|
||||||
* SPDX-License-Identifier: Apache-2.0
|
* SPDX-License-Identifier: Apache-2.0
|
||||||
*/
|
*/
|
||||||
|
|
||||||
import { describe, it, expect, beforeEach } from 'vitest';
|
import { describe, it, expect, beforeEach, vi } from 'vitest';
|
||||||
import { LoopDetectionService } from './loopDetectionService.js';
|
import { LoopDetectionService } from './loopDetectionService.js';
|
||||||
import {
|
import {
|
||||||
GeminiEventType,
|
GeminiEventType,
|
||||||
|
@ -12,15 +12,26 @@ import {
|
||||||
ServerGeminiToolCallRequestEvent,
|
ServerGeminiToolCallRequestEvent,
|
||||||
} from '../core/turn.js';
|
} from '../core/turn.js';
|
||||||
import { ServerGeminiStreamEvent } from '../core/turn.js';
|
import { ServerGeminiStreamEvent } from '../core/turn.js';
|
||||||
|
import { Config } from '../config/config.js';
|
||||||
|
import * as loggers from '../telemetry/loggers.js';
|
||||||
|
|
||||||
|
vi.mock('../telemetry/loggers.js', () => ({
|
||||||
|
logLoopDetected: vi.fn(),
|
||||||
|
}));
|
||||||
|
|
||||||
const TOOL_CALL_LOOP_THRESHOLD = 5;
|
const TOOL_CALL_LOOP_THRESHOLD = 5;
|
||||||
const CONTENT_LOOP_THRESHOLD = 10;
|
const CONTENT_LOOP_THRESHOLD = 10;
|
||||||
|
|
||||||
describe('LoopDetectionService', () => {
|
describe('LoopDetectionService', () => {
|
||||||
let service: LoopDetectionService;
|
let service: LoopDetectionService;
|
||||||
|
let mockConfig: Config;
|
||||||
|
|
||||||
beforeEach(() => {
|
beforeEach(() => {
|
||||||
service = new LoopDetectionService();
|
mockConfig = {
|
||||||
|
getTelemetryEnabled: () => true,
|
||||||
|
} as unknown as Config;
|
||||||
|
service = new LoopDetectionService(mockConfig);
|
||||||
|
vi.clearAllMocks();
|
||||||
});
|
});
|
||||||
|
|
||||||
const createToolCallRequestEvent = (
|
const createToolCallRequestEvent = (
|
||||||
|
@ -48,6 +59,7 @@ describe('LoopDetectionService', () => {
|
||||||
for (let i = 0; i < TOOL_CALL_LOOP_THRESHOLD - 1; i++) {
|
for (let i = 0; i < TOOL_CALL_LOOP_THRESHOLD - 1; i++) {
|
||||||
expect(service.addAndCheck(event)).toBe(false);
|
expect(service.addAndCheck(event)).toBe(false);
|
||||||
}
|
}
|
||||||
|
expect(loggers.logLoopDetected).not.toHaveBeenCalled();
|
||||||
});
|
});
|
||||||
|
|
||||||
it(`should detect a loop on the TOOL_CALL_LOOP_THRESHOLD-th identical call`, () => {
|
it(`should detect a loop on the TOOL_CALL_LOOP_THRESHOLD-th identical call`, () => {
|
||||||
|
@ -56,6 +68,7 @@ describe('LoopDetectionService', () => {
|
||||||
service.addAndCheck(event);
|
service.addAndCheck(event);
|
||||||
}
|
}
|
||||||
expect(service.addAndCheck(event)).toBe(true);
|
expect(service.addAndCheck(event)).toBe(true);
|
||||||
|
expect(loggers.logLoopDetected).toHaveBeenCalledTimes(1);
|
||||||
});
|
});
|
||||||
|
|
||||||
it('should detect a loop on subsequent identical calls', () => {
|
it('should detect a loop on subsequent identical calls', () => {
|
||||||
|
@ -64,6 +77,7 @@ describe('LoopDetectionService', () => {
|
||||||
service.addAndCheck(event);
|
service.addAndCheck(event);
|
||||||
}
|
}
|
||||||
expect(service.addAndCheck(event)).toBe(true);
|
expect(service.addAndCheck(event)).toBe(true);
|
||||||
|
expect(loggers.logLoopDetected).toHaveBeenCalledTimes(2);
|
||||||
});
|
});
|
||||||
|
|
||||||
it('should not detect a loop for different tool calls', () => {
|
it('should not detect a loop for different tool calls', () => {
|
||||||
|
@ -91,6 +105,7 @@ describe('LoopDetectionService', () => {
|
||||||
for (let i = 0; i < CONTENT_LOOP_THRESHOLD - 1; i++) {
|
for (let i = 0; i < CONTENT_LOOP_THRESHOLD - 1; i++) {
|
||||||
expect(service.addAndCheck(event)).toBe(false);
|
expect(service.addAndCheck(event)).toBe(false);
|
||||||
}
|
}
|
||||||
|
expect(loggers.logLoopDetected).not.toHaveBeenCalled();
|
||||||
});
|
});
|
||||||
|
|
||||||
it(`should detect a loop on the CONTENT_LOOP_THRESHOLD-th identical content string`, () => {
|
it(`should detect a loop on the CONTENT_LOOP_THRESHOLD-th identical content string`, () => {
|
||||||
|
@ -99,6 +114,7 @@ describe('LoopDetectionService', () => {
|
||||||
service.addAndCheck(event);
|
service.addAndCheck(event);
|
||||||
}
|
}
|
||||||
expect(service.addAndCheck(event)).toBe(true);
|
expect(service.addAndCheck(event)).toBe(true);
|
||||||
|
expect(loggers.logLoopDetected).toHaveBeenCalledTimes(1);
|
||||||
});
|
});
|
||||||
|
|
||||||
it('should not detect a loop for different content strings', () => {
|
it('should not detect a loop for different content strings', () => {
|
||||||
|
@ -108,6 +124,7 @@ describe('LoopDetectionService', () => {
|
||||||
expect(service.addAndCheck(event1)).toBe(false);
|
expect(service.addAndCheck(event1)).toBe(false);
|
||||||
expect(service.addAndCheck(event2)).toBe(false);
|
expect(service.addAndCheck(event2)).toBe(false);
|
||||||
}
|
}
|
||||||
|
expect(loggers.logLoopDetected).not.toHaveBeenCalled();
|
||||||
});
|
});
|
||||||
});
|
});
|
||||||
|
|
||||||
|
|
|
@ -6,6 +6,9 @@
|
||||||
|
|
||||||
import { createHash } from 'crypto';
|
import { createHash } from 'crypto';
|
||||||
import { GeminiEventType, ServerGeminiStreamEvent } from '../core/turn.js';
|
import { GeminiEventType, ServerGeminiStreamEvent } from '../core/turn.js';
|
||||||
|
import { logLoopDetected } from '../telemetry/loggers.js';
|
||||||
|
import { LoopDetectedEvent, LoopType } from '../telemetry/types.js';
|
||||||
|
import { Config } from '../config/config.js';
|
||||||
|
|
||||||
const TOOL_CALL_LOOP_THRESHOLD = 5;
|
const TOOL_CALL_LOOP_THRESHOLD = 5;
|
||||||
const CONTENT_LOOP_THRESHOLD = 10;
|
const CONTENT_LOOP_THRESHOLD = 10;
|
||||||
|
@ -24,6 +27,11 @@ export class LoopDetectionService {
|
||||||
private lastRepeatedSentence: string = '';
|
private lastRepeatedSentence: string = '';
|
||||||
private sentenceRepetitionCount: number = 0;
|
private sentenceRepetitionCount: number = 0;
|
||||||
private partialContent: string = '';
|
private partialContent: string = '';
|
||||||
|
private config: Config;
|
||||||
|
|
||||||
|
constructor(config: Config) {
|
||||||
|
this.config = config;
|
||||||
|
}
|
||||||
|
|
||||||
private getToolCallKey(toolCall: { name: string; args: object }): string {
|
private getToolCallKey(toolCall: { name: string; args: object }): string {
|
||||||
const argsString = JSON.stringify(toolCall.args);
|
const argsString = JSON.stringify(toolCall.args);
|
||||||
|
@ -59,7 +67,14 @@ export class LoopDetectionService {
|
||||||
this.lastToolCallKey = key;
|
this.lastToolCallKey = key;
|
||||||
this.toolCallRepetitionCount = 1;
|
this.toolCallRepetitionCount = 1;
|
||||||
}
|
}
|
||||||
return this.toolCallRepetitionCount >= TOOL_CALL_LOOP_THRESHOLD;
|
if (this.toolCallRepetitionCount >= TOOL_CALL_LOOP_THRESHOLD) {
|
||||||
|
logLoopDetected(
|
||||||
|
this.config,
|
||||||
|
new LoopDetectedEvent(LoopType.CONSECUTIVE_IDENTICAL_TOOL_CALLS),
|
||||||
|
);
|
||||||
|
return true;
|
||||||
|
}
|
||||||
|
return false;
|
||||||
}
|
}
|
||||||
|
|
||||||
private checkContentLoop(content: string): boolean {
|
private checkContentLoop(content: string): boolean {
|
||||||
|
@ -94,6 +109,10 @@ export class LoopDetectionService {
|
||||||
}
|
}
|
||||||
|
|
||||||
if (this.sentenceRepetitionCount >= CONTENT_LOOP_THRESHOLD) {
|
if (this.sentenceRepetitionCount >= CONTENT_LOOP_THRESHOLD) {
|
||||||
|
logLoopDetected(
|
||||||
|
this.config,
|
||||||
|
new LoopDetectedEvent(LoopType.CHANTING_IDENTICAL_SENTENCES),
|
||||||
|
);
|
||||||
return true;
|
return true;
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
|
|
|
@ -15,6 +15,7 @@ import {
|
||||||
ApiResponseEvent,
|
ApiResponseEvent,
|
||||||
ApiErrorEvent,
|
ApiErrorEvent,
|
||||||
FlashFallbackEvent,
|
FlashFallbackEvent,
|
||||||
|
LoopDetectedEvent,
|
||||||
} from '../types.js';
|
} from '../types.js';
|
||||||
import { EventMetadataKey } from './event-metadata-key.js';
|
import { EventMetadataKey } from './event-metadata-key.js';
|
||||||
import { Config } from '../../config/config.js';
|
import { Config } from '../../config/config.js';
|
||||||
|
@ -33,6 +34,7 @@ const api_response_event_name = 'api_response';
|
||||||
const api_error_event_name = 'api_error';
|
const api_error_event_name = 'api_error';
|
||||||
const end_session_event_name = 'end_session';
|
const end_session_event_name = 'end_session';
|
||||||
const flash_fallback_event_name = 'flash_fallback';
|
const flash_fallback_event_name = 'flash_fallback';
|
||||||
|
const loop_detected_event_name = 'loop_detected';
|
||||||
|
|
||||||
export interface LogResponse {
|
export interface LogResponse {
|
||||||
nextRequestWaitMs?: number;
|
nextRequestWaitMs?: number;
|
||||||
|
@ -448,6 +450,18 @@ export class ClearcutLogger {
|
||||||
});
|
});
|
||||||
}
|
}
|
||||||
|
|
||||||
|
logLoopDetectedEvent(event: LoopDetectedEvent): void {
|
||||||
|
const data = [
|
||||||
|
{
|
||||||
|
gemini_cli_key: EventMetadataKey.GEMINI_CLI_LOOP_DETECTED_TYPE,
|
||||||
|
value: JSON.stringify(event.loop_type),
|
||||||
|
},
|
||||||
|
];
|
||||||
|
|
||||||
|
this.enqueueLogEvent(this.createLogEvent(loop_detected_event_name, data));
|
||||||
|
this.flushIfNeeded();
|
||||||
|
}
|
||||||
|
|
||||||
logEndSessionEvent(event: EndSessionEvent): void {
|
logEndSessionEvent(event: EndSessionEvent): void {
|
||||||
const data = [
|
const data = [
|
||||||
{
|
{
|
||||||
|
|
|
@ -150,6 +150,13 @@ export enum EventMetadataKey {
|
||||||
|
|
||||||
// Logs the total number of Google accounts ever used.
|
// Logs the total number of Google accounts ever used.
|
||||||
GEMINI_CLI_GOOGLE_ACCOUNTS_COUNT = 37,
|
GEMINI_CLI_GOOGLE_ACCOUNTS_COUNT = 37,
|
||||||
|
|
||||||
|
// ==========================================================================
|
||||||
|
// Loop Detected Event Keys
|
||||||
|
// ===========================================================================
|
||||||
|
|
||||||
|
// Logs the type of loop detected.
|
||||||
|
GEMINI_CLI_LOOP_DETECTED_TYPE = 38,
|
||||||
}
|
}
|
||||||
|
|
||||||
export function getEventMetadataKey(
|
export function getEventMetadataKey(
|
||||||
|
|
|
@ -25,6 +25,7 @@ import {
|
||||||
ToolCallEvent,
|
ToolCallEvent,
|
||||||
UserPromptEvent,
|
UserPromptEvent,
|
||||||
FlashFallbackEvent,
|
FlashFallbackEvent,
|
||||||
|
LoopDetectedEvent,
|
||||||
} from './types.js';
|
} from './types.js';
|
||||||
import {
|
import {
|
||||||
recordApiErrorMetrics,
|
recordApiErrorMetrics,
|
||||||
|
@ -288,3 +289,23 @@ export function logApiResponse(config: Config, event: ApiResponseEvent): void {
|
||||||
);
|
);
|
||||||
recordTokenUsageMetrics(config, event.model, event.tool_token_count, 'tool');
|
recordTokenUsageMetrics(config, event.model, event.tool_token_count, 'tool');
|
||||||
}
|
}
|
||||||
|
|
||||||
|
export function logLoopDetected(
|
||||||
|
config: Config,
|
||||||
|
event: LoopDetectedEvent,
|
||||||
|
): void {
|
||||||
|
ClearcutLogger.getInstance(config)?.logLoopDetectedEvent(event);
|
||||||
|
if (!isTelemetrySdkInitialized()) return;
|
||||||
|
|
||||||
|
const attributes: LogAttributes = {
|
||||||
|
...getCommonAttributes(config),
|
||||||
|
...event,
|
||||||
|
};
|
||||||
|
|
||||||
|
const logger = logs.getLogger(SERVICE_NAME);
|
||||||
|
const logRecord: LogRecord = {
|
||||||
|
body: `Loop detected. Type: ${event.loop_type}.`,
|
||||||
|
attributes,
|
||||||
|
};
|
||||||
|
logger.emit(logRecord);
|
||||||
|
}
|
||||||
|
|
|
@ -246,6 +246,23 @@ export class FlashFallbackEvent {
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
|
|
||||||
|
export enum LoopType {
|
||||||
|
CONSECUTIVE_IDENTICAL_TOOL_CALLS = 'consecutive_identical_tool_calls',
|
||||||
|
CHANTING_IDENTICAL_SENTENCES = 'chanting_identical_sentences',
|
||||||
|
}
|
||||||
|
|
||||||
|
export class LoopDetectedEvent {
|
||||||
|
'event.name': 'loop_detected';
|
||||||
|
'event.timestamp': string; // ISO 8601
|
||||||
|
loop_type: LoopType;
|
||||||
|
|
||||||
|
constructor(loop_type: LoopType) {
|
||||||
|
this['event.name'] = 'loop_detected';
|
||||||
|
this['event.timestamp'] = new Date().toISOString();
|
||||||
|
this.loop_type = loop_type;
|
||||||
|
}
|
||||||
|
}
|
||||||
|
|
||||||
export type TelemetryEvent =
|
export type TelemetryEvent =
|
||||||
| StartSessionEvent
|
| StartSessionEvent
|
||||||
| EndSessionEvent
|
| EndSessionEvent
|
||||||
|
@ -254,4 +271,5 @@ export type TelemetryEvent =
|
||||||
| ApiRequestEvent
|
| ApiRequestEvent
|
||||||
| ApiErrorEvent
|
| ApiErrorEvent
|
||||||
| ApiResponseEvent
|
| ApiResponseEvent
|
||||||
| FlashFallbackEvent;
|
| FlashFallbackEvent
|
||||||
|
| LoopDetectedEvent;
|
||||||
|
|
Loading…
Reference in New Issue