fix: implement model-specific rate limiting for gemini-embedding-001

- Add Gemini-specific constants for gemini-embedding-001:
  - Reduced batch token limit (20,000 vs 100,000)
  - Longer retry delays (2000ms vs 500ms)
  - Smaller batch size limit (10 items)
- Update OpenAICompatibleEmbedder to accept configurable rate limiting parameters
- Add inter-batch delays for models with stricter rate limits
- Update GeminiEmbedder to use model-specific configuration
- Fix test expectations to match new constructor signature

Fixes #5713: gemini-embedding-001 quota limit issues during indexing

Author: roomote

src/services/code-index/constants/index.ts

@@ -28,3 +28,6 @@ export const BATCH_PROCESSING_CONCURRENCY = 10
 
 /**Gemini Embedder */
 export const GEMINI_MAX_ITEM_TOKENS = 2048
+export const GEMINI_EMBEDDING_001_MAX_BATCH_TOKENS = 20000 // Reduced batch size for gemini-embedding-001
+export const GEMINI_EMBEDDING_001_RETRY_DELAY_MS = 2000 // Longer delay for gemini-embedding-001
+export const GEMINI_EMBEDDING_001_MAX_BATCH_SIZE = 10 // Smaller batch size for gemini-embedding-001

src/services/code-index/embedders/__tests__/gemini.spec.ts

@@ -38,6 +38,9 @@ describe("GeminiEmbedder", () => {
 				apiKey,
 				"gemini-embedding-001",
 				2048,
+				20000, // GEMINI_EMBEDDING_001_MAX_BATCH_TOKENS
+				2000, // GEMINI_EMBEDDING_001_RETRY_DELAY_MS
+				10, // GEMINI_EMBEDDING_001_MAX_BATCH_SIZE
 			)
 		})
 
@@ -55,6 +58,9 @@ describe("GeminiEmbedder", () => {
 				apiKey,
 				"text-embedding-004",
 				2048,
+				100000, // MAX_BATCH_TOKENS (default for text-embedding-004)
+				500, // INITIAL_RETRY_DELAY_MS (default for text-embedding-004)
+				undefined, // maxBatchSize (undefined for text-embedding-004)
 			)
 		})
 

src/services/code-index/embedders/gemini.ts

@@ -1,6 +1,13 @@
 import { OpenAICompatibleEmbedder } from "./openai-compatible"
 import { IEmbedder, EmbeddingResponse, EmbedderInfo } from "../interfaces/embedder"
-import { GEMINI_MAX_ITEM_TOKENS } from "../constants"
+import {
+	GEMINI_MAX_ITEM_TOKENS,
+	GEMINI_EMBEDDING_001_MAX_BATCH_TOKENS,
+	GEMINI_EMBEDDING_001_RETRY_DELAY_MS,
+	GEMINI_EMBEDDING_001_MAX_BATCH_SIZE,
+	MAX_BATCH_TOKENS,
+	INITIAL_RETRY_DELAY_MS,
+} from "../constants"
 import { t } from "../../../i18n"
 import { TelemetryEventName } from "@roo-code/types"
 import { TelemetryService } from "@roo-code/telemetry"
@@ -32,12 +39,21 @@ export class GeminiEmbedder implements IEmbedder {
 		// Use provided model or default
 		this.modelId = modelId || GeminiEmbedder.DEFAULT_MODEL
 
+		// Get model-specific configuration for gemini-embedding-001
+		const isGeminiEmbedding001 = this.modelId === "gemini-embedding-001"
+		const maxBatchTokens = isGeminiEmbedding001 ? GEMINI_EMBEDDING_001_MAX_BATCH_TOKENS : MAX_BATCH_TOKENS
+		const retryDelayMs = isGeminiEmbedding001 ? GEMINI_EMBEDDING_001_RETRY_DELAY_MS : INITIAL_RETRY_DELAY_MS
+		const maxBatchSize = isGeminiEmbedding001 ? GEMINI_EMBEDDING_001_MAX_BATCH_SIZE : undefined
+
 		// Create an OpenAI Compatible embedder with Gemini's configuration
 		this.openAICompatibleEmbedder = new OpenAICompatibleEmbedder(
 			GeminiEmbedder.GEMINI_BASE_URL,
 			apiKey,
 			this.modelId,
 			GEMINI_MAX_ITEM_TOKENS,
+			maxBatchTokens,
+			retryDelayMs,
+			maxBatchSize,
 		)
 	}
 

src/services/code-index/embedders/openai-compatible.ts

@@ -37,15 +37,29 @@ export class OpenAICompatibleEmbedder implements IEmbedder {
 	private readonly apiKey: string
 	private readonly isFullUrl: boolean
 	private readonly maxItemTokens: number
+	private readonly maxBatchTokens: number
+	private readonly retryDelayMs: number
+	private readonly maxBatchSize?: number
 
 	/**
 	 * Creates a new OpenAI Compatible embedder
 	 * @param baseUrl The base URL for the OpenAI-compatible API endpoint
 	 * @param apiKey The API key for authentication
 	 * @param modelId Optional model identifier (defaults to "text-embedding-3-small")
 	 * @param maxItemTokens Optional maximum tokens per item (defaults to MAX_ITEM_TOKENS)
+	 * @param maxBatchTokens Optional maximum tokens per batch (defaults to MAX_BATCH_TOKENS)
+	 * @param retryDelayMs Optional initial retry delay in milliseconds (defaults to INITIAL_DELAY_MS)
+	 * @param maxBatchSize Optional maximum number of items per batch
 	 */
-	constructor(baseUrl: string, apiKey: string, modelId?: string, maxItemTokens?: number) {
+	constructor(
+		baseUrl: string,
+		apiKey: string,
+		modelId?: string,
+		maxItemTokens?: number,
+		maxBatchTokens?: number,
+		retryDelayMs?: number,
+		maxBatchSize?: number,
+	) {
 		if (!baseUrl) {
 			throw new Error(t("embeddings:validation.baseUrlRequired"))
 		}
@@ -63,6 +77,9 @@ export class OpenAICompatibleEmbedder implements IEmbedder {
 		// Cache the URL type check for performance
 		this.isFullUrl = this.isFullEndpointUrl(baseUrl)
 		this.maxItemTokens = maxItemTokens || MAX_ITEM_TOKENS
+		this.maxBatchTokens = maxBatchTokens || MAX_BATCH_TOKENS
+		this.retryDelayMs = retryDelayMs || INITIAL_DELAY_MS
+		this.maxBatchSize = maxBatchSize
 	}
 
 	/**
@@ -124,7 +141,10 @@ export class OpenAICompatibleEmbedder implements IEmbedder {
 					continue
 				}
 
-				if (currentBatchTokens + itemTokens <= MAX_BATCH_TOKENS) {
+				if (
+					currentBatchTokens + itemTokens <= this.maxBatchTokens &&
+					(!this.maxBatchSize || currentBatch.length < this.maxBatchSize)
+				) {
 					currentBatch.push(text)
 					currentBatchTokens += itemTokens
 					processedIndices.push(i)
@@ -143,6 +163,12 @@ export class OpenAICompatibleEmbedder implements IEmbedder {
 				allEmbeddings.push(...batchResult.embeddings)
 				usage.promptTokens += batchResult.usage.promptTokens
 				usage.totalTokens += batchResult.usage.totalTokens
+
+				// Add delay between batches if there are more batches to process
+				// This helps with rate limiting, especially for gemini-embedding-001
+				if (remainingTexts.length > 0 && this.retryDelayMs > INITIAL_DELAY_MS) {
+					await new Promise((resolve) => setTimeout(resolve, this.retryDelayMs / 4))
+				}
 			}
 		}
 
@@ -299,7 +325,7 @@ export class OpenAICompatibleEmbedder implements IEmbedder {
 				// Check if it's a rate limit error
 				const httpError = error as HttpError
 				if (httpError?.status === 429 && hasMoreAttempts) {
-					const delayMs = INITIAL_DELAY_MS * Math.pow(2, attempts)
+					const delayMs = this.retryDelayMs * Math.pow(2, attempts)
 					console.warn(
 						t("embeddings:rateLimitRetry", {
 							delayMs,