Skip to content

An AI-powered search engine with a generative UI

fix(context-window): add gemini-3 models to the context-window table (#884)

Yoshiki Miura783b3335c0aa
diff --git a/lib/utils/__tests__/context-window.test.ts b/lib/utils/__tests__/context-window.test.ts
index 94ab8885..4e4abe5c 100644
--- a/lib/utils/__tests__/context-window.test.ts
+++ b/lib/utils/__tests__/context-window.test.ts
@@ -48,6 +48,14 @@ describe('context-window', () => {
const maxTokens = getMaxAllowedTokens(mockModel)
expect(maxTokens).toBeGreaterThanOrEqual(1000)
})
+
+ test('uses the real ~1M window for production Gemini models', () => {
+ // (1048576 - 65536) - floor(1048576 * 0.1) = 983040 - 104857 = 878183
+ for (const id of ['gemini-3-flash-preview', 'gemini-3.1-flash-lite']) {
+ const maxTokens = getMaxAllowedTokens({ ...mockModel, id })
+ expect(maxTokens).toBe(878183)
+ }
+ })
})
describe('shouldTruncateMessages', () => {
diff --git a/lib/utils/context-window.ts b/lib/utils/context-window.ts
index fb794c31..5b4b4add 100644
--- a/lib/utils/context-window.ts
+++ b/lib/utils/context-window.ts
@@ -24,6 +24,8 @@ const MODEL_CONTEXT_WINDOWS: Record<string, ModelContextInfo> = {
'claude-3-5-haiku-20241022': { contextWindow: 200000, outputTokens: 8192 },
// Google Models
+ 'gemini-3-flash-preview': { contextWindow: 1048576, outputTokens: 65536 },
+ 'gemini-3.1-flash-lite': { contextWindow: 1048576, outputTokens: 65536 },
'gemini-2.5-flash': { contextWindow: 1048576, outputTokens: 65536 },
'gemini-2.5-pro': { contextWindow: 1048576, outputTokens: 65536 },
@@ -55,7 +57,9 @@ const MODEL_TO_ENCODING: Record<string, TiktokenEncoding> = {
'claude-3-7-sonnet': 'cl100k_base',
'claude-3-7-sonnet-20250219': 'cl100k_base',
'claude-3-5-haiku-20241022': 'cl100k_base',
- 'gemini-2.5-flash': 'cl100k_base', // Use GPT-4 tokenizer as approximation for Gemini
+ 'gemini-3-flash-preview': 'cl100k_base', // Use GPT-4 tokenizer as approximation for Gemini
+ 'gemini-3.1-flash-lite': 'cl100k_base',
+ 'gemini-2.5-flash': 'cl100k_base',
'gemini-2.5-pro': 'cl100k_base',
'grok-4-0709': 'cl100k_base', // Use GPT-4 tokenizer as approximation for Grok
'grok-3': 'cl100k_base',