LLM/BrowserOS
1
0
Fork 0
mirror of https://github.com/browseros-ai/BrowserOS.git synced 2026-05-17 02:25:57 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix: context window size (#171)

Browse Source
This commit is contained in:
shivammittal274
2026-01-07 17:13:00 +05:30
committed by GitHub
parent 9c6bccd46e
commit 2ac655b69e
5 changed files with 23 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
apps/agent/entrypoints/options/ai-settings/NewProviderDialog.tsx
View File

@@ -245,8 +245,10 @@ export const NewProviderDialog: FC<NewProviderDialogProps> = ({
setIsCustomModel(false)
}
// Auto-fill context window when model changes
// Auto-fill context window when model changes (only for new providers)
useEffect(() => {
if (initialValues?.id) return
if (watchedModelId && watchedModelId !== 'custom') {
const contextLength = getModelContextLength(
watchedType as ProviderType,
@@ -256,7 +258,7 @@ export const NewProviderDialog: FC<NewProviderDialogProps> = ({
form.setValue('contextWindow', contextLength)
}
}
}, [watchedModelId, watchedType, form])
}, [watchedModelId, watchedType, form, initialValues?.id])
// Handle model selection (including custom option)
const handleModelChange = (value: string) => {

2
apps/agent/entrypoints/sidepanel/index/useChatSession.ts
View File

@@ -261,6 +261,8 @@ export const useChatSession = () => {
conversationId: conversationIdRef.current,
model: provider?.modelId ?? 'default',
mode: currentMode,
contextWindowSize: provider?.contextWindow,
temperature: provider?.temperature,
// Azure-specific
resourceName: provider?.resourceName,
// Bedrock-specific

2
apps/agent/lib/schedules/getChatServerResponse.ts
View File

@@ -78,6 +78,8 @@ export async function getChatServerResponse(
conversationId,
model: provider?.modelId ?? 'default',
mode: request.mode ?? 'agent',
contextWindowSize: provider?.contextWindow,
temperature: provider?.temperature,
resourceName: provider?.resourceName,
accessKeyId: provider?.accessKeyId,
secretAccessKey: provider?.secretAccessKey,

16
apps/server/src/agent/agent/gemini-agent.ts
View File

@@ -49,16 +49,22 @@ export class GeminiAgent {
// Calculate compression threshold based on context window size
const contextWindow =
config.contextWindowSize ?? AGENT_LIMITS.DEFAULT_CONTEXT_WINDOW
// Hybrid compression: ensure minimum headroom while capping ratio for large contexts
const headroomBasedRatio =
(contextWindow - AGENT_LIMITS.COMPRESSION_MIN_HEADROOM) / contextWindow
const compressionRatio = Math.min(
AGENT_LIMITS.COMPRESSION_MAX_RATIO,
Math.max(AGENT_LIMITS.COMPRESSION_MIN_RATIO, headroomBasedRatio),
)
const compressionThreshold =
(AGENT_LIMITS.DEFAULT_COMPRESSION_RATIO * contextWindow) /
AGENT_LIMITS.DEFAULT_CONTEXT_WINDOW
(compressionRatio * contextWindow) / AGENT_LIMITS.DEFAULT_CONTEXT_WINDOW
logger.info('Compression config', {
contextWindow,
compressionRatio,
compressionThreshold,
compressesAtTokens: Math.floor(
AGENT_LIMITS.DEFAULT_COMPRESSION_RATIO * contextWindow,
),
compressesAtTokens: Math.floor(compressionRatio * contextWindow),
})
logger.debug('MCP servers config', {

5
packages/shared/src/constants/limits.ts
View File

@@ -15,7 +15,10 @@ export const RATE_LIMITS = {
export const AGENT_LIMITS = {
MAX_TURNS: 100,
DEFAULT_CONTEXT_WINDOW: 1_000_000,
DEFAULT_COMPRESSION_RATIO: 0.75,
// Compression settings - hybrid approach with minimum headroom
COMPRESSION_MIN_HEADROOM: 10_000, // Always leave at least 10K tokens for tool responses
COMPRESSION_MAX_RATIO: 0.75, // Never wait longer than 75% for large models
COMPRESSION_MIN_RATIO: 0.4, // Never compress too early (before 40%)
} as const
export const PAGINATION = {