Files
bolt-diy/app/lib/modules/llm/providers/google.ts
Stijnus b5d9055851 🔧 Fix Token Limits & Invalid JSON Response Errors (#1934)
ISSUES FIXED:
-  Invalid JSON response errors during streaming
-  Incorrect token limits causing API rejections
-  Outdated hardcoded model configurations
-  Poor error messages for API failures

SOLUTIONS IMPLEMENTED:

🎯 ACCURATE TOKEN LIMITS & CONTEXT SIZES
- OpenAI GPT-4o: 128k context (was 8k)
- OpenAI GPT-3.5-turbo: 16k context (was 8k)
- Anthropic Claude 3.5 Sonnet: 200k context (was 8k)
- Anthropic Claude 3 Haiku: 200k context (was 8k)
- Google Gemini 1.5 Pro: 2M context (was 8k)
- Google Gemini 1.5 Flash: 1M context (was 8k)
- Groq Llama models: 128k context (was 8k)
- Together models: Updated with accurate limits

�� DYNAMIC MODEL FETCHING ENHANCED
- Smart context detection from provider APIs
- Automatic fallback to known limits when API unavailable
- Safety caps to prevent token overflow (100k max)
- Intelligent model filtering and deduplication

🛡️ IMPROVED ERROR HANDLING
- Specific error messages for Invalid JSON responses
- Token limit exceeded warnings with solutions
- API key validation with clear guidance
- Rate limiting detection and user guidance
- Network timeout handling

 PERFORMANCE OPTIMIZATIONS
- Reduced static models from 40+ to 12 essential
- Enhanced streaming error detection
- Better API response validation
- Improved context window display (shows M/k units)

🔧 TECHNICAL IMPROVEMENTS
- Dynamic model context detection from APIs
- Enhanced streaming reliability
- Better token limit enforcement
- Comprehensive error categorization
- Smart model validation before API calls

IMPACT:
 Eliminates Invalid JSON response errors
 Prevents token limit API rejections
 Provides accurate model capabilities
 Improves user experience with clear errors
 Enables full utilization of modern LLM context windows
2025-08-29 20:53:57 +02:00

128 lines
4.4 KiB
TypeScript

import { BaseProvider } from '~/lib/modules/llm/base-provider';
import type { ModelInfo } from '~/lib/modules/llm/types';
import type { IProviderSetting } from '~/types/model';
import type { LanguageModelV1 } from 'ai';
import { createGoogleGenerativeAI } from '@ai-sdk/google';
export default class GoogleProvider extends BaseProvider {
name = 'Google';
getApiKeyLink = 'https://aistudio.google.com/app/apikey';
config = {
apiTokenKey: 'GOOGLE_GENERATIVE_AI_API_KEY',
};
staticModels: ModelInfo[] = [
/*
* Essential fallback models - only the most reliable/stable ones
* Gemini 1.5 Pro: 2M context, excellent for complex reasoning and large codebases
*/
{ name: 'gemini-1.5-pro', label: 'Gemini 1.5 Pro', provider: 'Google', maxTokenAllowed: 2000000 },
// Gemini 1.5 Flash: 1M context, fast and cost-effective
{ name: 'gemini-1.5-flash', label: 'Gemini 1.5 Flash', provider: 'Google', maxTokenAllowed: 1000000 },
];
async getDynamicModels(
apiKeys?: Record<string, string>,
settings?: IProviderSetting,
serverEnv?: Record<string, string>,
): Promise<ModelInfo[]> {
const { apiKey } = this.getProviderBaseUrlAndKey({
apiKeys,
providerSettings: settings,
serverEnv: serverEnv as any,
defaultBaseUrlKey: '',
defaultApiTokenKey: 'GOOGLE_GENERATIVE_AI_API_KEY',
});
if (!apiKey) {
throw `Missing Api Key configuration for ${this.name} provider`;
}
const response = await fetch(`https://generativelanguage.googleapis.com/v1beta/models?key=${apiKey}`, {
headers: {
['Content-Type']: 'application/json',
},
});
if (!response.ok) {
throw new Error(`Failed to fetch models from Google API: ${response.status} ${response.statusText}`);
}
const res = (await response.json()) as any;
if (!res.models || !Array.isArray(res.models)) {
throw new Error('Invalid response format from Google API');
}
// Filter out models with very low token limits and experimental/unstable models
const data = res.models.filter((model: any) => {
const hasGoodTokenLimit = (model.outputTokenLimit || 0) > 8000;
const isStable = !model.name.includes('exp') || model.name.includes('flash-exp');
return hasGoodTokenLimit && isStable;
});
return data.map((m: any) => {
const modelName = m.name.replace('models/', '');
// Get accurate context window from Google API
let contextWindow = 32000; // default fallback
if (m.inputTokenLimit && m.outputTokenLimit) {
// Use the input limit as the primary context window (typically larger)
contextWindow = m.inputTokenLimit;
} else if (modelName.includes('gemini-1.5-pro')) {
contextWindow = 2000000; // Gemini 1.5 Pro has 2M context
} else if (modelName.includes('gemini-1.5-flash')) {
contextWindow = 1000000; // Gemini 1.5 Flash has 1M context
} else if (modelName.includes('gemini-2.0-flash')) {
contextWindow = 1000000; // Gemini 2.0 Flash has 1M context
} else if (modelName.includes('gemini-pro')) {
contextWindow = 32000; // Gemini Pro has 32k context
} else if (modelName.includes('gemini-flash')) {
contextWindow = 32000; // Gemini Flash has 32k context
}
// Cap at reasonable limits to prevent issues
const maxAllowed = 2000000; // 2M tokens max
const finalContext = Math.min(contextWindow, maxAllowed);
return {
name: modelName,
label: `${m.displayName} (${finalContext >= 1000000 ? Math.floor(finalContext / 1000000) + 'M' : Math.floor(finalContext / 1000) + 'k'} context)`,
provider: this.name,
maxTokenAllowed: finalContext,
};
});
}
getModelInstance(options: {
model: string;
serverEnv: any;
apiKeys?: Record<string, string>;
providerSettings?: Record<string, IProviderSetting>;
}): LanguageModelV1 {
const { model, serverEnv, apiKeys, providerSettings } = options;
const { apiKey } = this.getProviderBaseUrlAndKey({
apiKeys,
providerSettings: providerSettings?.[this.name],
serverEnv: serverEnv as any,
defaultBaseUrlKey: '',
defaultApiTokenKey: 'GOOGLE_GENERATIVE_AI_API_KEY',
});
if (!apiKey) {
throw new Error(`Missing API key for ${this.name} provider`);
}
const google = createGoogleGenerativeAI({
apiKey,
});
return google(model);
}
}