import { z } from 'zod';
import Groq from 'groq-sdk';
import { Logger } from '../utils/logger.js';
import { RateLimiter } from '../utils/rate-limiter.js';
import { zodToJsonSchema } from 'zod-to-json-schema';
const VisionAnalysisSchema = z.object({
image_url: z.string().url('Valid image URL is required'),
prompt: z.string().optional(),
analysis_type: z.enum(['describe', 'ocr', 'technical', 'creative']).optional(),
detail_level: z.enum(['low', 'high']).optional(),
model: z.enum(['llama-4-scout-17b-instruct', 'llama-4-maverick-17b-instruct']).optional(),
max_tokens: z.number().min(1).max(4000).optional(),
json_mode: z.boolean().optional()
});
export class VisionAnalysisTool {
public readonly name = 'groq_vision_analysis';
public readonly description = 'Analyze images using Groq multimodal models';
public readonly inputSchema = zodToJsonSchema(VisionAnalysisSchema, { $refStrategy: 'none', target: 'jsonSchema7' });
constructor(
private groq: Groq,
private logger: Logger,
private rateLimiter: RateLimiter
) {}
async execute(args: z.infer<typeof VisionAnalysisSchema>): Promise<any> {
const validated = VisionAnalysisSchema.parse(args);
const model = validated.model || 'llama-4-scout-17b-instruct';
// Apply rate limiting for vision models
await this.rateLimiter.checkLimit(`vision_${model}`);
const prompt = this.buildPrompt(validated);
try {
this.logger.info('Starting vision analysis', {
model,
analysisType: validated.analysis_type,
imageUrl: validated.image_url.substring(0, 50) + '...'
});
const messages = [
{
role: 'user' as const,
content: [
{
type: 'text',
text: prompt
},
{
type: 'image_url',
image_url: {
url: validated.image_url,
detail: validated.detail_level || 'high'
}
}
]
}
];
const requestParams: any = {
model,
messages,
max_tokens: validated.max_tokens || 1000,
temperature: 0.3 // Lower temperature for more consistent vision analysis
};
if (validated.json_mode) {
requestParams.response_format = { type: 'json_object' };
}
const completion = await this.groq.chat.completions.create(requestParams);
const result = {
analysis: completion.choices[0]?.message?.content || '',
model,
analysis_type: validated.analysis_type,
image_url: validated.image_url,
usage: completion.usage,
timestamp: new Date().toISOString()
};
this.logger.info('Vision analysis completed', {
model,
analysisLength: result.analysis.length
});
return result;
} catch (error) {
const errorMessage = error instanceof Error ? error.message : String(error);
// ✅ CORREÇÃO: Usar 'this.name'
this.logger.error(`Tool ${this.name} failed`, { error: errorMessage });
}
}
private buildPrompt(args: z.infer<typeof VisionAnalysisSchema>): string {
if (args.prompt) return args.prompt;
const prompts = {
describe: 'Describe this image in detail, including objects, people, setting, colors, and overall composition.',
ocr: 'Extract and transcribe all text visible in this image. Organize the text logically and indicate its position/context.',
technical: 'Provide a technical analysis of this image including composition, lighting, quality, and any technical aspects.',
creative: 'Provide a creative interpretation of this image, including mood, artistic elements, and storytelling aspects.'
};
return prompts[args.analysis_type || 'describe'];
}
}
