@just-every/ensemble/dist/utils/image_to_text.js

LLM provider abstraction layer with unified streaming interface

import { ensembleRequest, findModel } from '../index.js';
const imageDescriptionCache = {};
function generateImageHash(imageData) {
    const sample = imageData.substring(0, 100);
    const length = imageData.length;
    return `${sample}_${length}`;
}
export async function convertImageToText(imageData, modelId) {
    if (!imageData.startsWith('data:image/')) {
        return imageData;
    }
    console.log(`Converting image to text description for model ${modelId}`);
    const imageHash = generateImageHash(imageData);
    if (imageDescriptionCache[imageHash]) {
        console.log(`Using cached image description for ${modelId}`);
        return imageDescriptionCache[imageHash];
    }
    try {
        const stream = ensembleRequest([
            {
                type: 'message',
                role: 'system',
                content: 'Please describe the following image in a couple of sentences. Focus on the main visual elements and key details that someone would need to understand what is shown in the image.',
            },
            {
                type: 'message',
                role: 'user',
                content: imageData,
            },
        ], {
            modelClass: 'vision_mini',
        });
        for await (const event of stream) {
            if (event.type === 'message_complete' && 'content' in event) {
                imageDescriptionCache[imageHash] = event.content.trim();
                return imageDescriptionCache[imageHash];
            }
        }
    }
    catch (error) {
        console.error('Error generating image description:', error);
    }
    return 'Image found, but could not be converted to text';
}
export async function convertImageToTextIfNeeded(imageData, modelId) {
    if (!imageData.startsWith('data:image/')) {
        return false;
    }
    if (modelId && findModel(modelId)?.features?.input_modality?.includes('image')) {
        return false;
    }
    return await convertImageToText(imageData, modelId || 'unknown');
}
//# sourceMappingURL=image_to_text.js.map