apps/api/src/scripts/calibrate-whisper.ts

import 'dotenv/config';
import fs from 'fs';
import path from 'path';
import { execSync } from 'child_process';
import { aiService } from '../services/ai';
import { normalizeWolof } from './normalizeWolof';
import { calculateWER, formatError } from '../utils/metrics';

const DATA_DIR = path.join(__dirname, '../../data');
const STATS_PATH = path.join(DATA_DIR, 'calibration_stats.json');
const HF_SAMPLES_PATH = path.join(DATA_DIR, 'hf_samples.json');
const PY_SCRIPT = path.join(__dirname, 'fetch_hf_audio.py');


export async function runCalibration() {
    console.log("🚀 Starting Whisper Confidence Calibration Stress-Test...");

    // Ensure data dir exists
    if (!fs.existsSync(DATA_DIR)) {
        fs.mkdirSync(DATA_DIR, { recursive: true });
    }

    // Run Python fetching script if samples not already present
    if (!fs.existsSync(HF_SAMPLES_PATH)) {
        console.log("📥 Calling Python datasets library to download Hugging Face audio...");
        try {
            execSync(`python3 ${PY_SCRIPT} --output ${DATA_DIR}`, { stdio: 'inherit' });
        } catch (e) {
            console.error("❌ Python script failed to fetch samples. Please check HF_TOKEN or network.");
        }
    } else {
        console.log("♻️ Using cached Hugging Face samples...");
    }

    if (!fs.existsSync(HF_SAMPLES_PATH)) {
        console.error("❌ No samples mapped. Exiting calibration.");
        return;
    }

    const samples = JSON.parse(fs.readFileSync(HF_SAMPLES_PATH, 'utf-8'));
    console.log(`\n🎧 Processing ${samples.length} samples through Whisper STT...`);

    const results = [];
    let redCount = 0;
    let orangeCount = 0;
    let greenCount = 0;
    let totalConfidence = 0;
    let totalRawWER = 0;
    let totalNormWER = 0;

    let totalProcessed = 0;

    for (let i = 0; i < samples.length; i++) {
        const sample = samples[i];
        console.log(`[${sample.source} ${i + 1}/${samples.length}] Transcribing...`);

        try {
            const audioBuffer = Buffer.from(sample.audio_base64, 'base64');
            // Pass to Whisper
            const { text, confidence } = await aiService.transcribeAudio(audioBuffer, `sample_${i}.wav`, 'WOLOF');

            // Normalize
            const normResult = normalizeWolof(text);

            // Compute WER
            const rawWER = calculateWER(sample.original_text, text);
            const normWER = calculateWER(sample.original_text, normResult.normalizedText);

            totalConfidence += confidence;
            totalRawWER += rawWER;
            totalNormWER += normWER;
            totalProcessed++;

            if (confidence <= 50) redCount++;
            else if (confidence <= 80) orangeCount++;
            else greenCount++;

            results.push({
                source: sample.source,
                index: i,
                hfOriginalText: sample.original_text,
                transcribedText: text,
                normalizedText: normResult.normalizedText,
                confidenceScore: confidence,
                rawWER,
                normalizedWER: normWER,
                status: confidence <= 50 ? 'RED' : confidence <= 80 ? 'ORANGE' : 'GREEN'
            });

        } catch (err: unknown) {
            console.error(`Error processing sample ${i} from ${sample.source}: ${formatError(err)}`);
        }
    }

    const averageConfidence = totalProcessed > 0 ? Math.round(totalConfidence / totalProcessed) : 0;
    const averageRawWER = totalProcessed > 0 ? totalRawWER / totalProcessed : 0;
    const averageNormalizedWER = totalProcessed > 0 ? totalNormWER / totalProcessed : 0;

    // Calculate Dictionary Efficiency (Improvement in WER relative to Raw WER)
    let dictionaryEfficiency = 0;
    if (averageRawWER > 0) {
        // If WER goes down, efficiency is positive.
        dictionaryEfficiency = ((averageRawWER - averageNormalizedWER) / averageRawWER) * 100;
    }

    const stats = {
        totalProcessed,
        averageConfidence,
        averageRawWER,
        averageNormalizedWER,
        dictionaryEfficiency,
        distribution: {
            red: { count: redCount, percentage: totalProcessed ? Math.round((redCount / totalProcessed) * 100) : 0 },
            orange: { count: orangeCount, percentage: totalProcessed ? Math.round((orangeCount / totalProcessed) * 100) : 0 },
            green: { count: greenCount, percentage: totalProcessed ? Math.round((greenCount / totalProcessed) * 100) : 0 }
        },
        samples: results,
        updatedAt: new Date().toISOString()
    };

    fs.writeFileSync(STATS_PATH, JSON.stringify(stats, null, 2));
    console.log(`\n✅ Calibration finished! Stats saved to ${STATS_PATH}`);
    console.log(`Average Confidence: ${averageConfidence}%`);
    console.log(`Raw WER: ${(averageRawWER * 100).toFixed(2)}% | Normalized WER: ${(averageNormalizedWER * 100).toFixed(2)}%`);
    console.log(`Dictionary Efficiency Gain: ${dictionaryEfficiency.toFixed(2)}%`);
    console.log(`Red (<=50%): ${redCount} | Orange (51-80%): ${orangeCount} | Green (>80%): ${greenCount}`);
}

// Allow running directly from command line
runCalibration().then(() => process.exit(0)).catch(err => {
    console.error(err);
    process.exit(1);
});