data-access

This document explains how to access transcript data from the k3s-deployed YouTube Transcript project.

Architecture

The YouTube Transcript project runs in the k3s cluster:

• Namespace: youtube-transcript
• Worker Pod: youtube-transcript-worker - GPU-powered transcription using Whisper
• Database: SQLite with Prisma ORM
• Replication: Litestream → MinIO (S3-compatible storage)

Prerequisites

• Access to the ~/Projects/uptownhr/agents directory
• Litestream installed (~/bin/litestream)
• MinIO credentials (minio/minio123)

Database Location

The live database runs inside the worker pod:

/app/prisma/youtube-transcripts.db

Litestream continuously replicates to MinIO:

s3://youtube-transcripts/db/

Restoring the Database Locally

1. Navigate to the project directory

cd ~/Projects/uptownhr/agents/packages/youtube-transcript

2. Use the litestream restore config

The litestream-restore.yml config is already set up:

dbs:
  - path: /tmp/youtube-transcripts.db
    replicas:
      - type: s3
        bucket: youtube-transcripts
        path: db
        endpoint: http://minio.local
        force-path-style: true

3. Run the restore

AWS_ACCESS_KEY_ID=minio AWS_SECRET_ACCESS_KEY=minio123 \
  ~/bin/litestream restore -config litestream-restore.yml /tmp/youtube-transcripts.db

This downloads the latest database snapshot to /tmp/youtube-transcripts.db.

Database Schema

The SQLite database uses Prisma with two main tables:

videos table

transcripts table

Querying the Database

Using Python (recommended):

import sqlite3

conn = sqlite3.connect('/tmp/youtube-transcripts.db')
cursor = conn.cursor()

# Count videos with transcripts
cursor.execute('''
    SELECT COUNT(*) FROM videos v
    JOIN transcripts t ON v.id = t.videoId
    WHERE t.correctedText IS NOT NULL
''')
print(f"Videos with transcripts: {cursor.fetchone()[0]}")

# Sample query
cursor.execute('''
    SELECT v.title, v.channelName, LENGTH(t.correctedText) as length
    FROM videos v
    JOIN transcripts t ON v.id = t.videoId
    ORDER BY t.createdAt DESC
    LIMIT 5
''')
for row in cursor.fetchall():
    print(f"{row[0]} ({row[1]}) - {row[2]} chars")

conn.close()

Data Statistics

As of December 2025:

• 658 total videos tracked
• 646 transcripts with full corrected text
• 646 takeaways with AI-generated analysis
• Average transcript length: ~15,000 characters

K8s Direct Access (Alternative)

If you need to access the live database directly in the cluster:

cd ~/Projects/uptownhr/agents
KUBECONFIG=.kube/config kubectl -n youtube-transcript exec -it deploy/youtube-transcript-worker -- /bin/sh

# Inside the pod
sqlite3 /app/prisma/youtube-transcripts.db ".tables"

Note: Prefer litestream restore for local work to avoid impacting the running service.