AI-curated, API-validated soundtracks for any scene, book, or feeling.
Give VibeCheck a book title, scene description, or feeling — it generates a cross-era playlist, verifies every track exists on YouTube Music, and gives you a single "Play All" link.
The key distinction from other AI playlist tools: the LLM is treated as a suggestion engine, not a source of truth. Every generated song title is cross-referenced against the YouTube Music API before it reaches the UI. Hallucinated tracks are caught and replaced via fallback queries, not silently shown.
1. Hallucination filter with confidence scoring
LLMs routinely invent plausible-sounding song titles. Instead of trusting the model's output directly, I parse each suggested track, search YouTube Music, and compute a string similarity score (SequenceMatcher) between the returned artist/title and the expected values. Tracks below the confidence threshold are retried with fallback queries, then dropped if all fail.
2. Parallel API validation
The original implementation made 40+ ytmusicapi calls sequentially — one per track. Replacing the loop with ThreadPoolExecutor (10 workers) reduced validation time by ~8–10x for large mixes, with a live progress bar updating as futures resolve.
3. Structured LLM output via system/user message separation
The prompt enforces a strict JSON schema (primary_query, fallback_queries, reason, era per track) using Groq's response_format: json_object. The system message carries the Music Supervisor persona; the user message carries constraints. This separation keeps the model's output consistent across varied inputs.
4. Session feedback loop
Per-track 👍/👎 signals are stored in st.session_state and injected into the next generation prompt as explicit liked/disliked context. The model adjusts its selections without needing a fine-tune or vector store.
5. Reasoning effort chosen by measurement, with a fallback
Model and settings were picked by running VibeCheck's own prompt and validator over the same 8 scenes and checking every track against YouTube Music on both artist and title. gpt-oss-120b at high reasoning effort scored 97% and always returned the full count; default effort 92%; the newer qwen3.8-27b 58%, with most requests rate-limited. High effort costs 30–80 s and ~10k tokens, so a refusal or a run past 50 s falls back to default effort rather than failing the request. The same run found the prompt asked for 12 tracks while its mixing rules capped a mix at 7 — the rules are now proportions of the requested count.
| Layer | Technology |
|---|---|
| UI | Streamlit |
| LLM | gpt-oss-120b via Groq Cloud API (high reasoning effort, default-effort fallback) |
| Music validation | ytmusicapi (YouTube Music) |
| Concurrency | concurrent.futures.ThreadPoolExecutor |
| Deployment | Streamlit Community Cloud |
git clone https://github.com/YashNirwan/VibeCheck.git
cd VibeCheck
pip install -r requirements.txtAdd your Groq API key to .streamlit/secrets.toml:
GROQ_API_KEY = "your_key_here"streamlit run app.pyGet a free Groq API key at console.groq.com.