Gaming YouTube Search

Search for tips
Search for techniques
Search for opinions

🎮 Inspiration

Game guides and reviews are scattered across hours-long videos. We wanted a tool that feels like “Ctrl + F for YouTube,” surfacing the exact 30 seconds you need—no scrubbing, no spoilers.

🛠️ How We Built It

Data pipeline
- Scraped metadata & transcripts of videos for ~1 000 games.
- Stored raw docs + embeddings in MongoDB Atlas; chunked transcripts for time-level recall.
Three purpose-built vector indexes
- games – reliable name matching
- tips-tricks – guides & walkthroughs
- reviews – critiques & first-looks
AI retrieval workflow
- Prompt Refiner (Gemini-Flash) extracts game titles + intent from the user’s NL query.
- Vector search (Atlas Vector Search) fetches top-k chunks, optionally pre-filtered by gameId.
- Neighboring chunks are fetched from Firestore to give context without extra Atlas storage.
- Response Refiner (Gemini-Flash) ranks chunks, adjusts timestamps, and explains relevance (RAG).
Frontend
- Built in Flutter Web → snappy, single-page UX.
- Deployed on Firebase Hosting; queries hit a Cloud Functions worker that orchestrates the flow.

🤓 What We Learned

Designing multi-index RAG: balancing precision (vector search) vs. recall (context expansion).
Cost-savvy hybrid storage: Atlas for search-critical data, Firestore for cold context.
Prompt engineering for timestamp alignment inside video transcripts.

🚧 Challenges

Synonym explosion in game titles (localized names, DLCs). Solved with a lightweight embedding-based matcher.
Keeping latency < 2 s despite two LLM hops—caching embeddings and parallelizing index hits did the trick.