The Problem: Videos Are Black Boxes
You watched a 2-hour lecture last week. Somewhere in there, the speaker explained a concept you need right now. Good luck finding it.
Videos are the fastest-growing content format, yet they remain the hardest to search. Unlike text documents, you can't Ctrl+F a video. Your options:
- Scrub manually: Drag the timeline and hope you recognize the right frame
- Rely on timestamps: If someone bothered to add chapters (most don't)
- Rewatch everything: The nuclear option—and the most common one
For organizations managing video libraries—training departments, media companies, educational platforms—this problem scales painfully. Thousands of hours of content, completely unsearchable.
What if you could search your entire video library with a simple question, like "when does he explain the pricing model"?
What We Built: Semantic Video Search Engine
We built a platform that makes video content as searchable as text. Upload a video, and AI analyzes every moment. Then search across your entire library using natural language.
How It Works
| Step | What Happens | Details |
|---|---|---|
| 1. Upload | Add a YouTube URL or video link | Any publicly accessible video |
| 2. AI Analysis | Gemini 2.5 Flash processes the video | Extracts timestamps, events, summaries |
| 3. Embedding | Each timestamp gets a vector embedding | Enables semantic similarity matching |
| 4. Search | Type a natural language query | Returns ranked results with timestamps |
What the AI Extracts
For each video, the system automatically identifies:
- Key moments with precise timestamps
- Event summaries describing what happens at each point
- Keywords for improved search matching
- Category classification across 15 content types (education, tech, news, entertainment, etc.)
The Search Experience
Natural Language, Not Keywords
Traditional search requires exact keyword matches. Our semantic search understands meaning:
| Query | What It Finds |
|---|---|
| "when does the pricing discussion start" | The exact moment pricing is mentioned, even if the word "pricing" isn't spoken |
| "the part about customer complaints" | Segments discussing customer feedback, issues, or dissatisfaction |
| "technical architecture overview" | Moments where system design or architecture is explained |
Ranked Results with Context
Each search result returns:
- Timestamp — Jump directly to the relevant moment
- Context summary — What's happening at that point
- Confidence score — How closely it matches your query
- Source video — Which video in your library contains it
Technical Architecture
The AI Pipeline
Video Processing:
- Video submitted via URL (YouTube or direct link)
- Google Gemini 2.5 Flash analyzes visual and audio content
- AI extracts structured timestamps with descriptions
- Each timestamp segment gets vectorized using Gemini embeddings
Search:
- User query converted to vector embedding
- PostgreSQL + pgvector performs similarity search across all indexed timestamps
- Results ranked by cosine similarity score
- Configurable thresholds filter low-confidence matches
Why Gemini + pgvector?
| Component | Choice | Why |
|---|---|---|
| Video AI | Gemini 2.5 Flash | Native video understanding—no need to extract frames or transcribe audio first |
| Vector DB | Supabase + pgvector | Production-ready, scalable, and runs alongside your existing Postgres data |
| API | FastAPI (Python) | Async processing for concurrent video analysis jobs |
| Frontend | Next.js 15 | Real-time progress tracking via Supabase subscriptions |
Real-Time Processing Feedback
Users see live progress as videos are analyzed—no waiting blindly for a job to finish. The frontend subscribes to database changes via Supabase real-time, updating the UI as each timestamp is extracted.
Built as a SaaS Product
Credit-Based Pricing
The platform runs on a sustainable credit model:
| Tier | Price | Credits/Month | Best For |
|---|---|---|---|
| Free | $0 | 600 (~1 hr of video) | Testing the platform |
| Starter | $12/mo | 14,400 | Individual creators |
| Pro | $29/mo | 36,000 + unlimited search | Teams and educators |
| Business | $99/mo | 150,000 + unlimited search | Media companies |
Stripe Integration
Full payment processing with subscription management, credit top-ups, and usage tracking. Activity logs track every search and analysis operation.
API Access
Power users get API keys for programmatic access—integrate video search into their own applications, batch process YouTube channels, or build custom workflows.
Results
| Metric | Value |
|---|---|
| Search accuracy | Semantic matching across meaning, not just keywords |
| Processing speed | ~2 min per hour of video content |
| Query response | Sub-second search across thousands of timestamps |
| Content categories | 15 automatic classifications |
What an Education Platform Said:
"We had 500+ hours of lecture recordings that students couldn't navigate. Now they search 'explain the Krebs cycle' and jump to the exact moment. Support tickets about finding content dropped to near zero."
— Online learning platform, Southeast Asia
Who This Is For
This solution works for:
- Education platforms making lecture libraries searchable
- Media companies indexing broadcast archives
- Training departments building searchable knowledge bases
- Content creators helping audiences navigate long-form content
- Research teams analyzing interview and focus group recordings
- Legal teams searching deposition and testimony videos
If your team has more than 10 hours of video content that people need to reference, this platform pays for itself in the first week.