Vector Database

Storing and searching millions of embeddings at speed

🔴 MUST

🧠 Understand👁️ Visualize💻 Code🔬 Experiment⚖️ Compare🗣️ Explain

⚡ Interactive Visualizer HNSW Graph & IVF Voronoi Simulator

Approximate Nearest Neighbor (ANN) Index Simulator

Click anywhere inside the 2D vector space below to place a Query Vector ($\vec{q}$). Watch how different indexing structures prune the search space to find Top-$K$ nearest neighbors in sub-millisecond time.

Click to move Query Vector $\vec{q}$ (🟡) Index: HNSW Graph ($O(\log N)$)

Top-3 Retrieval Results 92% Comparisons Saved

Performance Stats:

Comparisons: 8 / 60
Recall Rate: 98.5%
Search Latency: 0.12 ms
Index Memory: ~1.4 MB

🔍 Behind-the-Scenes: HNSW vs. IVF-PQ Trade-offs in Production

Index Type	Build Time	Query Latency	RAM Overhead	Best Used For
HNSW	Slow ($O(N \log N)$)	Ultra-Fast (<2ms)	High (+150% RAM)	Low-latency real-time RAG (<10M vectors)
IVF-PQ	Moderate (k-means)	Fast (<10ms)	Minimal (-80% RAM)	Billion-scale web search (e.g. Pinecone, Milvus)
Flat (Brute Force)	Instant (0s)	Linear $O(N)$ (Slow)	Zero overhead	Small collections (<50k vectors) with 100% recall

8 subtopics — click to study

Why Vector DB? 🔴 Store Embeddings 🔴 Vector Index 🟡 Similarity Search 🔴 Metadata 🟡 Metadata Filtering 🟡 Top-K Retrieval 🔴 Retrieval Performance 🟡

← Back to Roadmap