MinVector
Embeddings that adapt to scale, cost, and context
A backend-agnostic system for compressing, slicing, and retrieving embeddings across semantic granularity built for efficient, flexible, and long-lived AI systems.
What is MinVector?
What if a single embedding could work at multiple resolutions — without retraining your entire system?
Retrieve Fast
Use shorter embedding prefixes for lightning-fast retrieval when speed matters most.
Retrieve Deep
Scale up to full precision when accuracy is critical — without changing your API.
Dynamic Trade-offs
Balance cost, latency, and quality on the fly based on your system's needs.
Nested Representations
Instead of fixed-length vectors, MinVector structures embeddings so shorter prefixes remain meaningful while longer versions add precision.
How MinVector Works
Built around three core principles that enable flexible, efficient retrieval without compromising on quality.
Nested Representations
Each embedding is structured so that smaller prefixes remain semantically meaningful, while larger slices add detail.
Backend Agnostic
MinVector doesn't lock you into a specific model or provider. Any embedding backend can be wrapped with the same interface.
Optional Training
Use runtime slicing immediately, or train a lightweight projection head on top of frozen encoders for better alignment.
The same embedding works at multiple resolutions
What MinVector Is (and Isn't)
An infrastructure primitive — meant to be built upon.
MinVector is
- A research-oriented OSS project
- A foundation for efficient retrieval systems
- A composable building block
- An infrastructure primitive
MinVector is not
- A hosted SaaS
- A vector database replacement
- A magic accuracy booster
- A black box solution
Early Open-Source Release
MinVector is in v0.1 — a solid foundation with more to come.
Available Now
- Clean adapter API for nested embeddings
- Optional training via projection head
- CLI utilities for experimentation
- Strong foundation for extensions
On the Roadmap
- Indexing strategies for multi-resolution recall
- Evaluation tooling across slice depths
- Additional backends and storage integrations
- Deeper theoretical exploration
Explore the Code
If you're curious about embedding compression, retrieval efficiency, or AI-native memory structures — you'll likely find MinVector interesting.
View on GitHub