Chassis

Chassis is an embeddable, on-disk vector storage engine written in Rust.

It is designed to be used as a local storage component for vector similarity search. Chassis runs in-process, stores data on disk, and does not require a server or external dependencies.

The project is early-stage and focused on establishing a correct, stable storage core.

Current Capabilities

Chassis provides a high-performance vector storage, graph construction, and search core:

Python bindings

• Native bindings: Full Python support via the chassis package (sources in pychassis/; version tracks the Rust release, currently v0.6.3).
• Zero-copy: NumPy integration for high-throughput vector transfer.
• Install: pip install . from pychassis/, or install a matching wheel when published for your platform.

Universal Interface

• Stable C ABI: A fully compliant C-compatible FFI layer enables Chassis to be embedded in C, C++, Node.js, and Go.
• Safety Fortress: The ffi_guard architecture guarantees that Rust panics never crash the host process.
• Opaque Handle Design: Uses the "Pimpl" pattern to ensure ABI stability while hiding internal Rust implementation details.

High-Level API

• VectorIndex Facade: A clean, unified entry point that orchestrates storage, compute, and graph operations.
• Consistency Orchestration: Automates the "Register Last" insertion protocol to guarantee readers never see uninitialized data.
• Ghost Node Recovery: Automatically detects and recovers from partial writes during power loss (ADR-005).

Storage Layer

• Zero-Copy I/O: Memory-mapped vectors allow accessing 1536d embeddings in nanoseconds.
• ACID Persistence: Explicit fsync-backed commit strategy.
• Fixed-Width Geometry: O(1) deterministic addressing for all on-disk lookups.

Graph Layer

• HNSW Construction: Fully persistent, crash-safe graph topology builder.
• Bidirectional Linking: Maintains graph navigability with "Small World" guarantees.
• Diversity Heuristics: Implements robust neighbor pruning (Heuristic 2) to prevent clustering.
• Crash Consistency: Atomic write ordering ensures the graph structure is never corrupted, even on power loss.

Search Layer

• SIMD Acceleration: Hardware-accelerated distance kernels (AVX2 for x86, NEON for ARM) provide ~23 Gelem/s throughput.
• Zero-Allocation Traversal: The hot search path allocates no heap memory, ensuring consistent P99 latency.
• High Performance: Achieves sub-50µs latency for 1536d vectors (OpenAI embeddings) on commodity hardware.

Design Principles

Chassis prioritizes:

• Correctness over feature breadth
• Explicit invariants over implicit behavior
• Local-first operation with predictable performance
• Simple, inspectable file formats

The storage layer is intentionally conservative. Durability, growth strategy, and concurrency semantics are defined explicitly and documented.

Non-Goals

Chassis does not aim to be:

• A database server
• A cloud service
• A distributed system
• A query engine

These concerns are intentionally left to the embedding application.

Status

v0.6.3 (Stable) — May 2026

Patch release: SPDX workspace license, deny.toml for cargo deny, and rand bump (RUSTSEC-2026-0097). See CHANGELOG.md.

The core storage engine, C FFI layer, and Python bindings are feature-complete and ready for use. Release history and per-version notes live in CHANGELOG.md.

License

Chassis is dual licensed under:

• Apache License 2.0
• MIT License

You may use either license at your option.

Contributing

Contributions and design discussion are welcome.

The project currently prioritizes correctness, simplicity, and clear invariants over feature breadth. See CONTRIBUTING.md for details.