Design decisions

ELIPS targets C++23 to use std::span, structured bindings, designated initialisers, and concepts. The Core Guidelines are non-negotiable.

No server, no daemon. A database is a directory, a process opens it, and a file lock coordinates writers.

Hierarchical Navigable Small World offers the best recall / latency trade-off at the embedded scale ELIPS targets.

Record-framed with CRC32C. Recovery truncates at the first invalid record. Logical, not physical.

Single-writer, multi-reader. Atomic batched transactions buffer in memory and commit at once.

Lightweight, zero-overhead bindings that mirror the C++ surface 1:1 plus a typed wrapper for ergonomics.

The snapshot is the consistent baseline; the WAL is the log of mutations since. Together they make checkpoint atomic.

flock-based advisory locks. RAII-bound. No background coordination.

Payloads are typed variants, not a fixed schema. Filters operate on the same type system.

A small expression-oriented language with a lexer, parser, AST, and executor. EQL output is the same Filter / QueryPlan the SDK produces.

Backends live behind a GpuBackend abstraction, selected once at startup by BackendSelector. Metal is the implemented path; CUDA and HIP/ROCm are experimental; CPU is the universal fallback.

ELIPS ships its own distance and top-k kernels rather than depending on vendor graph libraries. Metal implements distance/top-k; richer GPU index kernels are future work.

Index build can use the GPU even when serving stays on CPU. The transfer manager handles the handoff.

A pool allocator keeps allocations stable across queries; the memory manager owns the lifetime.

QueryBatcher coalesces concurrent single-vector queries into one launch to amortise kernel overhead.

On Apple silicon ELIPS skips host↔device copies entirely — the Metal backend's main advantage.

FP32 by default; mixed precision is opt-in per index, never silent.

Backend → device → CPU. A named-but-absent backend raises BackendUnavailableError unless fallback is requested.