Conceptual overview

This page is descriptive, not normative. It explains bijux-vex in plain language without APIs or enums.

What bijux-vex is

• A vector execution engine: it runs vector similarity computations under explicit contracts.
• Contracts come in two flavors:
• Deterministic: exact algorithms, replayable outputs, bit-stable provenance.
• Non-deterministic (experimental): ANN/approximate algorithms, bounded divergence, declared randomness.
• Every run is tied to an execution plan, provenance, and artifacts that record how results were produced.

Vector execution vs. vector DB

• Vector DBs focus on storage and serving; execution is implicit and often approximate by default.
• bijux-vex focuses on execution semantics: what was run, under which guarantees, and how to replay or compare results.
• Artifacts, plans, and provenance are first-class; storage backends are just sources and ledgers.

Core objects (conceptual)

• Execution plan: the declared algorithm, contract, and parameters.
• Execution artifact: the outputs plus provenance and signatures.
• Provenance: the audit trail that explains how the artifact was produced.
• Randomness profile: where nondeterminism enters and how it is bounded.

How determinism works here

• Deterministic runs use exact search and must replay identically.
• Non-deterministic runs must emit approximation reports and randomness metadata; replay checks envelopes, not equality.

How to approach the docs

• If you want rules: go to spec/system_contract.md and spec/execution_contracts.md.
• If you want rationale: see design/why_vector_execution.md and design/contracts.md.
• If you want to run something: start with examples/overview.md.