Mock Embeddings

Plugin ID: nexus.embeddings.mock

A deterministic, zero-I/O embeddings.provider for tests and offline development. Same text in → same vector out. No network, no API key, sub-millisecond per call.

The vectors are derived from the SHA-256 hash of the input text. Two different inputs map to two different (but stable) vectors; the same input always maps to the same vector. The output is unit-normalized so backends like chromem-go that require normalized inputs accept it directly.

This plugin ships in the main binary alongside production providers — same pattern as nexus.io.test. It is opt-in; to use it just include it in plugins.active.

Configuration

nexus.embeddings.mock:
  # dimensions: 128            # default
  # model: mock-embedding      # default; echoed back in EmbeddingsRequest.Model

When to use it

• Integration tests. The RAG integration tests under tests/integration/rag_test.go use this plugin so the suite runs without an API key and without spending money. See Integration Testing.
• Adapter and storage development. When iterating on a new vector store backend, you don’t want a real embedding API in the loop. Wire up nexus.embeddings.mock to keep the focus on the storage side.
• Pipeline smoke tests. Boot a CLI ingest with mock embeddings to confirm chunking, persistence paths, and watch-mode wiring all work, then swap in a real provider.

When not to use it

The mock has no semantic structure — “cat” and “kitten” produce vectors as different as “cat” and “transistor radio”. Tests that depend on similarity rankings for meaning will give wrong-shaped signals. The included integration tests work around this by querying with the exact stored content (which has bit-for-bit-identical vectors and produces similarity ≈ 1.0).

Don’t use this plugin in production. Don’t use it for anything where retrieval quality matters.

Events

Same as any embeddings.provider: subscribes to embeddings.request, fills the result in place. See Embedding Providers for the full payload shape.