Embeddings (semantic re-rank)

A local semantic signal layered into wf query's ranking: cosine(query, claim) fused with the lexical score on the top-40 candidates. Ships off by default — the measured value is conditional, see when it earns its keep.

Setup

uv tool install wiki-fabric --with fastembed    # ONNX runtime + model (once)

integrations:
  embeddings:
    enabled: true
    model: all-MiniLM-L6-v2    # default; ~30 MB, cached offline

What changes when it's on

• wf query builds (or loads) a local vector index over claim/pattern statements, scores every top-40 lexical candidate semantically, and fuses into the ordering — RETRIEVAL["fuse_lex"]/["fuse_sem"] (0.5/0.5). ~5 ms/query once vectors are cached; no network (fastembed ONNX, fully offline).
• wf status shows Embeddings: enabled (vectors N).
• The System One fusion rerank (when the judgment tier is on) applies on top — embeddings widen the candidate pool semantically; the decision model judges which candidates answer.

Determinism note

The lexical base and the graph expansion remain byte-identical and re-runnable; embeddings add one more numeric signal to the fusion — same corpus + same query ⇒ same fused order (model is pinned, vectors cached). For determinism audits that can't accept an ML component at all: leave it off (the default).

When it earns its keep

• Corpus > ~2k claims: lexical collisions multiply and embedding space still separates topics.
• Cross-project word divergence is common (same concept, different vocabulary per repo).

Skip it when: small corpus; strictly within-topic queries. Measured in the exp-embeddings-spike-2026-09-28 experiment: at ~500 claims the gain was thin (+3 rel@10, none at @5) — the value conditions above are the result of that spike, not vibes. Re-benchmark when the corpus shape changes materially.