foley.eval.embedder#

A deterministic, numpy-only text embedder for the retrieval eval gate.

HashingBowEmbedder is a hashing bag-of-words text embedder — the shippable sibling of the test suite’s FakeEmbedder. It exists so the nDCG PR gate is deterministic and CLAP-free: unlike FakeEmbedder’s random embed_audio (whose content-noise vectors make RRF rankings platform-dependent — the flakiness documented in tests/test_bootstrap.py), the eval harness injects embed_text(caption + tags) as each clip’s vector, so both search legs (vector cosine and BM25) carry the same caption bag-of-words. Every golden query then shares 2–3 tokens with its answer clip → integer rank-1 in both legs → a bit-exact, cross-platform-stable nDCG.

It is a text-only embedder: embed_audio is intentionally unimplemented (the gate never calls it — vectors are injected via foley.index.library.SoundLibrary.add()).

Module Attributes

DEFAULT_DIM

Default embedding width (matches the test FakeEmbedder for parity).

Classes

HashingBowEmbedder(*[, dim])

Deterministic hashing bag-of-words text embedder (L2-normalized).

foley.eval.embedder.DEFAULT_DIM = 64#: Default embedding width (matches the test FakeEmbedder for parity).

class foley.eval.embedder.HashingBowEmbedder(*, dim=64)[source]#

Bases: object

Deterministic hashing bag-of-words text embedder (L2-normalized).

Two texts sharing a token share a dimension, so cosine similarity tracks lexical overlap — plausible and reproducible (a stable hashlib hash, never the salted builtin hash). Suitable only for the eval gate / offline retrieval regression; real retrieval uses CLAP.

embed_text(text)[source]#: Embed text (or a list of texts) -> a (n, dim) float32 array.