Switch the embed model to multilingual-e5-small for the European corpus

Measured on the live roles against the "Copenhagen engineering" agent's
intent: the English-only all-MiniLM-L6-v2 separates engineering from
non-engineering roles with an AUC of 0.68 overall but only 0.74 restricted
to Danish-country roles (a Danish engineering role beats the rest's
non-engineering median only 68% of the time). multilingual-e5-small reaches
0.70 overall and 1.00 on Danish-country roles, for about 3x the CPU time.
paraphrase-multilingual-MiniLM-L12-v2 measured worse than the English-only
baseline (AUC 0.60) and was dropped.

e5 needs "query: "/"passage: " prefixes; app/jobs/embed.py adds them by
matching the model name so a model change can't leave a stale prefix
setting behind. Added tests for the prefixing and for the two invariants
the model change relies on: embed drops the old model's vectors before
re-embedding, and an agent's cached intent vector is keyed by model so a
model change is a cache miss, never a stale-dimension read.

Co-Authored-By: Claude Sonnet <noreply@anthropic.com>
This commit is contained in:
WolfandClaude Sonnet committed 2026-09-29 13:59:47 +02:00
1 parent 0558ef19ed
commit b560cdce2a
7 files changed
+214 -15

No files matched your search

+144 -2
View File
@@ -124,6 +124,147 @@ def test_embed_without_sentence_transformers_says_so_and_does_nothing(
assert store.vector_counts() == {}
# -------------------------------------------------------- e5 prefixes
@pytest.mark.parametrize("model, wants", [
("intfloat/multilingual-e5-small", True),
("intfloat/e5-base-v2", True),
("intfloat/e5-large", True),
("sentence-transformers/all-MiniLM-L6-v2", False),
("sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2", False),
("fake", False),
("some-org/e5models-are-not-this", False),
])
def test_wants_e5_prefix_matches_the_e5_family_by_name(model: str, wants: bool) -> None:
assert embed.wants_e5_prefix(model) == wants
def test_role_text_and_intent_text_add_the_e5_prefixes_only_for_e5(
store: Store) -> None:
uid = role(store, "1", "Backend Engineer", body="Python services")
r = store.role(uid)
assert r is not None
assert embed.role_text("fake", r) == embed.document(r)
assert embed.role_text("intfloat/multilingual-e5-small", r) == f"passage: {embed.document(r)}"
assert embed.intent_text("fake", "python backend") == "python backend"
assert embed.intent_text("intfloat/multilingual-e5-small", "python backend") == \
"query: python backend"
class FakeE5Embedder:
"""A second fake, standing in for a genuinely different model: its own
(narrower) vector width and an e5-shaped model id, so it exercises both
the prefixing and the re-embedding-on-model-change paths without
downloading anything."""
model_id = "intfloat/multilingual-e5-small"
def __init__(self) -> None:
self.calls: list[list[str]] = []
def encode(self, texts: Sequence[str]) -> np.ndarray:
self.calls.append(list(texts))
rows = []
for text in texts:
lower = text.lower()
vec = np.asarray([lower.count(w) for w in WORDS[:2]] + [0.1, 0.2], dtype=np.float32)
rows.append(vec / np.linalg.norm(vec))
return np.vstack(rows)
def test_embed_prefixes_role_text_for_an_e5_model(
store: Store, tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> None:
e5 = FakeE5Embedder()
monkeypatch.setattr(embed, "load_embedder", lambda model: e5)
config = Config(database=str(tmp_path / "roledesk.db"),
embed={"model": "intfloat/multilingual-e5-small"})
role(store, "1", "Backend Engineer", body="Python services")
run(config, store, embed.run_embed)
assert e5.calls and e5.calls[0][0].startswith("passage: Backend Engineer")
def test_score_prefixes_the_intent_for_an_e5_model(
store: Store, tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> None:
e5 = FakeE5Embedder()
monkeypatch.setattr(embed, "load_embedder", lambda model: e5)
config = Config(database=str(tmp_path / "roledesk.db"),
embed={"model": "intfloat/multilingual-e5-small"})
role(store, "1")
store.save_agent("Backend", "python backend")
run(config, store, score.run_score)
assert e5.calls == [["query: python backend"]]
# ------------------------------------------------------- model changes
def test_changing_the_embed_model_drops_old_vectors_and_reembeds(
store: Store, tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> None:
"""The doc says a model change re-embeds; this is the guarantee that
actually makes that true, and the one that stops a run ever mixing two
models' vectors, which would be two different geometries under one
dimension count and garbage cosines."""
fake_a, fake_e5 = FakeEmbedder(), FakeE5Embedder()
embedders = {"fake": fake_a, "intfloat/multilingual-e5-small": fake_e5}
monkeypatch.setattr(embed, "load_embedder", lambda model: embedders[model])
role(store, "1")
role(store, "2")
config_a = Config(database=str(tmp_path / "roledesk.db"), embed={"model": "fake"})
run(config_a, store, embed.run_embed)
assert store.vector_counts() == {"fake": 2}
dim_a = next(iter(store.vectors("fake").values())).shape[-1]
config_e5 = Config(database=str(tmp_path / "roledesk.db"),
embed={"model": "intfloat/multilingual-e5-small"})
result = run(config_e5, store, embed.run_embed)
# The old model's rows are gone, not just shadowed: `vectors(model)`
# filters by model, so a leftover row could never be read back mixed in,
# but drop_vectors is what stops it existing at all.
assert store.vector_counts() == {"intfloat/multilingual-e5-small": 2}
assert store.vectors("fake") == {}
assert result["dropped"] == 2
dim_e5 = next(iter(store.vectors("intfloat/multilingual-e5-small").values())).shape[-1]
assert dim_a != dim_e5, "the fakes must differ in width for this test to mean anything"
def test_changing_the_embed_model_recomputes_the_agents_cached_intent_vector(
store: Store, tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> None:
"""`intent_vector` only returns a cache hit when `intent_key` -- which
hashes the model in -- matches what is stored, so a model change is a
cache miss and `_intent` re-encodes under the new model, never handing a
stale-dimension vector into a run's cosines."""
fake_a, fake_e5 = FakeEmbedder(), FakeE5Embedder()
embedders = {"fake": fake_a, "intfloat/multilingual-e5-small": fake_e5}
monkeypatch.setattr(embed, "load_embedder", lambda model: embedders[model])
uid = role(store, "1")
agent = store.save_agent("Backend", "python backend")
config_a = Config(database=str(tmp_path / "roledesk.db"), embed={"model": "fake"})
run(config_a, store, embed.run_embed)
run(config_a, store, score.run_score)
vec_a = store.intent_vector(agent.id, score.intent_key("fake", "python backend"))
assert vec_a is not None
config_e5 = Config(database=str(tmp_path / "roledesk.db"),
embed={"model": "intfloat/multilingual-e5-small"})
run(config_e5, store, embed.run_embed)
run(config_e5, store, score.run_score)
# The old key no longer resolves (the row moved on); the new key does,
# at the new model's own width, and scoring this role did not raise.
assert store.intent_vector(agent.id, score.intent_key("fake", "python backend")) is None
vec_e5 = store.intent_vector(
agent.id, score.intent_key("intfloat/multilingual-e5-small", "python backend"))
assert vec_e5 is not None
assert vec_e5.shape != vec_a.shape
assert uid in store.scores(agent.id)
# ------------------------------------------------------------------ terms
def test_recency_halves_every_fourteen_days() -> None:
@@ -361,8 +502,9 @@ def test_a_closed_liked_role_still_shapes_the_taste(
# ------------------------------------------------------------ fit scaling
def test_scale_maps_the_closest_to_one_and_the_farthest_to_zero() -> None:
# The spread the last real run showed under `(cos + 1) / 2`.
assert score.scale({"a": 0.55, "b": 0.79, "c": 0.67}) == pytest.approx(
# The spread the last real run showed under `(cos + 1) / 2`, on
# multilingual-e5-small.
assert score.scale({"a": 0.87, "b": 0.95, "c": 0.91}) == pytest.approx(
{"a": 0.0, "b": 1.0, "c": 0.5})
assert score.scale({"only": 0.3}) == {"only": 1.0}
assert score.scale({"a": 0.4, "b": 0.4}) == {"a": 1.0, "b": 1.0}