Add RAG basic implementation

This commit is contained in:
2026-11-09 08:43:30 +03:00
parent 447464c16c
commit 7fa1d774a0
76 changed files with 3030 additions and 124 deletions
+139
View File
@@ -0,0 +1,139 @@
import subprocess
import sys
import httpx
import pytest
from estimator.domain.embeddings import EMBEDDING_DIM
from estimator.domain.errors import LlmAuthError, LlmOutputError
from estimator.infra.embeddings import HashEmbedder, OpenAIEmbeddings
from tests.unit.llm.conftest import Recorder
E = HashEmbedder()
def dot(a, b):
return sum(x * y for x, y in zip(a, b, strict=True))
# ----------------------------------------------------------------- hash embedder
def test_vectors_have_the_schema_dimension_and_unit_length():
v = E.embed_one("Интеграция с amoCRM")
assert len(v) == EMBEDDING_DIM and dot(v, v) == pytest.approx(1.0)
def test_related_texts_are_closer_than_unrelated_ones():
a = E.embed_one("Интеграция с amoCRM")
related = E.embed_one("интеграции amoCRM и сделок")
unrelated = E.embed_one("Развёртывание и мониторинг")
assert dot(a, related) > 0.6
assert dot(a, unrelated) < 0.1
def test_case_and_russian_endings_do_not_matter():
assert dot(E.embed_one("Тестирование бота"), E.embed_one("ТЕСТИРОВАНИЯ БОТА")) > 0.99
def test_text_without_usable_words_still_gives_a_valid_vector():
for text in ("", " ", "и для на", "!!!"):
v = E.embed_one(text)
assert dot(v, v) == pytest.approx(1.0)
def test_embedding_is_deterministic_across_processes_and_hash_seeds():
code = (
"from estimator.infra.embeddings import HashEmbedder; import hashlib;"
"v = HashEmbedder().embed_one('Интеграция amoCRM'); print(hashlib.sha256(repr(v).encode()).hexdigest())"
)
out = [
subprocess.run([sys.executable, "-c", code], env={"PYTHONHASHSEED": seed, "PATH": ""}, capture_output=True, text=True, check=True).stdout
for seed in ("1", "2")
]
assert out[0] and out[0] == out[1]
async def test_async_interface_matches_the_sync_one():
[v] = await E.embed(["тест"])
assert v == E.embed_one("тест") and E.model == "hash-v1"
# ------------------------------------------------------------ OpenAI-compatible
def ok(vectors, order=None):
items = [{"index": i, "embedding": v} for i, v in enumerate(vectors)]
if order:
items = [items[i] for i in order]
return httpx.Response(200, json={"data": items})
def unit(i):
v = [0.0] * EMBEDDING_DIM
v[i] = 2.0 # deliberately not normalised
return v
def client(rec, sleeps, model="text-embedding-3-small", **kw):
return OpenAIEmbeddings(
model=model, base_url="https://emb.test/v1", api_key="sk-emb", transport=rec.transport,
sleep=sleeps, rng=lambda: 0.5, repair_attempts=0, **kw,
) # fmt: skip
async def test_request_shape_for_a_text_embedding_3_model(sleeps):
rec = Recorder(ok([unit(0), unit(1)]))
await client(rec, sleeps).embed(["один", "два"])
req = rec.requests[0]
assert str(req.url) == "https://emb.test/v1/embeddings" and req.headers["authorization"] == "Bearer sk-emb"
body = rec.body()
assert body["model"] == "text-embedding-3-small" and body["input"] == ["один", "два"]
assert body["encoding_format"] == "float" and body["dimensions"] == EMBEDDING_DIM
async def test_dimensions_are_not_sent_to_models_that_do_not_support_them(sleeps):
rec = Recorder(ok([unit(0)]))
await client(rec, sleeps, model="text-embedding-ada-002").embed(["x"])
assert "dimensions" not in rec.body()
async def test_vectors_are_normalised_and_returned_in_input_order(sleeps):
rec = Recorder(ok([unit(0), unit(1), unit(2)], order=[2, 0, 1])) # the provider answered out of order
vectors = await client(rec, sleeps).embed(["a", "b", "c"])
assert [v.index(1.0) for v in vectors] == [0, 1, 2]
assert all(dot(v, v) == pytest.approx(1.0) for v in vectors)
async def test_batches_follow_the_configured_size(sleeps):
rec = Recorder(*[ok([unit(i % 5) for i in range(n)]) for n in (64, 64, 22)])
vectors = await client(rec, sleeps, batch_size=64).embed([f"t{i}" for i in range(150)])
assert len(vectors) == 150 and len(rec.requests) == 3
assert [len(rec.body(i)["input"]) for i in range(3)] == [64, 64, 22]
async def test_no_request_for_no_texts(sleeps):
rec = Recorder()
assert await client(rec, sleeps).embed([]) == [] and rec.requests == []
async def test_very_long_and_empty_texts_are_made_safe(sleeps):
rec = Recorder(ok([unit(0), unit(1)]))
await client(rec, sleeps).embed(["я" * 20_000, ""])
sent = rec.body()["input"]
assert len(sent[0]) == 8000 and sent[1] == " "
async def test_wrong_vector_count_and_size_are_errors(sleeps):
with pytest.raises(LlmOutputError, match="asked for 2"):
await client(Recorder(ok([unit(0)])), sleeps).embed(["a", "b"])
short = httpx.Response(200, json={"data": [{"index": 0, "embedding": [0.1, 0.2]}]})
with pytest.raises(LlmOutputError, match=f"needs {EMBEDDING_DIM}"):
await client(Recorder(short), sleeps).embed(["a"])
async def test_provider_errors_use_the_shared_handling(sleeps):
with pytest.raises(LlmAuthError):
await client(Recorder(httpx.Response(401, json={"error": {"message": "bad"}})), sleeps).embed(["a"])
rec = Recorder(httpx.Response(429, headers={"retry-after": "2"}, json={}), ok([unit(0)]))
assert len(await client(rec, sleeps).embed(["a"])) == 1 and sleeps.calls == [2.0]