UnisKB/apps/knowledge/test_embedding_params.py

74 lines
3.5 KiB
Python
Raw Blame History

This file contains ambiguous Unicode characters!

This file contains ambiguous Unicode characters that may be confused with others in your current locale. If your use case is intentional and legitimate, you can safely ignore this warning. Use the Escape button to highlight these characters.

import unittest
from types import SimpleNamespace
from knowledge.embedding_params import ConfiguredEmbedding, validate_params, knowledge_params
class RecordingModel:
batch_size = 10
def __init__(self):
self.calls = []
def embed_query(self, text):
self.calls.append(('query', text))
return [3., 4.]
def embed_documents(self, texts):
self.calls.append(('documents', texts))
return [[3., 4.] for _ in texts]
class EmbeddingParamsTests(unittest.TestCase):
def test_inference_uses_separate_instructions_and_same_normalization(self):
base = RecordingModel()
model = ConfiguredEmbedding(base, {'normalize': True, 'query_prefix': '检索:', 'document_prefix': '正文:'})
self.assertEqual(model.embed_query('内容'), [.6, .8])
self.assertEqual(model.embed_documents(['内容']), [[.6, .8]])
self.assertEqual(base.calls, [('query', '检索:内容'), ('documents', ['正文:内容'])])
def test_batch_size_is_applied_to_real_calls(self):
base = RecordingModel()
model = ConfiguredEmbedding(base, {'batch_size': 2})
self.assertEqual(len(model.embed_documents(['a', 'b', 'c', 'd', 'e'])), 5)
self.assertEqual([len(call[1]) for call in base.calls], [2, 2, 1])
def test_knowledge_overrides_do_not_leak_between_shared_models(self):
base = RecordingModel()
first = ConfiguredEmbedding(base, {'query_prefix': 'A'})
second = ConfiguredEmbedding(base, {'query_prefix': 'B'})
first.embed_query('x'); second.embed_query('x'); first.embed_query('x')
self.assertEqual(base.calls, [('query', 'Ax'), ('query', 'Bx'), ('query', 'Ax')])
self.assertEqual(base.batch_size, 10)
def test_invalid_inputs_are_rejected(self):
for params in ([], {'batch_size': True}, {'batch_size': 0}, {'batch_size': 257},
{'batch_size': 2.5}, {'normalize': 'true'}, {'query_prefix': 1},
{'document_prefix': 'x' * 2001}, {'': 'value'}, {'custom': {}},
{'custom': ['value']}, {'x': 1, 'X': 2}, {'custom': float('nan')},
{'a' * 65: 'value'}, {str(i): 'value' for i in range(51)}):
with self.subTest(params=params), self.assertRaises(ValueError):
validate_params(params)
def test_custom_names_are_saved_but_unsupported_options_do_not_reach_inference(self):
params = {'领域': '电力', 'dimensions': 512, '备注': '', 'batch_size': 2, 'query_prefix': '检索:'}
self.assertEqual(validate_params(params), params)
active = knowledge_params(SimpleNamespace(meta={'embedding_params': params}))
self.assertEqual(active, {'batch_size': 2, 'query_prefix': '检索:'})
base = RecordingModel()
model = ConfiguredEmbedding(base, active)
model.embed_query('问题')
self.assertEqual(base.calls, [('query', '检索:问题')])
def test_legacy_and_empty_params_keep_original_behavior(self):
self.assertEqual(knowledge_params(SimpleNamespace(meta={})), {})
self.assertEqual(validate_params({'batch_size': None, 'query_prefix': ''}), {})
base = RecordingModel()
self.assertEqual(ConfiguredEmbedding(base, {}).embed_query('x'), [3., 4.])
self.assertEqual(ConfiguredEmbedding(base, {}).embed_documents([]), [])
def test_zero_vector_is_safe(self):
base = RecordingModel()
base.embed_query = lambda _: [0., 0.]
self.assertEqual(ConfiguredEmbedding(base, {'normalize': True}).embed_query('x'), [0., 0.])