Engram-protocol/engram

1"""Tests for kvcos.engram.embedder — unified fingerprint embedding."""2 3import pytest4import torch5import torch.nn.functional as F6 7from kvcos.engram.embedder import (8    HashEmbedder,9    get_embedder,10    get_fingerprint,11    reset_embedder,12)13 14 15class TestHashEmbedder:16    def test_deterministic(self):17        emb = HashEmbedder(dim=128)18        fp1 = emb.embed("hello")19        fp2 = emb.embed("hello")20        assert torch.allclose(fp1, fp2)21 22    def test_different_text(self):23        emb = HashEmbedder(dim=128)24        fp1 = emb.embed("hello")25        fp2 = emb.embed("world")26        assert not torch.allclose(fp1, fp2)27 28    def test_normalized(self):29        emb = HashEmbedder(dim=128)30        fp = emb.embed("test")31        norm = torch.norm(fp).item()32        assert abs(norm - 1.0) < 0.0133 34    def test_dimension(self):35        emb = HashEmbedder(dim=256)36        fp = emb.embed("test")37        assert fp.shape == (256,)38        assert emb.dim == 25639 40    def test_source_tag(self):41        emb = HashEmbedder()42        assert emb.source == "hash-fallback"43 44 45class TestGetFingerprint:46    def test_returns_tensor_and_source(self):47        fp, source = get_fingerprint("test text")48        assert isinstance(fp, torch.Tensor)49        assert isinstance(source, str)50        assert source in ("llama_cpp", "sbert", "hash-fallback")51 52    def test_deterministic(self):53        fp1, _ = get_fingerprint("same text")54        fp2, _ = get_fingerprint("same text")55        assert torch.allclose(fp1, fp2)56 57 58class TestSBertEmbedder:59    """Test sbert if available (installed in this venv)."""60 61    def test_sbert_available(self):62        """Verify sentence-transformers is usable."""63        try:64            from kvcos.engram.embedder import SBertEmbedder65            emb = SBertEmbedder()66            assert emb.source == "sbert"67            assert emb.dim == 38468        except ImportError:69            pytest.skip("sentence-transformers not installed")70 71    def test_semantic_discrimination(self):72        """Related texts should be more similar than unrelated."""73        try:74            from kvcos.engram.embedder import SBertEmbedder75            emb = SBertEmbedder()76        except ImportError:77            pytest.skip("sentence-transformers not installed")78 79        fp_a = emb.embed("machine learning neural network training")80        fp_b = emb.embed("deep learning model optimization")81        fp_c = emb.embed("chocolate cake baking recipe")82 83        sim_ab = F.cosine_similarity(fp_a.unsqueeze(0), fp_b.unsqueeze(0)).item()84        sim_ac = F.cosine_similarity(fp_a.unsqueeze(0), fp_c.unsqueeze(0)).item()85 86        assert sim_ab > sim_ac, (87            f"Related topics ({sim_ab:.4f}) should be more similar "88            f"than unrelated ({sim_ac:.4f})"89        )90 91 92class TestGetEmbedder:93    def test_singleton(self):94        reset_embedder()95        e1 = get_embedder()96        e2 = get_embedder()97        assert e1 is e298 99    def test_reset(self):100        reset_embedder()101        e1 = get_embedder()102        reset_embedder()103        e2 = get_embedder()104        # After reset, a new instance is created105        # (may or may not be same object depending on strategy)106        assert e2 is not None107