jina_v5

math-tutor/backend/app/config.py

@@ -33,8 +33,8 @@ def get_embedding_settings() -> EmbeddingSettings:
         return EmbeddingSettings(
             embedding_type=embedding_type,
             model=os.getenv("SENTENCE_TRANSFORMER_MODEL",
-                            "jinaai/jina-embeddings-v4"),
-            target_dim=int(os.getenv("EMBEDDING_DIM", "512")),
+                            "jinaai/jina-embeddings-v5-text-small-retrieval"),
+            target_dim=int(os.getenv("EMBEDDING_DIM", "1024")),
         )
     if embedding_type == "openai-like":
         base_url = os.getenv("OPENAI_BASE_URL")

math-tutor/backend/app/deterministic_services/embeddings.py

@@ -158,7 +158,9 @@ class SentenceTransformerEmbeddings(BaseEmbeddings):
         """Liefert das SentenceTransformer-Modell (lazy load)."""
         if self._model is None:
             self._model = SentenceTransformer(
-                self.model_name, trust_remote_code=True)
+                self.model_name,
+                trust_remote_code=True,
+            )
             self._model.max_seq_length = 512
         return self._model
 
@@ -167,7 +169,7 @@ class SentenceTransformerEmbeddings(BaseEmbeddings):
         passage_embeddings = self.model.encode(
             sentences=texts,
             task="retrieval",
-            prompt_name="passage",
+            prompt_name="document",
         )
         return [self._truncate([float(x) for x in emb]) for emb in passage_embeddings]
 

math-tutor/backend/app/deterministic_services/vector_store.py

@@ -216,6 +216,10 @@ def upsert_docs(pg_url: str, docs: List[DocRecord], embeddings: List[List[float]
 
     rows = []
     for doc, emb in zip(docs, embeddings):
+        if len(emb) != embedding_dim:
+            raise ValueError(
+                f"Embedding dimension {len(emb)} does not match configured target_dim {embedding_dim}"
+            )
         m = doc.metadata
         rows.append(
             {

math-tutor/backend/test/embeddings_jina_v5_unit_test.py

+from __future__ import annotations
+
+import importlib.util
+import os
+import sys
+from pathlib import Path
+import unittest
+from unittest.mock import MagicMock, patch
+
+
+BACKEND_ROOT = Path(__file__).resolve().parents[1]
+
+
+def _load_module(name: str, relative_path: str):
+    path = BACKEND_ROOT / relative_path
+    spec = importlib.util.spec_from_file_location(name, path)
+    if spec is None or spec.loader is None:
+        raise RuntimeError(f"Failed to load module spec for {path}")
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    return module
+
+
+config = _load_module("backend_config_test_module", "app/config.py")
+
+fake_sentence_transformers = type(sys)("sentence_transformers")
+fake_sentence_transformers.SentenceTransformer = object
+sys.modules.setdefault("sentence_transformers", fake_sentence_transformers)
+embeddings = _load_module("backend_embeddings_test_module", "app/deterministic_services/embeddings.py")
+
+
+class SentenceTransformerJinaV5Test(unittest.TestCase):
+    def test_config_defaults_to_jina_v5(self) -> None:
+        with patch.dict(os.environ, {"EMBEDDING_TYPE": "sentence-transformer"}, clear=False):
+            settings = config.get_embedding_settings()
+
+        self.assertEqual(settings.model, "jinaai/jina-embeddings-v5-text-small-retrieval")
+        self.assertEqual(settings.target_dim, 1024)
+
+    def test_embedder_uses_document_and_query_prompts(self) -> None:
+        fake_model = MagicMock()
+        fake_model.encode.side_effect = [
+            [[0.1, 0.2, 0.3, 0.4]],
+            [[0.4, 0.3, 0.2, 0.1]],
+        ]
+
+        with patch.object(embeddings, "SentenceTransformer", return_value=fake_model) as ctor:
+            embedder = embeddings.SentenceTransformerEmbeddings(
+                embeddings.SentenceTransformerConfig(
+                    model="jinaai/jina-embeddings-v5-text-small-retrieval",
+                    target_dim=4,
+                )
+            )
+
+            docs = embedder.embed_documents(["doc text"])
+            query = embedder.embed_query("query text")
+
+        ctor.assert_called_once()
+        self.assertEqual(fake_model.encode.call_args_list[0].kwargs["prompt_name"], "document")
+        self.assertEqual(fake_model.encode.call_args_list[1].kwargs["prompt_name"], "query")
+        self.assertEqual(len(docs[0]), 4)
+        self.assertEqual(len(query), 4)
+
+
+if __name__ == "__main__":
+    unittest.main()