Initial commit: Tasteby - YouTube restaurant map service

Backend (FastAPI + Oracle ADB), Frontend (Next.js), daemon worker.
Features: channel/video/restaurant management, semantic search,
Google OAuth, user reviews.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

backend/core/vector.py

@@ -0,0 +1,97 @@
+"""Vector embedding generation and storage for restaurant semantic search."""
+
+from __future__ import annotations
+
+import array
+import os
+
+import oci
+from oci.generative_ai_inference import GenerativeAiInferenceClient
+from oci.generative_ai_inference.models import (
+    EmbedTextDetails,
+    OnDemandServingMode,
+)
+
+from core.db import conn
+
+
+def _embed_texts(texts: list[str]) -> list[list[float]]:
+    config = oci.config.from_file()
+    client = GenerativeAiInferenceClient(
+        config,
+        service_endpoint=os.environ["OCI_GENAI_ENDPOINT"],
+    )
+    model_id = os.environ.get("OCI_EMBED_MODEL_ID", "cohere.embed-v4.0")
+    compartment_id = os.environ["OCI_COMPARTMENT_ID"]
+
+    details = EmbedTextDetails(
+        inputs=texts,
+        serving_mode=OnDemandServingMode(model_id=model_id),
+        compartment_id=compartment_id,
+        input_type="SEARCH_DOCUMENT",
+    )
+    response = client.embed_text(details)
+    return response.data.embeddings
+
+
+def _to_vec(embedding: list[float]) -> array.array:
+    return array.array("f", embedding)
+
+
+def save_restaurant_vectors(restaurant_id: str, chunks: list[str]) -> list[str]:
+    """Embed and store text chunks for a restaurant.
+
+    Returns list of inserted row IDs.
+    """
+    if not chunks:
+        return []
+
+    embeddings = _embed_texts(chunks)
+    inserted: list[str] = []
+
+    sql = """
+        INSERT INTO restaurant_vectors (restaurant_id, chunk_text, embedding)
+        VALUES (:rid, :chunk, :emb)
+        RETURNING id INTO :out_id
+    """
+    import oracledb
+    with conn() as c:
+        cur = c.cursor()
+        for chunk, emb in zip(chunks, embeddings):
+            out_id = cur.var(oracledb.STRING)
+            cur.execute(sql, {
+                "rid": restaurant_id,
+                "chunk": chunk,
+                "emb": _to_vec(emb),
+                "out_id": out_id,
+            })
+            inserted.append(out_id.getvalue()[0])
+    return inserted
+
+
+def search_similar(query: str, top_k: int = 10) -> list[dict]:
+    """Semantic search: find restaurants similar to query text.
+
+    Returns list of dicts: restaurant_id, chunk_text, distance.
+    """
+    embeddings = _embed_texts([query])
+    query_vec = _to_vec(embeddings[0])
+
+    sql = """
+        SELECT rv.restaurant_id, rv.chunk_text,
+               VECTOR_DISTANCE(rv.embedding, :qvec, COSINE) AS dist
+        FROM restaurant_vectors rv
+        ORDER BY dist
+        FETCH FIRST :k ROWS ONLY
+    """
+    with conn() as c:
+        cur = c.cursor()
+        cur.execute(sql, {"qvec": query_vec, "k": top_k})
+        return [
+            {
+                "restaurant_id": r[0],
+                "chunk_text": r[1].read() if hasattr(r[1], "read") else r[1],
+                "distance": r[2],
+            }
+            for r in cur.fetchall()
+        ]