MoSchmidt
diff --git a/‎backend/app/core/database.py‎
Lines changed: 8 additions & 6 deletions b/‎backend/app/core/database.py‎
Lines changed: 8 additions & 6 deletions
diff --git a/‎backend/app/repositories/search_repository.py‎
Lines changed: 25 additions & 10 deletions b/‎backend/app/repositories/search_repository.py‎
Lines changed: 25 additions & 10 deletions
diff --git a/‎backend/requirements.txt‎
Lines changed: 2 additions & 1 deletion b/‎backend/requirements.txt‎
Lines changed: 2 additions & 1 deletion
@@ -4,7 +4,7 @@
 from importlib import import_module
 from typing import AsyncGenerator
 
-from sqlalchemy.ext.asyncio import AsyncSession, create_async_engine, async_sessionmaker
+from sqlalchemy.ext.asyncio import AsyncSession, async_sessionmaker, create_async_engine
 from sqlalchemy.orm import declarative_base
 
 from app.core.config import settings
@@ -14,10 +14,12 @@
 Base = declarative_base()
 
 # Use async engine with asyncpg or aiomysql driver
+# Set probes for the vector index for every connection
 engine = create_async_engine(
     settings.DATABASE_URL,
     echo=(settings.ENVIRONMENT == "dev"),
     future=True,
+    connect_args={"server_settings": {"ivfflat.probes": "50"}},
 )
 
 async_session_local = async_sessionmaker(
@@ -32,11 +34,11 @@
 async def init_db() -> None:
     """Automatically create or update tables based on SQLAlchemy models."""
     for module in (
-            "app.models.user",
-            "app.models.project",
-            "app.models.paper",
-            "app.models.project_paper",
-            "app.models.paper_content",
+        "app.models.user",
+        "app.models.project",
+        "app.models.paper",
+        "app.models.project_paper",
+        "app.models.paper_content",
     ):
         import_module(module)
 
 
@@ -1,6 +1,7 @@
 from datetime import date
 from typing import List, Optional, Tuple, Union
 
+import numpy as np
 from sqlalchemy import ColumnElement, and_, or_, select
 from sqlalchemy.ext.asyncio import AsyncSession
 
@@ -34,26 +35,40 @@ async def search_papers_by_embeddings(
         Returns a list of (PaperModel, avg_distance) tuples ordered by ascending distance.
         Optionally applies advanced search filters (year range, text conditions).
         """
+        # Compute centroid of all extracted search queries, so we can use the IVF index for
+        # efficient vector searches
+        vecs = np.array(embeddings, dtype=np.float32)
+        q = vecs.mean(axis=0)
+        q /= np.linalg.norm(q) + 1e-12
+        q_list = q.tolist()
+
+        cand_stmt = (
+            select(PaperModel.paper_id)
+            .order_by(PaperModel.embedding.cosine_distance(q_list))
+            .limit(500)
+        )
+
+        if search_filter:
+            clauses = SearchRepository._build_filter_clauses(search_filter)
+            if clauses:
+                cand_stmt = cand_stmt.where(and_(*clauses))
 
-        # Build distance expressions
+        cand_ids = [r[0] for r in (await db.execute(cand_stmt)).all()]
+
+        # Rerank the retrieved results based on the avg. distance to all search queries
+        # This drastically improves the quality of the responses with negligible more runtime
         distance_exprs = [PaperModel.embedding.cosine_distance(emb) for emb in embeddings]
-        total_distance = sum(distance_exprs)
-        avg_distance = total_distance / len(embeddings)
+        avg_distance = sum(distance_exprs) / len(distance_exprs)
 
         stmt = (
             select(PaperModel, avg_distance.label("avg_distance"))
-            .where(avg_distance < threshold)
+            .where(avg_distance < threshold, PaperModel.paper_id.in_(cand_ids))
             .order_by(avg_distance.asc())
             .limit(limit)
         )
 
-        if search_filter:
-            clauses = SearchRepository._build_filter_clauses(search_filter)
-            if clauses:
-                stmt = stmt.where(and_(*clauses))
+        rows = (await db.execute(stmt)).all()
 
-        result = await db.execute(stmt)
-        rows = result.fetchall()
         return [(paper, float(dist)) for paper, dist in rows]
 
     @staticmethod
 
@@ -34,9 +34,10 @@ transformers==4.51.3
 tokenizers==0.21.4
 adapters==1.2.0
 
+numpy==1.26.4
 pandas>=2.2.0
 pyarrow>=22.0.0
 pypdf>=4.0.0
 
 # --- PDF to Markdown Conversion ---
-docling==2.67.0
+docling==2.67.0