Merge pull request #387 from ggozad/feat/cache_reranker

Cache reranker on the client instead of rebuilding per search

CHANGELOG.md

@@ -8,6 +8,7 @@
 ### Changed
 
 - A successful DELETE job auto-prunes dead jobs with the same `(source_id, uri)`. New `JobRepo.prune_dead(source_id, uri)`.
+- Reranker built once per `HaikuRAG` client instead of per `search()` call.
 
 ## [0.50.0] - 2026-05-27
 

haiku_rag_slim/haiku/rag/client/__init__.py

@@ -7,6 +7,7 @@ import tempfile
 from collections.abc import AsyncGenerator
 from datetime import datetime
 from enum import Enum
+from functools import cached_property
 from pathlib import Path
 from typing import TYPE_CHECKING, overload
 from urllib.parse import urlparse
@@ -31,6 +32,7 @@ if TYPE_CHECKING:
     from PIL import Image as PILImage
 
     from haiku.rag.ingester.sources.base import Source
+    from haiku.rag.reranking.base import RerankerBase
     from haiku.rag.sandbox import AnalysisResult
     from haiku.rag.store.models.citation import Citation
 
@@ -87,6 +89,15 @@ class HaikuRAG:
         """Whether the client is in read-only mode."""
         return self.store.is_read_only
 
+    @cached_property
+    def reranker(self) -> "RerankerBase | None":
+        """The configured reranker, built once and reused across searches.
+
+        None when reranking is disabled. Local rerankers load model weights on
+        construction, so building per search would reload them on every query.
+        """
+        return get_reranker(config=self._config)
+
     async def __aenter__(self):
         """Async context manager entry — initializes store and repositories."""
         self.store = Store(

haiku_rag_slim/haiku/rag/client/search.py

@@ -1,7 +1,6 @@
 import base64
 from typing import TYPE_CHECKING
 
-from haiku.rag.reranking import get_reranker
 from haiku.rag.store.models.chunk import Chunk, SearchResult, SearchType
 from haiku.rag.store.models.document_item import PICTURE_REF_PREFIX
 
@@ -42,7 +41,7 @@ async def search(
         if search_type is None:
             search_type = "hybrid"
 
-        reranker = get_reranker(config=client._config)
+        reranker = client.reranker
 
         if reranker is None:
             chunk_results = await client.chunk_repository.search(

tests/test_search.py

@@ -395,6 +395,37 @@ async def test_search_with_bytes_query_uses_multimodal_embedder(
     assert received_kwargs["query"] == ""
 
 
+@pytest.mark.asyncio
+async def test_reranker_built_once_across_searches(temp_db_path, monkeypatch):
+    """The reranker is constructed once per client and reused across searches,
+    rather than rebuilt (reloading model weights) on every query."""
+    from haiku.rag.store.models.chunk import Chunk
+
+    build_count = 0
+
+    class StubReranker:
+        async def rerank(self, query, chunks, top_n):
+            return [(chunk, 1.0) for chunk in chunks][:top_n]
+
+    def fake_get_reranker(config):
+        nonlocal build_count
+        build_count += 1
+        return StubReranker()
+
+    monkeypatch.setattr("haiku.rag.client.get_reranker", fake_get_reranker)
+
+    async def fake_chunk_search(query, limit, search_type, filter):
+        return [(Chunk(content="x", metadata={}), 0.5)]
+
+    async with HaikuRAG(temp_db_path, create=True) as rag:
+        rag.chunk_repository.search = fake_chunk_search  # type: ignore[method-assign]
+        await rag.search("first", include_images=False)
+        await rag.search("second", include_images=False)
+        await rag.search("third", include_images=False)
+
+    assert build_count == 1
+
+
 @pytest.mark.asyncio
 async def test_search_with_pil_image_works_like_bytes(temp_db_path, monkeypatch):
     from PIL import Image as PILImageModule