Separate fetching from ranking in search

`search` fetched, reranked and truncated in one pass, with the reranker's
over-fetch and the reranking itself interleaved in the same branch. Searching
several databases needs to fuse their candidates before anything is ranked, so
the phases have to be separable.

`_fetch` returns one database's candidates, over-fetching only when a reranker
will re-order them. `_rank` orders and cuts them, leaving an image query's vector
ranking alone since there is no text for a reranker to score against. The
over-fetch multiplier is named rather than a literal 10 at the point of use.

Both check the query type before reading `client.reranker`, which is a
cached_property that builds the reranker on first access and loads model weights
for a local one. An image query never used it and must not start.

No behaviour change: the same suite passes, and the search outputs digest
identically to before.

haiku_rag_slim/haiku/rag/client/search.py

@@ -38,40 +38,8 @@ async def search(
     if limit is None:
         limit = client._config.search.limit
 
-    if isinstance(query, str):
-        if search_type is None:
-            search_type = "hybrid"
-
-        reranker = client.reranker
-
-        if reranker is None:
-            chunk_results = await client.chunk_repository.search(
-                query, limit, search_type, filter
-            )
-        else:
-            search_limit = limit * 10
-            raw_results = await client.chunk_repository.search(
-                query, search_limit, search_type, filter
-            )
-            chunks = [chunk for chunk, _ in raw_results]
-            if client._config.reranking.multimodal:
-                await _attach_picture_data(client, chunks)
-            chunk_results = await reranker.rerank(query, chunks, top_n=limit)
-    else:
-        embedder = client.embedder
-        if not embedder.supports_images:
-            raise ValueError(
-                "Image queries require a multimodal embedder. Set "
-                "embeddings.model.multimodal: true on a vllm, voyageai, or cohere "
-                "model."
-            )
-        query_vector = await embedder.embed_image(query)
-        chunk_results = await client.chunk_repository.search(
-            query="",
-            limit=limit,
-            filter=filter,
-            query_vector=query_vector,
-        )
+    candidates = await _fetch(client, query, limit, search_type, filter)
+    chunk_results = await _rank(client, query, candidates, limit)
 
     results = [SearchResult.from_chunk(chunk, score) for chunk, score in chunk_results]
     results = _dedup_picture_chunks(results)
@@ -82,6 +50,73 @@ async def search(
     return results
 
 
+# Candidates per requested result when a reranker will re-order them.
+_RERANK_OVERFETCH = 10
+
+
+async def _fetch(
+    client: "HaikuRAG",
+    query: "str | bytes | PILImage.Image",
+    limit: int,
+    search_type: SearchType | None,
+    filter: str | None,
+) -> list[tuple[Chunk, float]]:
+    """Candidates from one database, ranked by that database.
+
+    Over-fetches when a reranker will re-order them. Separate from `_rank` so a
+    caller searching several databases can fuse their candidates before anything
+    is ranked or enriched.
+    """
+    if isinstance(query, str):
+        if search_type is None:
+            search_type = "hybrid"
+        fetch_limit = limit * _RERANK_OVERFETCH if client.reranker else limit
+        return await client.chunk_repository.search(
+            query, fetch_limit, search_type, filter
+        )
+
+    embedder = client.embedder
+    if not embedder.supports_images:
+        raise ValueError(
+            "Image queries require a multimodal embedder. Set "
+            "embeddings.model.multimodal: true on a vllm, voyageai, or cohere "
+            "model."
+        )
+    query_vector = await embedder.embed_image(query)
+    return await client.chunk_repository.search(
+        query="",
+        limit=limit,
+        filter=filter,
+        query_vector=query_vector,
+    )
+
+
+async def _rank(
+    client: "HaikuRAG",
+    query: "str | bytes | PILImage.Image",
+    candidates: list[tuple[Chunk, float]],
+    limit: int,
+) -> list[tuple[Chunk, float]]:
+    """Order candidates and cut them to `limit`.
+
+    An image query carries no text for a reranker to score against, so its
+    candidates keep the vector ranking. Its type is checked before
+    `client.reranker`, which builds the reranker on first access and loads model
+    weights for a local one.
+    """
+    if not isinstance(query, str):
+        return candidates[:limit]
+
+    reranker = client.reranker
+    if reranker is None:
+        return candidates[:limit]
+
+    chunks = [chunk for chunk, _ in candidates]
+    if client._config.reranking.multimodal:
+        await _attach_picture_data(client, chunks)
+    return await reranker.rerank(query, chunks, top_n=limit)
+
+
 async def _attach_picture_data(client: "HaikuRAG", chunks: list[Chunk]) -> None:
     """Attach picture bytes to synthetic picture chunks in-place, so a
     multimodal reranker can score the pixels instead of just the chunk's

tests/test_search_phases.py

@@ -0,0 +1,60 @@
+import pytest
+
+from haiku.rag.client import HaikuRAG
+from haiku.rag.client.search import _fetch, _rank
+from haiku.rag.store.models import Chunk
+
+
+@pytest.fixture
+def exploding_reranker(monkeypatch):
+    """A reranker that cannot be built, so any access fails loudly."""
+
+    def boom(self):
+        raise AssertionError("reranker built for a query that cannot use it")
+
+    monkeypatch.setattr(HaikuRAG, "reranker", property(boom), raising=True)
+
+
+@pytest.mark.asyncio
+async def test_image_ranking_never_builds_the_reranker(
+    temp_db_path, exploding_reranker
+):
+    """Local rerankers load model weights on construction, so an image query,
+    which has no text to score against, must not touch one."""
+    async with HaikuRAG(temp_db_path, create=True) as rag:
+        candidates = [
+            (Chunk(id="a", document_id="d", content="one"), 0.9),
+            (Chunk(id="b", document_id="d", content="two"), 0.8),
+        ]
+
+        ranked = await _rank(rag, b"image-bytes", candidates, limit=1)
+
+        assert [c.id for c, _ in ranked] == ["a"]
+
+
+@pytest.mark.asyncio
+async def test_image_fetch_never_builds_the_reranker(
+    temp_db_path, exploding_reranker, monkeypatch
+):
+    async with HaikuRAG(temp_db_path, create=True) as rag:
+        seen = {}
+
+        async def fake_search(
+            query, limit, search_type="hybrid", filter=None, query_vector=None
+        ):
+            seen["limit"] = limit
+            return []
+
+        async def fake_embed_image(self, image):
+            return [0.1] * 8
+
+        monkeypatch.setattr(rag.chunk_repository, "search", fake_search)
+        monkeypatch.setattr(type(rag.embedder), "embed_image", fake_embed_image)
+        monkeypatch.setattr(
+            type(rag.embedder), "supports_images", property(lambda self: True)
+        )
+
+        await _fetch(rag, b"image-bytes", 5, None, None)
+
+        # No over-fetch: nothing will re-rank these.
+        assert seen["limit"] == 5