Honor configured vector metric during search

CHANGELOG.md

@@ -6,6 +6,9 @@
 
 - Migration to 0.38.0 no longer fails with `UnicodeDecodeError` on a
   `docling_document` blob written as zstd.
+- Vector search explicitly uses `search.vector_index_metric`, so flat and indexed
+  searches rank with the same configured distance metric. The unused `dot` option
+  is removed; `cosine` and `l2` remain.
 
 ## [0.79.0] - 2026-08-28
 

docs/cli.md

@@ -391,7 +391,7 @@ haiku-rag create-index [--db /path/to/your.lancedb]
 
 **Requirements:**
 - Minimum 256 chunks required for index creation (LanceDB training data requirement)
-- Creates an IVF_PQ index using the configured `search.vector_index_metric` (cosine/l2/dot)
+- Creates an IVF_PQ index using the configured `search.vector_index_metric` (cosine/l2)
 
 **When to use:**
 - After ingesting documents (indexes are not created automatically)

docs/configuration/index.md

@@ -114,7 +114,7 @@ qa:
 search:
   limit: 5                     # Default number of results to return
   max_context_chars: 5000     # Maximum characters in expanded context
-  vector_index_metric: cosine  # cosine, l2, or dot
+  vector_index_metric: cosine  # cosine or l2
   vector_refine_factor: 30
 
 doctor:

docs/configuration/storage.md

@@ -300,7 +300,7 @@ Configure vector search settings:
 
 ```yaml
 search:
-  vector_index_metric: cosine  # cosine, l2, or dot
+  vector_index_metric: cosine  # cosine or l2
   vector_refine_factor: 30     # Re-ranking factor for accuracy
 ```
 
@@ -309,7 +309,6 @@ For search behavior settings (`limit`, `max_context_chars`), see [Search and Que
 - **vector_index_metric**: Distance metric for vector similarity:
   - `cosine`: Cosine similarity (default, best for most embeddings)
   - `l2`: Euclidean distance
-  - `dot`: Dot product similarity
 - **vector_refine_factor**: Improves accuracy when using a vector index by retrieving `refine_factor * limit` candidates (using approximate search) and re-ranking them with exact distances. Higher values increase accuracy but slow down queries. Default: 30
   - **Only applies with a vector index** - has no effect on brute-force search, which already returns exact results
 

haiku_rag_slim/haiku/rag/config/models.py

@@ -304,7 +304,7 @@ class ProcessingConfig(ConfigModel):
 class SearchConfig(ConfigModel):
     limit: int = Field(default=5, gt=0)
     max_context_chars: int = Field(default=5000, gt=0)
-    vector_index_metric: Literal["cosine", "l2", "dot"] = "cosine"
+    vector_index_metric: Literal["cosine", "l2"] = "cosine"
     vector_refine_factor: int = Field(default=30, gt=0)
 
 

haiku_rag_slim/haiku/rag/store/repositories/chunk.py

@@ -254,6 +254,7 @@ class ChunkRepository:
                 self.store.chunks_table.query()
                 .nearest_to(query_embedding)
                 .column("vector")
+                .distance_type(self.store._config.search.vector_index_metric)
                 .refine_factor(self.store._config.search.vector_refine_factor)
             )
             # An image query has no text to match, so it stays vector-only.

tests/test_chunk.py

@@ -570,6 +570,82 @@ async def test_chunk_search_with_precomputed_vector_skips_text_query(temp_db_pat
         assert any(c.document_id == doc.id for c, _ in results)
 
 
+@pytest.mark.parametrize(
+    "metric,expected_ids",
+    [
+        ("cosine", ["cosine-best", "l2-best"]),
+        ("l2", ["l2-best", "cosine-best"]),
+    ],
+)
+async def test_vector_metric_ranking_matches_before_and_after_index(
+    temp_db_path, metric, expected_ids
+):
+    """Flat and indexed searches honor the configured metric.
+
+    The two leading vectors are deliberately not unit-normalized: cosine and
+    L2 must rank them in opposite orders, so LanceDB's flat-search L2 default
+    cannot accidentally satisfy the cosine case.
+    """
+    from datetime import timedelta
+
+    from lancedb.index import IvfPq
+
+    config = get_config().model_copy(deep=True)
+    config.embeddings.model.vector_dim = 8
+    config.search.vector_index_metric = metric
+
+    def row(chunk_id, vector, order):
+        return {
+            "id": chunk_id,
+            "document_id": "doc-1",
+            "content": chunk_id,
+            "content_fts": chunk_id,
+            "metadata": "{}",
+            "order": order,
+            "vector": vector,
+        }
+
+    rows = [
+        row("cosine-best", [10.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], 0),
+        row("l2-best", [1.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], 1),
+    ]
+    # IVF-PQ needs 256 training rows. These are far from the query under both
+    # metrics and vary enough to train the quantizer without entering Top-K.
+    rows.extend(
+        row(
+            f"filler-{i}",
+            [0.0, *[100.0 + i + j for j in range(7)]],
+            i + 2,
+        )
+        for i in range(254)
+    )
+
+    async with HaikuRAG(temp_db_path, config=config, create=True) as client:
+        await client.store.chunks_table.add(rows)
+        query_vector = [1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0]
+
+        async def top_ids():
+            results = await client.chunk_repository.search(
+                "", limit=2, search_type="vector", query_vector=query_vector
+            )
+            return [chunk.id for chunk, _score in results]
+
+        flat_ids = await top_ids()
+
+        await client.store.chunks_table.create_index(
+            "vector",
+            config=IvfPq(distance_type=metric, num_partitions=1, num_sub_vectors=1),
+            replace=True,
+        )
+        await client.store.chunks_table.wait_for_index(
+            ["vector_idx"], timeout=timedelta(minutes=1)
+        )
+        indexed_ids = await top_ids()
+
+    assert flat_ids == expected_ids
+    assert indexed_ids == expected_ids
+
+
 async def test_get_chunk_ids_by_self_ref_grouped_without_documents(temp_db_path):
     async with HaikuRAG(
         db_path=temp_db_path, config=get_config(), create=True

tests/test_config.py

@@ -670,6 +670,7 @@ def test_unknown_keys_are_rejected(data, bad_key):
         {"processing": {"converter": "docling-loca"}},
         {"processing": {"chunker": "docling-remote"}},
         {"processing": {"chunker_type": "semantic"}},
+        {"search": {"vector_index_metric": "dot"}},
     ],
 )
 def test_finite_switches_reject_unknown_values(data):