benchmarks: fix and extend HNSW+DiskANN recompute vs no-recompute; docs: add fresh numbers and DiskANN notes

benchmarks/benchmark_no_recompute.py

@@ -84,6 +84,80 @@ def main():
     )
     print("Expectation: no-recompute should be faster but larger on disk.")
 
+    # DiskANN quick benchmark (final rerank vs no-recompute)
+    try:
+        index_path_diskann_nr = str(base / "diskann_nr.leann")
+        index_path_diskann_r = str(base / "diskann_r.leann")
+
+        # Build DiskANN no-recompute (keeps full disk index)
+        if not (
+            Path(index_path_diskann_nr).parent / f"{Path(index_path_diskann_nr).stem}.meta.json"
+        ).exists():
+            b = LeannBuilder(
+                backend_name="diskann",
+                embedding_model=os.getenv("LEANN_EMBED_MODEL", "facebook/contriever"),
+                embedding_mode=os.getenv("LEANN_EMBED_MODE", "sentence-transformers"),
+                graph_degree=32,
+                complexity=64,
+                num_threads=4,
+                is_recompute=False,
+            )
+            for i in range(5000):
+                b.add_text(f"DiskANN NR test doc {i} for quick benchmark.")
+            b.build_index(index_path_diskann_nr)
+
+        # Build DiskANN recompute (enables partition; prunes redundant files)
+        if not (
+            Path(index_path_diskann_r).parent / f"{Path(index_path_diskann_r).stem}.meta.json"
+        ).exists():
+            b = LeannBuilder(
+                backend_name="diskann",
+                embedding_model=os.getenv("LEANN_EMBED_MODEL", "facebook/contriever"),
+                embedding_mode=os.getenv("LEANN_EMBED_MODE", "sentence-transformers"),
+                graph_degree=32,
+                complexity=64,
+                num_threads=4,
+                is_recompute=True,
+            )
+            for i in range(5000):
+                b.add_text(f"DiskANN R test doc {i} for quick benchmark.")
+            b.build_index(index_path_diskann_r)
+
+        # Measure size per build prefix
+        def _size_for(prefix: str) -> int:
+            p = Path(prefix)
+            base_dir = p.parent
+            stem = p.stem
+            total = 0
+            for f in base_dir.iterdir():
+                if f.is_file() and f.name.startswith(stem):
+                    total += f.stat().st_size
+            return total
+
+        size_diskann_nr = _size_for(index_path_diskann_nr)
+        size_diskann_r = _size_for(index_path_diskann_r)
+
+        # Speed on recompute-build (final rerank vs no-recompute)
+        s = LeannSearcher(index_path_diskann_r)
+        _ = s.search("DiskANN R test doc 123", top_k=10, complexity=64, recompute_embeddings=False)
+        _ = s.search("DiskANN R test doc 123", top_k=10, complexity=64, recompute_embeddings=True)
+
+        t0 = time.time()
+        _ = s.search("DiskANN R test doc 123", top_k=10, complexity=64, recompute_embeddings=False)
+        t_diskann_nr = time.time() - t0
+
+        t0 = time.time()
+        _ = s.search("DiskANN R test doc 123", top_k=10, complexity=64, recompute_embeddings=True)
+        t_diskann_r = time.time() - t0
+
+        print("\nBenchmark results (DiskANN):")
+        print(f"  build(recompute=False): size={size_diskann_nr / 1024 / 1024:.1f}MB")
+        print(f"  build(recompute=True, partition): size={size_diskann_r / 1024 / 1024:.1f}MB")
+        print(f"  search recompute=False: {t_diskann_nr:.3f}s (on recompute-build)")
+        print(f"  search recompute=True (final rerank): {t_diskann_r:.3f}s (on recompute-build)")
+    except Exception as e:
+        print(f"DiskANN quick benchmark skipped due to: {e}")
+
 
 if __name__ == "__main__":
     main()

docs/configuration-guide.md

@@ -363,12 +363,23 @@ Trade-offs:
 Real-world quick benchmark (HNSW, 5k texts; script `benchmarks/benchmark_no_recompute.py`):
 
 ```text
-recompute=True:  ~6.58s; size ~1.1MB
-recompute=False: ~0.10s; size ~16.6MB
+recompute=True:  ~7.55s; size ~1.1MB
+recompute=False: ~0.11s; size ~16.6MB
 
 Conclusion: no-recompute is much faster but uses more storage; recompute is smaller but has higher first-hop latency.
 ```
 
+DiskANN (5k texts; same script, final rerank strategy):
+
+```text
+build(recompute=False): size ~24.8MB
+build(recompute=True, partition): size ~5.7MB
+search recompute=False: ~0.250s (on recompute-build)
+search recompute=True (final rerank): ~0.120s (on recompute-build)
+
+Conclusion: DiskANN's recompute-build enables partitioning to reduce storage; enabling final rerank further improves latency while keeping traversal PQ-fast.
+```
+
 
 ## Further Reading
 

packages/leann-backend-diskann/leann_backend_diskann/diskann_backend.py

@@ -442,8 +442,14 @@ class DiskannSearcher(BaseSearcher):
             use_global_pruning = True
 
         # Perform search with suppressed C++ output based on log level
-        use_deferred_fetch = kwargs.get("USE_DEFERRED_FETCH", True)
+        # Strategy:
+        # - Traversal always uses PQ distances
+        # - If recompute_embeddings=True, do a single final rerank via deferred fetch
+        #   (fetch embeddings for the final candidate set only)
+        # - Do not recompute neighbor distances along the path
+        use_deferred_fetch = True if recompute_embeddings else False
         recompute_neighors = False
+
         with suppress_cpp_output_if_needed():
             labels, distances = self._index.batch_search(
                 query,

packages/leann-core/src/leann/chat.py

@@ -422,7 +422,6 @@ class LLMInterface(ABC):
                 top_k=10,
                 complexity=64,
                 beam_width=8,
-                USE_DEFERRED_FETCH=True,
                 skip_search_reorder=True,
                 recompute_beighbor_embeddings=True,
                 dedup_node_dis=True,
@@ -434,7 +433,6 @@ class LLMInterface(ABC):
         Supported kwargs:
             - complexity (int): Search complexity parameter (default: 32)
             - beam_width (int): Beam width for search (default: 4)
-            - USE_DEFERRED_FETCH (bool): Enable deferred fetch mode (default: False)
             - skip_search_reorder (bool): Skip search reorder step (default: False)
             - recompute_beighbor_embeddings (bool): Enable ZMQ embedding server for neighbor recomputation (default: False)
             - dedup_node_dis (bool): Deduplicate nodes by distance (default: False)