Merge main into gatekeeper/fix-2325-1

docs/src/python/python.md

@@ -200,6 +200,33 @@ listing a storage directory.
 
 ::: lancedb.materialized_view.MaterializedViewDefinition
 
+::: lancedb.vector_dedup
+
+::: lancedb.materialized_view.VectorDedupSource
+
+For an existing indexed source, declare and materialize a dedup result without
+writing SQL. Creation captures the source version once and waits for the complete
+result; remote connections use the service's MV jobs.
+
+```python
+from lancedb import vector_dedup
+
+view = db.create_materialized_view(
+    "images_clean",
+    vector_dedup("images", column="phash", distance_threshold=4),
+)
+cleaned = view.table
+```
+
+To submit without waiting, call `db.create_materialized_view_async(...)` with the
+same source expression and wait on the returned job. To declare only, pass
+`with_no_data=True`, then use `view.refresh_async().wait()` when ready. Pass
+`dataset_version=...` to `vector_dedup` for an explicit snapshot. The default
+policy retains direct representatives (A-B-C without A-C keeps A and C), preserves
+the source, and materializes its original columns. Only within-segment/partition
+pairs are covered, using the index's distance representation.
+
+
 ## Views
 
 ::: lancedb.view.ViewDescription
@@ -381,6 +408,29 @@ still work. Queries return descriptors. Call
 
 ## Reranking
 
+`TypeSafeReranker` supports opt-in request batching:
+
+```python
+from lancedb.rerankers import TypeSafeReranker
+
+reranker = TypeSafeReranker(batch_size=40, max_concurrency=8)
+```
+
+The default `batch_size=1` keeps the query and document in request state and
+sends one request per non-null candidate. With `batch_size=40`, 80 non-null
+candidates require two requests. `max_concurrency` still limits simultaneous
+requests, and the SDK handles retries.
+
+In batched mode, state contains only `{"query": query}`. Each independent
+question contains `{"question": instructions, "document": document}` in its
+structured instructions, so it sees only its own document and the shared query.
+Custom instructions and criteria are kept verbatim: adapt prompts that explicitly
+reference request-state fields such as `state.document` before enabling batching.
+Null documents retain a zero score without an API call; empty strings are scored.
+API failures, mismatched answer IDs, and invalid probabilities raise errors.
+Batching changes the payload and can affect model scores; compare quality and
+latency on your workload before opting in.
+
 ::: lancedb.rerankers
     options:
       show_root_heading: false