docs: Update README and plan to clarify effects of overquery_factor on recall and latency

Author: tae898

bindings/python/examples/benchmark-vector/README.md

@@ -13,7 +13,8 @@
 
 ### 17 January 2026 Update
 
-- Ran the experiments with varying `overquery_factor` again to see how it affects recall and latency. Increasing `overquery_factor` should improve recall at the cost of latency. as it increases the number of candidates considered during search (more efforts on search)
+- Ran the experiments with varying `overquery_factor` again to see how it affects recall and latency.
+- Increasing `overquery_factor` should improve recall at the cost of latency, as it increases the number of candidates considered during search (more efforts on search)
 
 #### MSMARCO-1M (1000 queries, Recall@50)
 

plan_update_python_bindings.md

@@ -68,5 +68,5 @@ The vector benchmark scripts and other examples have been updated.
         pq_center_globally: Optional[bool] = None,
         pq_training_limit: Optional[int] = None,
 - [x] Tests: add PQ/approx search coverage (params pass-through, k results, overquery handling, persistence after reopen, invalid PQ config errors).
-- [ ] Benchmarks/examples: extend vector examples and MSMARCO bench scripts to include PQ/approx runs and note recall trade-offs for random vs real embeddings.
 - [ ] Docs: update Python examples and mkdocs, once all the vector stuff is done.
+- [ ] Run the vector index & search on 10M and 20M datasets to see if it's fesaible with 16G and 32G heap budgets, respectively. (Wikipedia has over 7.1 million articles, as of 17 January 2026, and I wish I can index wikipedia with the same number of vectors.)