diff --git a/api/requirements.txt b/api/requirements.txt
index a63ff42..f38813f 100644
--- a/api/requirements.txt
+++ b/api/requirements.txt
@@ -10,6 +10,5 @@ sentry-sdk[fastapi]>=2.20.0
 slowapi>=0.1.9
 tiktoken>=0.7.0
 faiss-cpu>=1.8.0
-rank-bm25>=0.2.2
 filelock>=3.12.0
 asyncpg>=0.29.0
diff --git a/pyproject.toml b/pyproject.toml
index d7de1f2..632be9c 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -22,7 +22,6 @@ dependencies = [
     "slowapi>=0.1.9",
     "tiktoken>=0.7.0",
     "faiss-cpu>=1.8.0",
-    "rank-bm25>=0.2.2",
     "filelock>=3.12.0",
     "asyncpg>=0.29.0"
 ]
@@ -34,7 +33,8 @@ dev = [
     "pytest-mock>=3.12.0",
     "coverage>=7.6.0",
     "pytest-cov>=5.0.0",
-    "pandas>=2.0.0"
+    "pandas>=2.0.0",
+    "bm25s>=0.3.0",
 ]
 
 [tool.pytest.ini_options]
diff --git a/scripts/benchmark_beir.py b/scripts/benchmark_beir.py
index c04add8..3ebd0ed 100644
--- a/scripts/benchmark_beir.py
+++ b/scripts/benchmark_beir.py
@@ -180,12 +180,17 @@ def compute_metrics(
 
 
 def build_lexical_index(corpus: dict[str, str]):
-    """Build lightweight token-based BM25 index."""
-    from rank_bm25 import BM25Okapi
+    """Build lightweight token-based BM25 index.
+
+    Tokenization is intentionally plain lowercase whitespace splitting so
+    lexical scores stay comparable across runs (no stemming).
+    """
+    import bm25s
 
     doc_ids = list(corpus.keys())
     tokenized_corpus = [corpus[did].lower().split() for did in doc_ids]
-    bm25 = BM25Okapi(tokenized_corpus)
+    bm25 = bm25s.BM25()
+    bm25.index(tokenized_corpus)
     return bm25, doc_ids