fix: prune deleted BM25 vocabulary entries (#11997)

Author: hxaxd

haystack/document_stores/in_memory/document_store.py

@@ -507,6 +507,9 @@ def delete_documents(self, document_ids: list[str]) -> None:
             doc_len = doc_stats.doc_len
 
             self._freq_vocab_for_idf.subtract(Counter(freq.keys()))
+            for token in freq:
+                if self._freq_vocab_for_idf[token] <= 0:
+                    del self._freq_vocab_for_idf[token]
             try:
                 self._avg_doc_len = (self._avg_doc_len * (len(self._bm25_attr) + 1) - doc_len) / len(self._bm25_attr)
             except ZeroDivisionError:

releasenotes/notes/prune-zero-frequency-bm25-vocabulary-2669497a2350057b.yaml

@@ -0,0 +1,5 @@
+fixes:
+  - |
+    Remove zero-frequency vocabulary entries after deleting documents from
+    ``InMemoryDocumentStore`` so previous document history no longer changes
+    BM25Okapi scores for the active corpus.

test/document_stores/test_in_memory.py

@@ -763,6 +763,26 @@ def test_bm25_avg_doc_len_after_delete(self, in_memory_doc_store):
         # After removing "hello world" (2 tokens), only "foo bar baz" (3 tokens) remains
         assert in_memory_doc_store._avg_doc_len == pytest.approx(3.0)
 
+    def test_bm25_okapi_scores_do_not_depend_on_deleted_documents(self):
+        active_documents = [
+            Document(id="d1", content="common alpha"),
+            Document(id="d2", content="common beta"),
+            Document(id="d3", content="common gamma"),
+        ]
+        fresh_store = InMemoryDocumentStore(bm25_algorithm="BM25Okapi", shared=False)
+        reused_store = InMemoryDocumentStore(bm25_algorithm="BM25Okapi", shared=False)
+        deleted_document = Document(content="one two three four five six seven eight nine ten")
+
+        reused_store.write_documents([deleted_document])
+        reused_store.delete_documents([deleted_document.id])
+        fresh_store.write_documents(active_documents)
+        reused_store.write_documents(active_documents)
+
+        fresh_scores = {doc.id: doc.score for doc in fresh_store.bm25_retrieval(query="common", top_k=3)}
+        reused_scores = {doc.id: doc.score for doc in reused_store.bm25_retrieval(query="common", top_k=3)}
+
+        assert reused_scores == pytest.approx(fresh_scores)
+
 
 class TestMemoryDocumentStoreNotShared(TestMemoryDocumentStore):
     """