diff --git a/daras_ai_v2/vector_search.py b/daras_ai_v2/vector_search.py
index 87471948b..110f760fd 100644
--- a/daras_ai_v2/vector_search.py
+++ b/daras_ai_v2/vector_search.py
@@ -242,7 +242,7 @@ def get_top_k_references(
 
 def vespa_search_results_to_refs(
     search_result: dict,
-) -> typing.Iterable[SearchReference]:
+) -> typing.Iterable[tuple[str, SearchReference]]:
     for hit in search_result["root"].get("children", []):
         try:
             ref = EmbeddingsReference.objects.get(vespa_doc_id=hit["fields"]["id"])
@@ -250,7 +250,6 @@ def vespa_search_results_to_refs(
         except EmbeddingsReference.DoesNotExist:
             continue
         if "text/html" in ref.embedded_file.metadata.mime_type:
-            # logger.debug(f"Generating fragments {ref['url']} as it is a HTML file")
             ref.url = generate_text_fragment_url(url=ref.url, text=ref.snippet)
         yield (
             ref_key,
diff --git a/scripts/cleanup_vespa_db.py b/scripts/cleanup_vespa_db.py
new file mode 100644
index 000000000..5536626da
--- /dev/null
+++ b/scripts/cleanup_vespa_db.py
@@ -0,0 +1,66 @@
+import typing
+from datetime import timedelta
+
+from django.utils import timezone
+
+from daras_ai_v2 import settings
+from daras_ai_v2.vector_search import get_vespa_app
+from embeddings.models import EmbeddedFile
+
+if typing.TYPE_CHECKING:
+    from vespa.io import VespaResponse
+
+
+STALENESS_THRESHOLD_DAYS = 90
+BATCH_SIZE = 1_000
+
+
+def cleanup_stale_cache():
+    vespa = get_vespa_app()
+
+    while True:
+        stale_qs = (
+            EmbeddedFile.objects.prefetch_related("embeddings_references")
+            .filter(
+                updated_at__lt=timezone.now() - timedelta(days=STALENESS_THRESHOLD_DAYS)
+            )
+            .order_by("updated_at")[:BATCH_SIZE]
+        )
+        stale_files = list(stale_qs)
+        if not stale_files:
+            break
+
+        docs_to_delete = (
+            {"id": ref.vespa_doc_id}
+            for ef in stale_files
+            for ref in ef.embeddings_references.all()
+        )
+        total_deleted = 0
+
+        def vespa_callback(response: "VespaResponse", id: str):
+            nonlocal total_deleted
+            if response.is_successful():
+                total_deleted += 1
+            else:
+                print(f"""\
+                    Failed to delete document {id} from Vespa.
+                    Status code: {response.get_status_code()}
+                    JSON: {response.get_json()}
+                    """)
+
+        vespa.feed_iterable(
+            docs_to_delete,
+            schema=settings.VESPA_SCHEMA,
+            operation_type="delete",
+            callback=vespa_callback,
+        )
+        print(f"Deleted {total_deleted} documents from Vespa.")
+
+        deleted_per_model = EmbeddedFile.objects.filter(
+            id__in=[ef.id for ef in stale_files]
+        ).delete()
+        print(f"Deleted EmbeddedFiles & related objects: {deleted_per_model}")
+
+
+def run():
+    cleanup_stale_cache()