diff --git a/backend/app/services/collections/providers/openai.py b/backend/app/services/collections/providers/openai.py
index 770e0773f..84993268c 100644
--- a/backend/app/services/collections/providers/openai.py
+++ b/backend/app/services/collections/providers/openai.py
@@ -47,8 +47,14 @@ def upload_files(
                     if doc.file_size_kb is None:
                         doc.file_size_kb = round(tmp.tell() / 1024, 2)
                     tmp.seek(0)
+                    if "." in doc.fname:
+                        filename, extension = doc.fname.rsplit(".", 1)
+                        normalized_fname = f"{filename}.{extension.lower()}"
+                    else:
+                        normalized_fname = doc.fname
                     uploaded = self.client.files.create(
-                        file=(doc.fname, tmp), purpose="assistants"
+                        file=(normalized_fname, tmp),
+                        purpose="assistants",
                     )
             except Exception as err:
                 logger.error(
diff --git a/backend/app/tests/services/collections/providers/test_openai_provider.py b/backend/app/tests/services/collections/providers/test_openai_provider.py
index d31fb0ae3..8d8cf42c0 100644
--- a/backend/app/tests/services/collections/providers/test_openai_provider.py
+++ b/backend/app/tests/services/collections/providers/test_openai_provider.py
@@ -296,15 +296,25 @@ def test_upload_files_empty_docs_is_noop() -> None:
     client.files.create.assert_not_called()
 
 
-def test_upload_files_file_name_matches_doc_fname() -> None:
-    """The file tuple passed to OpenAI must carry the original filename."""
+@pytest.mark.parametrize(
+    ("fname", "expected"),
+    [
+        ("report.pdf", "report.pdf"),
+        ("Report.PDF", "Report.pdf"),
+        ("README", "README"),
+    ],
+)
+def test_upload_files_file_name_lowercases_only_extension(
+    fname: str, expected: str
+) -> None:
+    """The file tuple passed to OpenAI keeps the stem's case, lowercases the extension."""
     client = MagicMock()
     client.files.create.return_value = MagicMock(id="file-abc")
     provider = OpenAIProvider(client=client)
 
     storage = _make_storage(b"data")
     doc = _make_doc(file_size_kb=1.0)
-    doc.fname = "report.pdf"
+    doc.fname = fname
 
     session_p, crud_p = _patch_session_and_crud()
     with session_p as MockSession, crud_p:
@@ -313,8 +323,8 @@ def test_upload_files_file_name_matches_doc_fname() -> None:
         provider.upload_files(storage, [doc], project_id=1)
 
     _, kwargs = client.files.create.call_args
-    fname, _ = kwargs["file"]
-    assert fname == "report.pdf"
+    uploaded_fname, _ = kwargs["file"]
+    assert uploaded_fname == expected
 
 
 def test_upload_files_raises_on_db_update_failure() -> None:
diff --git a/docs/wiki/modules/knowledge-base.md b/docs/wiki/modules/knowledge-base.md
index 4cefd5719..179b0c69d 100644
--- a/docs/wiki/modules/knowledge-base.md
+++ b/docs/wiki/modules/knowledge-base.md
@@ -35,6 +35,7 @@ All paths relative to `backend/app/`.
 ## Gotchas
 - `signed_url` behaviour is chosen per request by the `download` query param (default false = inline, as before). `put()` stores the upload's `Content-Type`, so a bare presigned URL for a PDF opens in a tab (CSV/XLSX datasets happen to download, which is why only KB docs looked broken). `get_signed_url(..., filename=...)` adds `ResponseContentDisposition: attachment`; `_signed_url()` in `services/documents/helpers.py` passes `fname` only when `download` is set, and the flag threads through `build_document_schema(s)` / `build_job_schema(s)` from `api/routes/documents.py`, `doc_transformation_job.py` and `collections.py`. An attachment URL will not render in an `<iframe>`, so preview callers must leave `download` off. The upload response and the doc-transform callback have no request to carry the flag and always sign inline. Leave `filename` unset for URLs meant to play inline (TTS audio in `api/routes/llm.py`).
 - Uploads de-duplicate by provider file ID (see deep dive §7).
+- Extension case: `get_file_format` lowercases the suffix, so `report.PDF` uploads fine and is stored with its original case in `fname` (the `download` attachment name depends on it). OpenAI file search may not match an uppercase suffix, so `services/collections/providers/openai.py` lowercases only the extension of the name it sends to `files.create`.
 - Collections are immutable-ish: deletion semantics in deep dive §10.
 - OpenAI file-batch id: the SDK's `file_batches.poll()` / `upload_and_poll()` final return deserializes a vector-store body, so its `.id` is the `vs_` id, not the `vsfb_` batch id. `crud/rag/open_ai.py` captures the batch id from `create()` before polling and uses it for `list_files`. Any failed file is a hard failure (whole vector store rolled back); partial indexing needs an add-documents endpoint first.
 - The SDK's `file_batches.poll()` never times out. `_poll_file_batch` polls `retrieve` in a loop with no internal deadline — the Celery soft time limit bounds it, and its `SoftTimeLimitExceeded` aborts the task. An earlier version took a deadline from the caller via a `task_budget` `ContextVar` (and a fixed `BATCH_POLL_TIMEOUT_SECONDS`); both were deleted. Don't reintroduce caller coupling here.