From ad9f6ce23b4d4d39f493b3997d3b9f095cc1f2b7 Mon Sep 17 00:00:00 2001
From: spions <spionsbbs@gmail.com>
Date: Fri, 7 Aug 2026 13:16:07 +0300
Subject: [PATCH] fix(logging): track token usage for rerank calls in spend
 logs

RerankResponse has no top-level `usage` field, only `meta.billed_units`/
`meta.tokens`, so get_usage_from_response_obj() always returned zeroed
Usage for every rerank provider, causing /ui/usage to show 0 tokens for
all rerank calls regardless of provider.
---
 litellm/litellm_core_utils/litellm_logging.py | 14 +++++++
 .../test_standard_logging_payload.py          | 37 +++++++++++++++++++
 2 files changed, 51 insertions(+)

diff --git a/litellm/litellm_core_utils/litellm_logging.py b/litellm/litellm_core_utils/litellm_logging.py
index 99721c3ffa2..c26c69c3d48 100644
--- a/litellm/litellm_core_utils/litellm_logging.py
+++ b/litellm/litellm_core_utils/litellm_logging.py
@@ -4719,6 +4719,20 @@ class StandardLoggingPayloadSetup:
             )
 
         usage: Final = response_obj.get("usage", None) or {}
+        if not usage:
+            # RerankResponse has no top-level `usage` field - token usage lives
+            # under `meta.billed_units`/`meta.tokens` instead (Cohere-style API).
+            meta = response_obj.get("meta")
+            if isinstance(meta, dict) and ("billed_units" in meta or "tokens" in meta):
+                billed_units = meta.get("billed_units") or {}
+                tokens = meta.get("tokens") or {}
+                total_tokens = billed_units.get("total_tokens", 0) or 0
+                prompt_tokens = tokens.get("input_tokens", total_tokens) or total_tokens
+                return Usage(
+                    prompt_tokens=prompt_tokens,
+                    completion_tokens=0,
+                    total_tokens=total_tokens,
+                )
         if usage is None or (not isinstance(usage, dict) and not isinstance(usage, Usage)):
             return Usage(
                 prompt_tokens=0,
diff --git a/tests/logging_callback_tests/test_standard_logging_payload.py b/tests/logging_callback_tests/test_standard_logging_payload.py
index f29b245b3be..79771ada9a4 100644
--- a/tests/logging_callback_tests/test_standard_logging_payload.py
+++ b/tests/logging_callback_tests/test_standard_logging_payload.py
@@ -127,6 +127,43 @@ def test_get_usage_from_image_generation_response():
     assert usage.completion_tokens_details.text_tokens == 100
 
 
+def test_get_usage_from_rerank_response():
+    """
+    RerankResponse has no top-level `usage` field - token usage lives under
+    `meta.billed_units`/`meta.tokens` (Cohere-style API), which is also what
+    hosted_vllm/other rerank servers return. Without this, /ui/usage always
+    showed 0 tokens for every rerank call.
+    """
+    response_obj = {
+        "id": "rerank-d618748e0f5543e8ba09ee7dd131ac59",
+        "results": [],
+        "meta": {
+            "billed_units": {"total_tokens": 42},
+            "tokens": {"input_tokens": 42},
+        },
+    }
+
+    usage = StandardLoggingPayloadSetup.get_usage_from_response_obj(response_obj)
+
+    assert usage.prompt_tokens == 42
+    assert usage.completion_tokens == 0
+    assert usage.total_tokens == 42
+
+
+def test_get_usage_from_rerank_response_no_meta():
+    """
+    A rerank response with no usable `meta` (and no `usage`) should still
+    fall back to all-zero usage instead of raising.
+    """
+    response_obj = {"id": "rerank-abc", "results": [], "meta": None}
+
+    usage = StandardLoggingPayloadSetup.get_usage_from_response_obj(response_obj)
+
+    assert usage.prompt_tokens == 0
+    assert usage.completion_tokens == 0
+    assert usage.total_tokens == 0
+
+
 def test_get_additional_headers():
     additional_headers = {
         "x-ratelimit-limit-requests": "2000",