refac

backend/open_webui/utils/middleware.py

@@ -107,6 +107,7 @@ from open_webui.utils.filter import (
 )
 from open_webui.utils.code_interpreter import execute_code_jupyter
 from open_webui.utils.payload import apply_system_prompt_to_body
+from open_webui.utils.response import normalize_usage
 from open_webui.utils.mcp.client import MCPClient
 
 
@@ -3121,10 +3122,11 @@ async def process_chat_response(
                                 else:
                                     choices = data.get("choices", [])
 
-                                    # 17421
+                                    # 17421 - Normalize usage data to standard format
                                     usage = data.get("usage", {}) or {}
                                     usage.update(data.get("timings", {}))  # llama.cpp
                                     if usage:
+                                        usage = normalize_usage(usage)
                                         await event_emitter(
                                             {
                                                 "type": "chat:completion",

backend/open_webui/utils/response.py

@@ -6,6 +6,47 @@ from open_webui.utils.misc import (
 )
 
 
+def normalize_usage(usage: dict) -> dict:
+    """
+    Normalize usage statistics to standard format.
+    Handles OpenAI, Ollama, and llama.cpp formats.
+
+    Adds standardized token fields to the original data:
+    - input_tokens: Number of tokens in the prompt
+    - output_tokens: Number of tokens generated
+    - total_tokens: Sum of input and output tokens
+    """
+    if not usage:
+        return {}
+
+    # Map various field names to standard names
+    input_tokens = (
+        usage.get("input_tokens")  # Already standard
+        or usage.get("prompt_tokens")  # OpenAI
+        or usage.get("prompt_eval_count")  # Ollama
+        or usage.get("prompt_n")  # llama.cpp
+        or 0
+    )
+
+    output_tokens = (
+        usage.get("output_tokens")  # Already standard
+        or usage.get("completion_tokens")  # OpenAI
+        or usage.get("eval_count")  # Ollama
+        or usage.get("predicted_n")  # llama.cpp
+        or 0
+    )
+
+    total_tokens = usage.get("total_tokens") or (input_tokens + output_tokens)
+
+    # Add standardized fields to original data
+    result = dict(usage)
+    result["input_tokens"] = int(input_tokens)
+    result["output_tokens"] = int(output_tokens)
+    result["total_tokens"] = int(total_tokens)
+
+    return result
+
+
 def convert_ollama_tool_call_to_openai(tool_calls: list) -> list:
     openai_tool_calls = []
     for tool_call in tool_calls: