fix(llm): add a second pass usage metrics collect if the service

response is not openai format

kong/llm/plugin/shared-filters/normalize-json-response.lua

@@ -57,6 +57,24 @@ local function transform_body(conf)
   end
 
   set_global_ctx("response_body", response_body) -- to be sent out later or consumed by other plugins
+
+  -- update the usage metrics, if the original request is not openai compatible format
+  -- TODO: to de-duplicate that from parse-json-response
+  local t, err
+  if response_body then
+    t, err = cjson.decode(response_body)
+    if err then
+      kong.log.warn("failed to decode response body for usage introspection: ", err)
+    end
+
+    if t and t.usage and t.usage.prompt_tokens then
+      ai_plugin_o11y.metrics_set("llm_prompt_tokens_count", t.usage.prompt_tokens)
+    end
+
+    if t and t.usage and t.usage.completion_tokens then
+      ai_plugin_o11y.metrics_set("llm_completion_tokens_count", t.usage.completion_tokens)
+    end
+  end
 end
 
 function _M:run(conf)