fix(ai-proxy): set content-length for non compressed response

kong/llm/drivers/shared.lua

@@ -175,7 +175,7 @@ _M.operation_map = {
 }
 
 _M.clear_response_headers = {
-  shared = {
+  shared = { -- deprecared, not using
     "Content-Length",
   },
   openai = {

kong/llm/plugin/base.lua

@@ -110,6 +110,8 @@ function MetaPlugin:header_filter(sub_plugin, conf)
       -- and seems nginx doesn't support it
 
     elseif get_global_ctx("accept_gzip") then
+      -- for gzip response, don't set content-length at all to align with upstream
+      kong.response.clear_header("Content-Length")
       kong.response.set_header("Content-Encoding", "gzip")
     end
 

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

@@ -2,7 +2,6 @@ local cjson = require("cjson")
 
 local ai_plugin_ctx = require("kong.llm.plugin.ctx")
 local ai_plugin_o11y = require("kong.llm.plugin.observability")
-local ai_shared = require("kong.llm.drivers.shared")
 
 local _M = {
   NAME = "normalize-json-response",
@@ -57,6 +56,8 @@ local function transform_body(conf)
   end
 
   set_global_ctx("response_body", response_body) -- to be sent out later or consumed by other plugins
+
+  return #response_body
 end
 
 function _M:run(conf)
@@ -81,8 +82,9 @@ function _M:run(conf)
   -- if not streaming, prepare the response body buffer
   -- this must be called before sending any response headers so that
   -- we can modify status code if needed
+  local body_length
   if not get_global_ctx("stream_mode") then
-    transform_body(conf)
+    body_length = transform_body(conf)
   end
 
   -- populate cost
@@ -94,12 +96,11 @@ function _M:run(conf)
     ai_plugin_o11y.metrics_set("llm_usage_cost", 0)
   end
 
-  -- clear shared restricted headers
-  for _, v in ipairs(ai_shared.clear_response_headers.shared) do
-    kong.response.clear_header(v)
+  if not get_global_ctx("accept_gzip") and not get_global_ctx("stream_mode") then
+    -- otherwise use our transformed body length
+    kong.response.set_header("Content-Length", body_length)
   end
 
-
   if ngx.var.http_kong_debug or conf.model_name_header then
     local model_t = ai_plugin_ctx.get_request_model_table_inuse()
     assert(model_t and model_t.name, "model name is missing")
@@ -109,4 +110,4 @@ function _M:run(conf)
   return true
 end
 
-return _M
+return _M