diff --git a/llmage/llmclient.py b/llmage/llmclient.py
index 2b493ee..8dbea65 100644
--- a/llmage/llmclient.py
+++ b/llmage/llmclient.py
@@ -93,6 +93,18 @@ async def uapi_request(request, llm, callerid, callerorgid, params_kw=None):
 		llmusage.accounting_status = 'created'
 		await write_llmusage(llmusage)
 	except Exception as e:
+		# Retry on duplicate key: append *1, *2, ...
+		retry = 0
+		while 'Duplicate entry' in str(e) and retry < 3:
+			retry += 1
+			llmusage.id = f'{luid}*{retry}'
+			try:
+				await write_llmusage(llmusage)
+				exception(f'write_llmusage retry {retry} succeeded with id={llmusage.id}')
+				e = None
+				break
+			except Exception as e2:
+				e = e2
 		# Refund balance reservation on failure
 		try:
 			from .balance import refund_balance
diff --git a/wwwroot/v1/chat/completions/index.dspy b/wwwroot/v1/chat/completions/index.dspy
index 6b726bf..87871bb 100644
--- a/wwwroot/v1/chat/completions/index.dspy
+++ b/wwwroot/v1/chat/completions/index.dspy
@@ -1,13 +1,24 @@
 async def gen():
 	env = request._run_ns.copy()
 	f = partial(inference_generator, request, params_kw=params_kw)
-	if params_kw.stream:
-		async for l in f():
-			yield f'data: {l}\\n'
-		yield 'data: [DONE]\\n\\n'
-	else:
-		async for l in f():
-			yield l
+	try:
+		if params_kw.stream:
+			async for l in f():
+				yield f'data: {l}\\n'
+			yield 'data: [DONE]\\n\\n'
+		else:
+			async for l in f():
+				yield l
+	except Exception as e:
+		luid = params_kw.get('_luid')
+		if luid:
+			try:
+				await env.refund_balance(luid)
+			except:
+				pass
+		import json as _json
+		err = _json.dumps({"error": str(e), "status": "FAILED"}, ensure_ascii=False)
+		yield f'data: {err}\\n' if params_kw.stream else err
 
 debug_params('params_kw', params_kw)
 catelogid = params_kw.catelogid or 't2t'