[GH-ISSUE #15 ] v3 repetitive function call ? #10

Author

Owner

@a-hariti commented on GitHub (Jan 3, 2025):

+1

I'm using it with the Vercel AI sdk and it keeps looping

@a-hariti commented on GitHub (Jan 3, 2025): +1 I'm using it with the Vercel AI sdk and it keeps looping

bsakaguchi commented

Author

Owner

@ninghao commented on GitHub (Jan 5, 2025):

same :)

@ninghao commented on GitHub (Jan 5, 2025): same :)

bsakaguchi commented

Author

Owner

@Devincool commented on GitHub (Jan 6, 2025):

+1

@Devincool commented on GitHub (Jan 6, 2025): +1

bsakaguchi commented

2026-10-26 21:10:32 +00:00

Author

Owner

@vincent-pli commented on GitHub (Jan 6, 2025):

+1， cannot support groupchat pattern in multiple agents with function call, I leverage autogen, not work as expected

@vincent-pli commented on GitHub (Jan 6, 2025): +1， cannot support groupchat pattern in multiple agents with function call, I leverage autogen, not work as expected

bsakaguchi commented

Author

Owner

@ro-hansolo commented on GitHub (Jan 7, 2025):

So the temporary fix I have come upon with, that seems to fix this, is appending a user message with the function response to the messages array before sending it to the model too.

function createMessagesWithToolUse(
	messages: readonly ChatCompletionMessageParam[],
	toolName: string,
	toolArgs: string,
	toolId: string,
	functionResponse: unknown,
	model: string
): readonly ChatCompletionMessageParam[] {
	const baseMessages = [
		...messages,
		{
			role: 'assistant',
			tool_calls: [
				{
					type: 'function',
					id: toolId,
					function: {
						name: toolName,
						arguments: toolArgs,
					},
				},
			],
		},
		{
			role: 'tool',
			content: JSON.stringify(functionResponse),
			tool_call_id: toolId,
		},
	] as readonly ChatCompletionMessageParam[];

	return model === 'deepseek-chat' ?
			[
				...baseMessages,
				{
					role: 'user',
					content: JSON.stringify(functionResponse),
				},
			]
		:	baseMessages;
}

This seems to get the model to respond too the tool use, but this is tacky and still sometimes creates 2 function calls before working. This needs to be fixed asap however

@ro-hansolo commented on GitHub (Jan 7, 2025): So the temporary fix I have come upon with, that seems to fix this, is appending a user message with the function response to the messages array before sending it to the model too. ```javascript function createMessagesWithToolUse( messages: readonly ChatCompletionMessageParam[], toolName: string, toolArgs: string, toolId: string, functionResponse: unknown, model: string ): readonly ChatCompletionMessageParam[] { const baseMessages = [ ...messages, { role: 'assistant', tool_calls: [ { type: 'function', id: toolId, function: { name: toolName, arguments: toolArgs, }, }, ], }, { role: 'tool', content: JSON.stringify(functionResponse), tool_call_id: toolId, }, ] as readonly ChatCompletionMessageParam[]; return model === 'deepseek-chat' ? [ ...baseMessages, { role: 'user', content: JSON.stringify(functionResponse), }, ] : baseMessages; } ``` This seems to get the model to respond too the tool use, but this is tacky and still sometimes creates 2 function calls before working. This needs to be fixed asap however

bsakaguchi commented

2026-10-26 21:10:32 +00:00

Author

Owner

@leslie2046 commented on GitHub (Jan 8, 2025):

+1

@leslie2046 commented on GitHub (Jan 8, 2025): +1 ![d1381ff68fc5afc4bd0db24e47c2346](https://github.com/user-attachments/assets/ec97c820-c74f-49bd-9f93-04a5d72119fd)

bsakaguchi commented

2026-10-26 21:10:33 +00:00

Author

Owner

@Devincool commented on GitHub (Jan 8, 2025):

So the temporary fix I have come upon with, that seems to fix this, is appending a user message with the function response to the messages array before sending it to the model too.

function createMessagesWithToolUse(
	messages: readonly ChatCompletionMessageParam[],
	toolName: string,
	toolArgs: string,
	toolId: string,
	functionResponse: unknown,
	model: string
): readonly ChatCompletionMessageParam[] {
	const baseMessages = [
		...messages,
		{
			role: 'assistant',
			tool_calls: [
				{
					type: 'function',
					id: toolId,
					function: {
						name: toolName,
						arguments: toolArgs,
					},
				},
			],
		},
		{
			role: 'tool',
			content: JSON.stringify(functionResponse),
			tool_call_id: toolId,
		},
	] as readonly ChatCompletionMessageParam[];

	return model === 'deepseek-chat' ?
			[
				...baseMessages,
				{
					role: 'user',
					content: JSON.stringify(functionResponse),
				},
			]
		:	baseMessages;
}

This seems to get the model to respond too the tool use, but this is tacky and still sometimes creates 2 function calls before working. This needs to be fixed asap however

Not work for me

@Devincool commented on GitHub (Jan 8, 2025): ![image](https://github.com/user-attachments/assets/90eb450b-e02a-4241-9533-fa353c026c71) > So the temporary fix I have come upon with, that seems to fix this, is appending a user message with the function response to the messages array before sending it to the model too. > > ```js > function createMessagesWithToolUse( > messages: readonly ChatCompletionMessageParam[], > toolName: string, > toolArgs: string, > toolId: string, > functionResponse: unknown, > model: string > ): readonly ChatCompletionMessageParam[] { > const baseMessages = [ > ...messages, > { > role: 'assistant', > tool_calls: [ > { > type: 'function', > id: toolId, > function: { > name: toolName, > arguments: toolArgs, > }, > }, > ], > }, > { > role: 'tool', > content: JSON.stringify(functionResponse), > tool_call_id: toolId, > }, > ] as readonly ChatCompletionMessageParam[]; > > return model === 'deepseek-chat' ? > [ > ...baseMessages, > { > role: 'user', > content: JSON.stringify(functionResponse), > }, > ] > : baseMessages; > } > ``` > > This seems to get the model to respond too the tool use, but this is tacky and still sometimes creates 2 function calls before working. This needs to be fixed asap however Not work for me

bsakaguchi commented

2026-10-26 21:10:33 +00:00

Author

Owner

@mgisr commented on GitHub (Jan 8, 2025):

same🧐

@mgisr commented on GitHub (Jan 8, 2025): same🧐

bsakaguchi commented

2026-10-26 21:10:34 +00:00

Author

Owner

@ro-hansolo commented on GitHub (Jan 9, 2025):

As I mentioned before it doesn't work on the first go always for me also, but the loop ends after a second function call atleast

However considering we're paying per token this needs to be fixed ASAP by the devs

@ro-hansolo commented on GitHub (Jan 9, 2025): As I mentioned before it doesn't work on the first go always for me also, but the loop ends after a second function call atleast However considering we're paying per token this needs to be fixed ASAP by the devs

bsakaguchi commented

2026-10-26 21:10:34 +00:00

Author

Owner

@LittleLittleCloud commented on GitHub (Jan 9, 2025):

same

@LittleLittleCloud commented on GitHub (Jan 9, 2025): same

bsakaguchi commented

2026-10-26 21:10:34 +00:00

Author

Owner

@DillionApple commented on GitHub (Jan 9, 2025):

Thanks for your feedback, DeepSeek Engineer here. The Function Calling feature is still unstable in this version. We will continue to work on the issue and hopefully it will be resolved in the next major version of DeepSeek-V3.

@DillionApple commented on GitHub (Jan 9, 2025): Thanks for your feedback, DeepSeek Engineer here. The Function Calling feature is still unstable in this version. We will continue to work on the issue and hopefully it will be resolved in the next major version of DeepSeek-V3.

bsakaguchi commented

2026-10-26 21:10:35 +00:00

Author

Owner

@Hairi81 commented on GitHub (Jan 11, 2025):

Thanks for your feedback, DeepSeek Engineer here. The Function Calling feature is still unstable in this version. We will continue to work on the issue and hopefully it will be resolved in the next major version of DeepSeek-V3.

Dear DillionApple,
If the function calling is still unstable then how does the search feature in chat.deepseek.com work?
Also, can you give an estimated time until the next release that will fix this issue?
Thank you for your attention and time

@Hairi81 commented on GitHub (Jan 11, 2025): > Thanks for your feedback, DeepSeek Engineer here. The Function Calling feature is still unstable in this version. We will continue to work on the issue and hopefully it will be resolved in the next major version of DeepSeek-V3. Dear DillionApple, If the function calling is still unstable then how does the search feature in chat.deepseek.com work? Also, can you give an estimated time until the next release that will fix this issue? Thank you for your attention and time

bsakaguchi commented

2026-10-26 21:10:35 +00:00

Author

Owner

@leinad59 commented on GitHub (Jan 15, 2025):

+1

@leinad59 commented on GitHub (Jan 15, 2025): +1

bsakaguchi commented

2026-10-26 21:10:36 +00:00

Author

Owner

@lilyjunefield commented on GitHub (Jan 15, 2025):

Ran into this myself creating a DeepSeek powered CLI tool. The tool would go into an infinite loop using the DeepSeek API when calling tools. Not always but about 80% of the time from my testing.

@lilyjunefield commented on GitHub (Jan 15, 2025): Ran into this myself creating a DeepSeek powered CLI tool. The tool would go into an infinite loop using the DeepSeek API when calling tools. Not always but about 80% of the time from my testing.

bsakaguchi commented

2026-10-26 21:10:36 +00:00

Author

Owner

@kermankohli commented on GitHub (Jan 16, 2025):

+1. wasted so much time on this.

@kermankohli commented on GitHub (Jan 16, 2025): +1. wasted so much time on this.

bsakaguchi commented

2026-10-26 21:10:36 +00:00

Author

Owner

@nils010485 commented on GitHub (Jan 16, 2025):

+1 Same issue, the model is unusable with the tools for now

@nils010485 commented on GitHub (Jan 16, 2025): +1 Same issue, the model is unusable with the tools for now

bsakaguchi commented

2026-10-26 21:10:37 +00:00

Author

Owner

@a-hariti commented on GitHub (Jan 17, 2025):

to be fair it responds with json when asked for one thing

I tried it with the Vercel AI SDK to generateObject and it works

It's not far off!

@a-hariti commented on GitHub (Jan 17, 2025): to be fair it responds with json when asked for one thing I tried it with the Vercel AI SDK to `generateObject` and it works It's not far off!

bsakaguchi commented

2026-10-26 21:10:37 +00:00

Author

Owner

@crixue commented on GitHub (Jan 19, 2025):

I meet the same issue. Even I tried different temperatures and refine the system prompt like:
You serve as an intelligent assistant. You must give your final answer in 1 rounds tool calls. And if you don't know the answer, just answer 'Sorry, I don't know'

It still repeat to call the tool until the input tokens exceed 64k context tokens limit. I leverage the langchain components to call agent to call tools. Qwen-max and Gpt-4o have a good understanding the various system prompt, but deepseek-v3 is not. It's so frustrating, even though the price is cheap.

@crixue commented on GitHub (Jan 19, 2025): I meet the same issue. Even I tried different temperatures and refine the system prompt like: You serve as an intelligent assistant. You must give your final answer in 1 rounds tool calls. And if you don't know the answer, just answer 'Sorry, I don't know' It still repeat to call the tool until the input tokens exceed 64k context tokens limit. I leverage the langchain components to call agent to call tools. Qwen-max and Gpt-4o have a good understanding the various system prompt, but deepseek-v3 is not. It's so frustrating, even though the price is cheap.

bsakaguchi commented

2026-10-26 21:10:38 +00:00

Author

Owner

@chendwww commented on GitHub (Jan 19, 2025):

+1

@chendwww commented on GitHub (Jan 19, 2025): +1

bsakaguchi commented

2026-10-26 21:10:38 +00:00

Author

Owner

@chendwww commented on GitHub (Jan 19, 2025):

+1

@chendwww commented on GitHub (Jan 19, 2025): +1

bsakaguchi commented

2026-10-26 21:10:39 +00:00

Author

Owner

@chendwww commented on GitHub (Jan 19, 2025):

+1

@chendwww commented on GitHub (Jan 19, 2025): +1

bsakaguchi commented

2026-10-26 21:10:39 +00:00

Author

Owner

@catusax commented on GitHub (Jan 20, 2025):

same problem, my solution is remove the tools section in the too response message, so the model will have no tools to call in the next round

@catusax commented on GitHub (Jan 20, 2025): same problem, my solution is remove the `tools` section in the too response message, so the model will have no tools to call in the next round

bsakaguchi commented

2026-10-26 21:10:40 +00:00

Author

Owner

@rovnyart commented on GitHub (Jan 20, 2025):

+1

@rovnyart commented on GitHub (Jan 20, 2025): +1

bsakaguchi commented

2026-10-26 21:10:40 +00:00

Author

Owner

@tempurai commented on GitHub (Jan 20, 2025):

+1

@tempurai commented on GitHub (Jan 20, 2025): +1

bsakaguchi commented

https://github.com/deepseek-ai/DeepSeek-V3/issues/15#issuecomment-2567432464

Author

Owner

@graysonchen commented on GitHub (Jan 21, 2025):

@DillionApple Thanks updated. So, approximately what time would that be?

@graysonchen commented on GitHub (Jan 21, 2025): https://github.com/deepseek-ai/DeepSeek-V3/issues/15#issuecomment-2567432464 @DillionApple Thanks updated. So, approximately what time would that be?

bsakaguchi commented

Author

Owner

@golforado commented on GitHub (Jan 21, 2025):

+1

@golforado commented on GitHub (Jan 21, 2025): +1

bsakaguchi commented

Author

Owner

@aaahariri commented on GitHub (Jan 22, 2025):

+1
Also noticed that it wasn’t reliable with tools. Any update on properly supporting tools/function calling?

@aaahariri commented on GitHub (Jan 22, 2025): +1 Also noticed that it wasn’t reliable with tools. Any update on properly supporting tools/function calling?

bsakaguchi commented

2026-10-26 21:10:42 +00:00

Author

Owner

@koko0v0 commented on GitHub (Jan 22, 2025):

+1
Would love to see if there is any update on this

@koko0v0 commented on GitHub (Jan 22, 2025): +1 Would love to see if there is any update on this

bsakaguchi commented

Author

Owner

@teddyzxcv commented on GitHub (Jan 23, 2025):

+1 encountered with this problem in tool calling for n8n workflow

@teddyzxcv commented on GitHub (Jan 23, 2025): +1 encountered with this problem in tool calling for n8n workflow

bsakaguchi commented

2026-10-26 21:10:42 +00:00

Author

Owner

@tahafatih commented on GitHub (Jan 27, 2025):

Same issue here

@tahafatih commented on GitHub (Jan 27, 2025): Same issue here

bsakaguchi commented

2026-10-26 21:10:43 +00:00

Author

Owner

@Sanket9225 commented on GitHub (Jan 27, 2025):

Facing same issue, please fix this asap, Without tool calling developers can not build anything from it. At most it will support RAG, but nothing agentic.

@Sanket9225 commented on GitHub (Jan 27, 2025): Facing same issue, please fix this asap, Without tool calling developers can not build anything from it. At most it will support RAG, but nothing agentic.

bsakaguchi commented

2026-10-26 21:10:44 +00:00

Author

Owner

@kevintsai1202 commented on GitHub (Jan 28, 2025):

The new model has been released, but the function calling feature is still not fixed. This is an essential functionality for developing agents. I hope it gets resolved soon.

@kevintsai1202 commented on GitHub (Jan 28, 2025): The new model has been released, but the function calling feature is still not fixed. This is an essential functionality for developing agents. I hope it gets resolved soon.

bsakaguchi commented

2026-10-26 21:10:44 +00:00

Author

Owner

@Hairi81 commented on GitHub (Jan 28, 2025):

Dear @mowentian @DillionApple

It's been over a month and you've released a new CoT reasoning model, R1, yet still haven't fixed this issue.
Does the team need the community help to optimize/post train multi-turn native function calling?
If you're able to provide more detail on how the function calling was trained (JSON with SFT/RL?)
perhaps the community can assists in fixing this issue

Best Regards,
Hairi

@Hairi81 commented on GitHub (Jan 28, 2025): Dear @mowentian @DillionApple It's been over a month and you've released a new CoT reasoning model, R1, yet still haven't fixed this issue. Does the team need the community help to optimize/post train multi-turn native function calling? If you're able to provide more detail on how the function calling was trained (JSON with SFT/RL?) perhaps the community can assists in fixing this issue Best Regards, Hairi

bsakaguchi commented

2026-10-26 21:10:45 +00:00

Author

Owner

@huangmouren2023 commented on GitHub (Jan 28, 2025):

是不是缓存命中的锅？我一直觉得，应该给这个机制增加一个可以选择配置开启关闭的开关

@huangmouren2023 commented on GitHub (Jan 28, 2025): 是不是缓存命中的锅？我一直觉得，应该给这个机制增加一个可以选择配置开启关闭的开关

bsakaguchi commented

2026-10-26 21:10:45 +00:00

Author

Owner

@Hairi81 commented on GitHub (Jan 28, 2025):

It could also be the cache hit but I've seen other OS LLM models experiencing similar infinite loop, they were having difficulties escaping the function call params inside the JSON and understanding delimiters

@Hairi81 commented on GitHub (Jan 28, 2025): It could also be the cache hit but I've seen other OS LLM models experiencing similar infinite loop, they were having difficulties escaping the function call params inside the JSON and understanding delimiters

bsakaguchi commented

2026-10-26 21:10:45 +00:00

Author

Owner

@heshengtao commented on GitHub (Jan 28, 2025):

调用deepseek API时遇到了同样的问题，在调用代码解释器工具后，工具已经返回了正确的结果，但是模型无法从工具结果中回复用户，而是强迫症一般重新调用工具。
希望贵团队可以加快修复该问题，这对我们基于deepseek开发智能体真的很重要，感谢贵团队的贡献！新年快乐！

@heshengtao commented on GitHub (Jan 28, 2025): 调用deepseek API时遇到了同样的问题，在调用代码解释器工具后，工具已经返回了正确的结果，但是模型无法从工具结果中回复用户，而是强迫症一般重新调用工具。希望贵团队可以加快修复该问题，这对我们基于deepseek开发智能体真的很重要，感谢贵团队的贡献！新年快乐！

bsakaguchi commented

Author

Owner

@kimifish commented on GitHub (Feb 3, 2025):

I'm getting openai.APIStatusError: Error code: 405 - {'detail': 'Tool calling is not supported for model: deepseek-ai/DeepSeek-V3'} while using langchain-deepseek library. Is that somehow connected with unstable function calls? Or that's another problem in my case?

@kimifish commented on GitHub (Feb 3, 2025): I'm getting `openai.APIStatusError: Error code: 405 - {'detail': 'Tool calling is not supported for model: deepseek-ai/DeepSeek-V3'}` while using langchain-deepseek library. Is that somehow connected with unstable function calls? Or that's another problem in my case?

bsakaguchi commented

Author

Owner

@SavetheTears commented on GitHub (Feb 5, 2025):

I'm getting openai.APIStatusError: Error code: 405 - {'detail': 'Tool calling is not supported for model: deepseek-ai/DeepSeek-V3'} while using langchain-deepseek library. Is that somehow connected with unstable function calls? Or that's another problem in my case?

i believe you did not initialise the model correctly, check your code where you used ChatOpenAI, the params should include model, base_url, and your deepseek key.

@SavetheTears commented on GitHub (Feb 5, 2025): > I'm getting `openai.APIStatusError: Error code: 405 - {'detail': 'Tool calling is not supported for model: deepseek-ai/DeepSeek-V3'}` while using langchain-deepseek library. Is that somehow connected with unstable function calls? Or that's another problem in my case? i believe you did not initialise the model correctly, check your code where you used ChatOpenAI, the params should include model, base_url, and your deepseek key.

bsakaguchi commented

Author

Owner

@2000calories commented on GitHub (Feb 5, 2025):

Same issue here

@2000calories commented on GitHub (Feb 5, 2025): Same issue here

bsakaguchi commented