Prompt Template Issue with Mixtral Model: "Conversation roles must alternate user/assistant"

[mdetommasi]

Hi everyone,
I'm encountering an error while using RAGFlow with a Mixtral model. The error seems to be related to chat conversation roles

ERROR: Error code: 400 - {'object': 'error', 'message': 'Conversation roles must alternate user/assistant/user/assistant/... Conversation roles must alternate user/assistant/user/assistant/...', 'type': 'BadRequestError', 'param': None, 'code': 400}

From my investigation, it appears Mixtral expects strict alternation between 'user' and 'assistant' roles, but RAGFlow might be sending a 'system' role instead, which breaks the expected pattern.
Has anybody else encountered this prompt template compatibility issue with Mixtral models in RAGFlow? If so, were you able to find a solution or workaround?
Thanks in advance for any insights!

[dosubot[bot]]

Hi @mdetommasi! I'm Dosu and I’m helping the ragflow team.

Yes, others have run into this error with Mixtral and other OpenAI-compatible models in RAGFlow. The root cause is that some RAGFlow workflows—especially knowledge graph extraction—can send chat message histories that don't strictly alternate between 'user' and 'assistant' roles, or include a 'system' role outside the initial position, which breaks the pattern required by Mixtral and similar models. There isn't a built-in config or code-level check in RAGFlow to enforce this alternation, so the most reliable workaround is to disable knowledge graph creation, which avoids sending problematic message histories. This issue is systemic in RAGFlow's knowledge graph logic and affects any OpenAI-compatible backend, not just Mixtral. If you need knowledge graph features, you'd have to patch the message construction logic to guarantee strict alternation before sending requests. More details and discussion are available in these threads: #9334, #3062.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}    

[mdetommasi]

@dosu I disabled the knowledge graph both in the Chat settings and Dataset configurations, but the problem persists.

[dosubot[bot]]

Disabling the knowledge graph doesn't always resolve the Mixtral prompt template error in RAGFlow—other workflows or components can still send message histories that break the required alternation between 'user' and 'assistant' roles, or include 'system' roles outside the initial position. This is a known systemic issue, especially after recent updates, and has been reported by multiple users even with knowledge graph fully disabled. Troubleshooting steps include verifying that all containers are running the same version, clearing browser and backend caches, and capturing the actual payload sent to the LLM API to inspect for role alternation issues. If the error persists, manual review or patching of the message construction logic may be needed to guarantee strict alternation before requests are sent. More details and user reports are available in issue #9334.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}    

[Magicbook1108]

I have raised a bug issue in #11021 for this discussion. And the answer is we currently insert system role during conversation, and mistral model strictly don't accept system role.