Use an agent in a third-party interface using API, with the same request and response format as OpenAI API

[Hekas1]

I am trying to use an agent in a third-party interface using API, with the same request and response format as OpenAI API. However, I get the result only when sending a message to the agent, its greeting, and when resending, only the greeting. Do I need to write an intermediate interaction functionality, the entire format does not contain the entire mechanism, do I need to do it myself?

[dosubot[bot]]

RAGFlow natively supports OpenAI-compatible API endpoints for agents, so you don't need to build your own intermediate context management if you use the API as documented. If you only get the agent's greeting each time, it's almost always because the client isn't sending the full message history or session/conversation context in each API call. The API expects you to include the entire conversation history (at least one message with the "user" role) in the messages array of your request body, just like OpenAI's API.

Check that your requests to /api/v1/agents_openai/{agent_id}/chat/completions include the full message history and any required session or conversation IDs. You can find the exact request/response format and requirements in the HTTP API reference. If you follow this pattern, RAGFlow will handle the conversational state for you—no extra logic needed on your end.

If you still see only the greeting, double-check your request structure and authentication token. If you want a minimal working example, let me know!

To reply, just mention @dosu.

---

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

[ZhenhangTung]

@Hekas1 Could you share your agent's orchestration?

BTW, you can wait for 0.20 for the whole Agent module is being redesigned.

[yongtenglei]

To me. First, it seems like your agent take action for each response. However, API send several events or messages to make one complete response. That's why I recommend you wait for a complete response then let agent decide what to do next.

The next,

The agent repeats the same thing to me many times.

Because, the agent never incorporate proper history messages to make the next action. It means, LLM never know you already made your selection, and never get your response. So, it keeps asking you, "Please select the topic that interests you:".

Adjust your workflow to make it work.

[Hekas1]

Thanks for the answer, I'll try. I have a relatively large structure, I'll create the simplest agent and try to expand it. I think I understood your idea and the reason for such agent behavior.

[Hekas1]

Okay, I created an agent consisting of three components:
begin - interact - generate

By sending a greeting message to the agent via a third-party interface, this is what I get

'Generate_0' is running...🕞PriPrivet! YesPrivet! Yes, I'm onlinePrivet! Yes, I'm online. HowPrivet! Yes, I'm online. How can I helpPrivet! Yes, I'm online. How can I help?

[yongtenglei]

You're welcome. I think you could try first to set stream option to false and see if the first issue could be mitigated.

[Hekas1]

Setting stream option to false at the simple agent level solved the problem. Thanks, waiting for the update