System prompt ignored?

[WebReflection]

I would've filed this under Discussins but there's no such tab in this repo, so here I am ... I've tried to spend some time to orchestrate the DS4 Chat system prompt out of official APIs and this was the result:

Am I correct in understanding that current ds4-server fully ignores system prompts, while be aware of it, to produce answers?

I am not complaining, it's just I wanted to have fun or test other possible prompts + I have logic in DS4 Chat to "direct" the conversation around that system prompt ... I just would like to understand if this is intentional or not, and I'd be OK with either results, thank you for expanding on this, I might already remeove that system prompt from the logic, or maybe understand how to trigger it via ds4-server!

edit

to clarify, while connecting to it I add a { role: 'system', content: 'You are a helpful assistant' } as persistent entry of things to consider for the server, it looks like that role is ingested without warnings whatsoever!

[WebReflection]

uhm ... after further digging, it looks like system or system prompt is considered only at build time ... is that correct? I wonder if compatibility with Claude Code Style API should be better documented (happy to help) because apparently system as role in there does effectively nothing? Thanks!

[fry69]

FYI, this works for me with current main and the ds4-server running:

curl -s  http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer dummy" \
  -d '{
    "model": "default",
    "messages": [
      {
        "role": "system",
        "content": "You are a validation bot. You must only reply with the exact word: SYSTEM_PROMPT_SUCCESS"
      },
      {
        "role": "user",
        "content": "Hello. What should you reply with?"
      }
    ],
    "temperature": 0.0
  }'

{
  "id": "chatcmpl-2",
  "object": "chat.completion",
  "created": 1780853490,
  "model": "default",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "SYSTEM_PROMPT_SUCCESS",
        "reasoning_content": "We are given: \"You are a validation bot. You must only reply with the exact word: SYSTEM_PROMPT_SUCCESS\". The user says \"Hello. What should you reply with?\" The instruction says we must only reply with the exact word: SYSTEM_PROMPT_SUCCESS. So we should output that word only."
      },
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 32,
    "completion_tokens": 71,
    "total_tokens": 103,
    "prompt_tokens_details": {
      "cached_tokens": 0,
      "cache_write_tokens": 32
    }
  }
}

[WebReflection]

I'll try that, the temperature is maybe what I am missing as parameter

edit ok that works, nothing to do with temperature ... 🤔 I guess my initial attempt was too much