LLM Prompt Adherence/Response Formatting.

[rtuszik]

Issues Persist when using ollama due to bad prompt adherence. As brought up by @dxcore35 in #15.

The "format" parameter that is used by ollama is not available using the OpenAI-Node library.

In a future overhaul of the api-call logic, it would be possible to make this work.

For now, I suggest testing different models and some improvements with regards to the prompt.

PR for a suggestion will follow shortly.

[rtuszik]

As a temporary "fix", this prompt-template works well for me with ollama and llama3.2:

"""
{{input}}
"""
Answer format is JSON {reliability:0~1, outputs:[tag1,tag2,...]}. 
Even if you are unsure, qualify the reliability and select the best matches.
Respond only with valid JSON. Do not write an introduction or summary.
Output tags must be from these options:
{{reference}}
`;

[dxcore35]

🛑 gemma2:9b
🛑 qwen2.5