Fix search query generation when model generates thinking tags

[William John Shipman (williamjshipman)]

I ran into an issue using LMStudio with the qwq-32b, deepseek-r1-distill-qwen-7b and deepseek-r1-llama-8b models. In the generate_query and reflect_on_summary steps, the models return text with thinking tags before the actual JSON. With the DeepSeek models, I also saw that DeepSeek recommended using to enclose an example of the desired output. This pull request updates the reflect_on_summary prompt with example tags and moves the stripping of think tags to earlier both reflect_on_summary and generate_query so the JSON part of the response can be parsed.

[Paul Cuciureanu (pcuci)]

Came across: https://ollama.com/MFDoom/deepseek-r1-tool-calling:32b

I wonder if this would also solve the structured output expectations(?)

[William John Shipman (williamjshipman)]

I gave this model a quick test, although just the 1.5b parameter version. Still fails to generate JSON.