We are building an image generator, and have the issue that certain models are best for specific applications, interior design vs. art, for example.
What is the best way to build a model “Router” for Pickaxe?
As infrastructure for this, does it make sense that I ask that Actions be made “parameterizable”? As in, yeah, we are going to use Replicate, but depending on what the user asks for, I want to change the model selection?
I have done some research and found some JS packages - Semantic-Router or LangChain RouterChain give you “automatic” model selection immediately, and you can graduate to learning-based systems (RoRF / RouteLLM) but it’s a heavy lift for us simpler Pickaxe users.
Hey @kenlyle I love how you’re thinking about this problem!
Theoretically, you can make this happen via a custom app (which should likely include the custom JS Paclases you found in the scripts) that can switch between models based on user query, and call on the custom app via an HTTP module via a make/n8n scenario.
Also, if your users want to upload a photo for processing, you can use a specific URL tag to handle images on Pickaxe.
The template for such a setup should look something like this:
Make webhook >>> HTTP >>> Webhook Response (Sends processed media back to your Pickaxe).
This is definitely advanced, and I’d be happy to explore it further with you over at the Pickaxe Prospectors Academy.
I just wrote this to the GerMerlin.in guys, but most of it applies to Pickaxe, too:
”
The selector and filters by function (Reasoning, Coding) are pretty good. The token cost indicator works, too.
I think the only missing item would be to pull performance data for the category and model, and offer a metric of price/performance.
For example, Claude models are great for coding and reasoning, but cost 2.5 times as much as GPT 5.2 High (150 to 60), but I wonder, is it 2.5 times as good?
I could almost see each user has a profile with a slider or sliders for price sensitivity.
In this case, I think if GPT were 80% as good for 40% of the price, I would pick it. In very special cases, I would choose Opus 4.6.
Maybe that’s the metric…I would always pick a model that has a price/performance ratio of more than 1.5 (50%) relative to the absolute best model.
Please give this some thought…it’s an obvious next step.
”
In Merlin, I really appreciate the giant context window of Gemini 3 Pro. Noting that, by this metric, Claude costs 2x what GPT does for very similar performance, and GPT even has a bigger context window.
Is this interesting to anyone else? Anyone making this kind of analysis?
It would be awesome if Pickaxe could monitor interactions for model selection, including context window, even setting input and output token counts.
I would love to be able to connect model to an option in Pickaxe. Even to make it a form field that is user selectable. With that I could create a alternative to Merlin or Straico right in Pickaxe.
