One of the huge blockers for us at the moment is the lack of advanced settings for OpenAI API.
For the workflow we need, we have to be able to set Temperature, Top_p / top_k for ChatGPT and choose pecific assistants. Ideally, this would be configurable “per column” (much like the choice of ChatGPT model at the moment)
Currently, this is very difficult to work around, mostly due to the webhooks issue (no option to select manual triggers or filter the column update trigger)
This issue renders Baserow not really competitive in the Ai space. While Airtable’s AI is way too expensive for anything other than causal use, they enable a sensible Make integration.
Hey @dev-rd, we plan to add these settings for OpenAI API, but it’s a low priority for us at the moment. As we see more users requesting these features, we’ll re-discuss their priority.
if webhooks could be retried (same way as the AI field), people could cook up their own solution using n8n or other integration framework. I actually prefer this solution. Otherwise you’re going to make baserow complex. THe average person doesn’t want to see advanced OpenAI parameters.
Custom plugin might be the option if there’s any prospect for a return on investment. @lucw What exactly do you mean by heavyweight development?
As for webhooks - I’ve seen a github issue to hopefully fix this. However, webhooks, as they are implemented right now, are practically useless for any no-code or almost-no-code scenario. Anything that maybe remotely makes sense from a financial standpoint is a self hosted N8N, but then again it is too limited in the community version. Furthermore it would complicate tables and more importantly, make it impossible for anyone outside of the “IT team” to set up a new “automation scenario”, which further complicates workflows.
Complexity may be measured differently for different projects - ultimately - it is much easier to teach people what the “ChatGPT temperature” is and when to tweak it, than it is to teach people how to create N8N scenarios (on top of that)…
As for Baserow complexity:
I may have suggested the wrong scope - temperature and top P are not actually “advanced” features for anyone with background in AI. Crucially, I really do not see how you could make Baserow overly complex by adding one or two parameters/sliders for the AI column or workspace setting. The AI feature is behind a paywall anyways.
I would also argue that Baserow might really benefit from people using it to do their work (work = paying customer) - what I’m getting at is that anyone seriously considering the use of AI will sooner rather than later want to tweak it a bit with at least a temperature setting. If they don’t find those solutions, they simply won’t stay as paying customers. Especially when this is relatively easily doable in Excel (via add-in), Airtable, etc, and even more so if it proves necessary for what they want to do with AI.
With custom plugins, you need to be a full stack web developer pretty much. You need to understand back-end and front-end code, and you need to know how to host your app when it’s ready. And if you need to pay someone to do this for you, it’s not going to be cheap as full stack developers can command a certain salary on the job market.
Though i’d like to understand why you consider webhooks useless for no-code ? What’s the issue exactly ? When a row is updated, your webhook target receives the update. You could then do some processing and publish back a field. It’s not complicated if you follow a tutorial.
Now that i read your message again, a thought comes into my mind. There is massive, massive amounts of focus on LLMs and their performance right now. Some aggregators like openrouter.ai allow you to chat with multiple LLMs at the same time. What if Baserow tried to become the front-end of choice for people wanting to repeatedly generate LLM output based on different rows of input ? Then it might make sense to support multiple parameters and different LLM engines.
openrouter.ai chose to use a clone of the OpenAI API for every single engine: Quick Start | OpenRouter
maybe that’s the way to go with Baserow also. If customization of the OpenAI API parameters was supported, then your use case (and many others) could be easily accomplished.
Summary for @bram@davide : if you allow customizing OpenAI API parameters and allow custom urls, you can now connect baserow to 100’s of LLMs through openrouter.ai: Models | OpenRouter
Thanks for the follow up on the plugin dev - This is more or less what I expected.
However, I started wondering if there are any Baserow specific challenges with plugins - as some of the posts are two or more years old and there is just a handful of them - and - there are still no official plugins.
Re webhooks:
They update on every column change - since we change a lot of fields and only need specific manual triggers for AI - this essentially exhausts any reasonable number of paid “operations” or flows, due to the needed filtering for what changed (if we go with no code solutions such as Make, Zapier, etc). That leaves self hosted N8N as the only semi-viable option.
N8N is “semi viable” for us, as “normal people” (normal - as in - standard user of MS Excel, no macros, comfortable with simple formulas) actually would need to be able to set up and trigger new AI columns without any help from the IT team. This is not really attainable in the long run with N8N.
It is, however, easily doable with current Baserow AI column and we love it. We just need a bit more control over the output (since we do not use AI as a chatbot, we need as much consistency as possible - default settings do not provide optimal output in our case).
getting back to webhooks - for N8N - If we need to spend a lot of time on IT maintenance per project, then the price gap to Airtable or Excel add-in solutions suddenly narrows to a point, where it could actually make more sense to just use those tools rather than Baserow, when you factor in AI features, platform maturity, UX, workflow efficiency, user ‘satisfaction’ (or rather - lack of user frustration with some bugs or UI), etc.
That is why we’d much rather have the OpenAI temp setting built into the Baserow native AI column.
Like I said before - I’d argue that any use case targeting consistency for AI outputs would benefit a lot even from a simple “temperature” setting. You’d probably make a lot of users very happy with such a setting and save them a lot of time.
