• EN
Question

[Megathread] AI Assist in FME Form 2023.1 - Feedback and Discussions

  • 1 September 2023
  • 30 replies
  • 173 views
Show first post

30 replies

mattwilkie
Rank
Badge +8

The AI Assist dialog is model, and thus blocks access to the rest of FME workbench application. As a consequence you can't, for example, copy an error message from Translation Log in order to feed back to the AI when it's code generated the error.

It also means you have to close the dialog in order to run any code it produces.

mattwilkie
Rank
Badge +8

A big missing feature, which I'm confident is in your roadmap already but I'll say it here to add emphasis:

AI HISTORY.

Currently I need to copy the question, generated code, and explanation to another location such as Onenote or Word in order to refer back to and (attempt to) understand why/how code X came to be. As such it's actually easier to use a 3rd party LLM front end and paste it's results into FME instead.

 

So with that in mind, what are the system prompts being fed to the LLM to prepare it for "all of the following are about FME"?

natalieatsafe
Rank
Badge +6

When using personal API key what are the valid model names? (where do we look these up?)

When I try to use `GPT-4` I'm told it's an invalid request. I have no problem using GPT-4 with other integrations (chatbotui.com and flux.paradigm.xyz for instance).

 

image

Hi @mattw1ilkie​ , thanks for your question. I know when I bring my own key, I usually type the model in lower case (eg. gpt-4) and that works well for me. If you have a look at this OpenAI Models document, under the Continuous Model Upgrades and the GPT-4 sections, you'll notice that OpenAI references their model names in lowercase. For the Model parameter of the web connection, I would suggest trying out one of the models listed in the GPT-4 section of the document linked above, being sure to enter the value exactly as its shown in the document. Just be sure that your API key has permissions for the model that you choose to try.

natalieatsafe
Rank
Badge +6

A big missing feature, which I'm confident is in your roadmap already but I'll say it here to add emphasis:

AI HISTORY.

Currently I need to copy the question, generated code, and explanation to another location such as Onenote or Word in order to refer back to and (attempt to) understand why/how code X came to be. As such it's actually easier to use a 3rd party LLM front end and paste it's results into FME instead.

 

So with that in mind, what are the system prompts being fed to the LLM to prepare it for "all of the following are about FME"?

@mattw1ilkie​ That's a great suggestion, and I've created FMEFORM-29591 to request this AI History functionality.

 

At this time, we are not publishing our system prompts for the AI Assist tool. This tool is still in technical preview status, and so its various components are subject to change as we test and refine them.

mattwilkie
Rank
Badge +8

A big missing feature, which I'm confident is in your roadmap already but I'll say it here to add emphasis:

AI HISTORY.

Currently I need to copy the question, generated code, and explanation to another location such as Onenote or Word in order to refer back to and (attempt to) understand why/how code X came to be. As such it's actually easier to use a 3rd party LLM front end and paste it's results into FME instead.

 

So with that in mind, what are the system prompts being fed to the LLM to prepare it for "all of the following are about FME"?

An example session of using external chat agent instead of embedded AI Assistant.

https://gist.githubusercontent.com/maphew/aa958ad2f2116e292b4e0e3ad2ee771d/raw/ecbc23a936383c5ef00e84dde2d477083bc5a4a8/a-session-output.html" target="_blank">https://htmlpreview.github.io/?https://gist.githubusercontent.com/maphew/aa958ad2f2116e292b4e0e3ad2ee771d/raw/ecbc23a936383c5ef00e84dde2d477083bc5a4a8/a-session-output.html

 

Goal was to write and FME PythonCaller function to report on python environment. Elapsed time was ~30-45 minutes. Output html is bit garbage since ChatbotUI doesn't have a share feature yet. The session export json and script to convert is also attached if curious. And the working PythonCaller code too of course.

File results at https://gist.github.com/maphew/aa958ad2f2116e292b4e0e3ad2ee771d

 

Comparing the time spent and resulting progress I experienced using AI Assistant and Chatbot, using chatbot is 20 to 30% more efficient for me (but would be challenging if I only had a single display).

Reply


Community Stats

28,807
Posts
109,814
Replies
36,623
Members
Terms & ConditionsCookie settings
A product of
Safe Software respectfully acknowledges that we live, learn and work on the traditional and unceded territories of the Kwantlen, Katzie, and Semiahmoo First Nations.