Why is Replit by default implementing depreciated LLM versions

Agent & Assistant
1

I find it really hard and a bit frustrating having to tell my agent to look for the latest and most relevant models using the API key of a provider. I literally have to go to the sites and feed it the latest model names, while it still tries to connect to GPT-3.5?

3 Likes
2

I understand the frustration with the knowledge cutoff issue.

Here are some workarounds I’ve found effective:

For Agents: When giving instructions, have the Agent read the current model information by using the link reading feature to access: https://platform.openai.com/docs/models

For Assistants: You can prevent model version changes by adding a system prompt in the Assistant Prompt like:

Use OpenAI model "gpt-4o" released in 2024. Do not modify the model name.
2 Likes
3

Known issue with LLM knowledge cutoffs, Agent’s integrations and tools mostly stick up-to-date info into the system prompts though. Can you send
your repl link to me in a DM and I can look at the trace to see where it tried to use GPT-3.5? I just checked the openai integration and it contains the up to date supported models

1 Like
4

Thanks! Is there a list of all the model information, e.g. OpenAI, Anthropic, Gemini, DeepSeek, …?

5

The quickest way to get information about LLM is to visit the official website.
This is because the implementation method varies depending on the company and model.
For example, if you want to use the latest gemini2.0 thinking model, give Agent the following URL

However, as kody-san said, if it is a major model or implementation method, the Agent should already have the knowledge.

2 Likes
6

It would be nice if Replit would build an internal reference document to the latest LLM models, software packages etc. for its agent to reference.

2 Likes
7

We have those in the agent tools, sometimes the models don’t follow them though haha and more “ABSOLUTELY USE THIS, DO NOT USE OTHERS” injects in the prompt is less space for user code

8

I generally go to google ai studio or OpenAI playground copy the basic code template and give it as instruction to the agent to use the syntax for Llm model calls …. And it just works.

1 Like
9

Yes! I did that as well in one case and it went flawlessly (this was for DeepSeek integration.)