Role "error" is not accepted by OpenAI, etc.

[michaelwechner]

When an error occurs, then a message is pushed using the role "error".

When the user then sends another message, then for example OpenAI will throw an error, because it does not know the role "error":

Object { message: "Invalid value: 'error'. Supported values are: 'system', 'assistant', 'user', 'function', and 'tool'.", type: "invalid_request_error", param: "messages[2].role", … }

I think error messages should be handled differently in the conversation history, such that this problem does not occur.

What do you think?

[michaelwechner]

I have implemented a quick fix inside src/app/chat-gpt.service.ts ... before sending I replace the role "error" by "assistant"

body['messages'] = this.replaceRoleError(messages);

+

• /**
• • Replace role "error" by "assistant", because most LLMs do not support role "error"
• */
• replaceRoleError(messages: Array): Array {
• //replaceRoleError(messages: Array<{ role: string; content: string }>): Array<{ role: string; content: string }> {
• let messagesToLLM: Array = [];
• console.info("Replace role 'error' by 'assistant' ...");
• for (var msg of messages) {
• console.info("Role: " + msg.role);
• var role = msg.role;
• if (msg.role == "error") {
• role = "assistant";
• }
• messagesToLLM.push({ role : role, content: msg.content });
• }
• return messagesToLLM;
• }

[alenkvakic]

Thanks for your input! I created this project back then when ChatGPT was quite new so I assume some things have changed in their API and the request/responses. As soon as I have time I will update this project to the latest angular version (with the new control flow syntax as well), and then will take a look at how its handled today.

Please feel free to create pull requests if you have fixes in the meantime (like this), as I can add those quicker.