The Discourse AI bot utilizes RAG (Retrieval Augmented Generation) for posts, and therefore if the bot has been enabled, each post should have an associated embedding, which is a vector space representation. It would be convenient if these embeddings could be accessed through the API.
As more of us begin to use AI in Discourse, we may desire features that aren’t currently available. Since an AI typically requires post embeddings, which are already present in Discourse sites with the AI bot active, it makes sense to utilize these existing embeddings instead of duplicating them.
I was hoping that the Discourse AI bot could access external data by now, but I haven’t come across any information regarding that capability. Therefore, I’m exploring the creation of a custom GPT for ChatGPT Plus and would like to leverage the pre-existing embeddings.
Even when the API is not available for this specific use-case, you could always create a SQL query with data explorer (tables beginning with ai_ like ai_topic_embeddings) and use API to get the results of the query.
This could use a tutorial on the significance of why this is important, how to construct queries, which seem to be a prompt, and how it could be used to enhance ones Discourse forum.
Also, if one changes the AI model say from OpenAI GPT to Claude, then the embeddings would be different so the impacts of such a change should be noted in a tutorial.
I do hang out on a Zulip forum (Lean) and surely miss the AI features that Discourse has. (Zulip feature matrix)
meaning the feature does not exist.
Hats off to the Discourse staff on augmenting with AI.
IIRC the use of this is free because the cost of embeddings is so cheap now that it is not worth charging. Obviously the Discourse staff need to clarify that statement but I am sure others might be surprised at the answer.
Out of curiosity I tried the URL on a Discourse site where I am just a Trust Level 1 (Basic user) and the query worked.
Since this seems to be enabled and available by default, in the back of my mind I tend to consider if such technology is of benefit to the blind trying to use the internet, this is obviously a win for them when the dots are connected meaning if they are made aware of this then they or others can make use of it to help them.
As such noted this in a topic for the blind on the OpenAI forum.
meaning the feature does not exist.
Obviously the Discourse staff need to clarify that statement but I am sure others might be surprised at the answer.