Soupy (soupy-solr.py) now has all of the following functions:
- Soupy has a rudimentary memory in the form of JSON databases.
- This memory is somewhat rudimentary. It works quite well, but it is pretty simple. It relies on keywords that are generated by ChatGPT based on messages it receives from chat history. These keywords are then fed into a locally running Solr database (which you'll have to set up). Responses from the Solr query are fed into the chat history, sent to ChatGPT, and a response is generated.
- Chat with soupy by @tagging it in any channel that it can access.
- Specify a specific channel where soupy responds to all messages.
- !generate an image based on user input.
- !transform an image based on user input.
- Describe the image attached to a message, just @tag the bot in the message. You can optionally give instructions for the image description, to make it more or less descriptive, and so on.
- !whattime will return the time in that city. For example, "!whattime Belgium" will return the time in Belgium.
- WORKS BEST WITH GPT-4o. Other versions of GPT are fine, but 4o is highly, highly recommended.
This is a major update for soupy. The new updates are contained in soupy-solr.py. In a nutshell, all of the below functionality works, plus some new stuff that you'll want to get set up. There are many small updates, but the main update is as follows:
- Soupy has a very rudimentary "memory" in the form of a Solr index based on a JSON database.
- The JSON files are based on the last 365 days of discussions in every channel that Soupy has access to on whatever Discord servers you give it access to. These JSON files will be stored in whatever directory soupy-solr.py is stored. Depending on the size of your message history, this could take several minutes, or even an hour, the first time you run it.
- Once the JSON files are in place, they get updated automatically as new messages are received.
- The long-memory is fairly rudimentary, but with the proper pre-prompting, it works pretty well.
- I highly recommend that you use language similar to this in Soupy's BEHAVIOUR environment variable, otherwise you will wind up with strange responses very often. This is because of the way that ChatGPT is going to interpret the history that you're sending to it:
"BEHAVIOUR="You are a helpful Discord chatbot named Soupy and you do have the ability to recall past interactions and conversations. Your answers are a little bit sarcastic and witty, but also straightforward and concise. You give human-like answers rather than lists as your responses. If someone is asking you about their past opinions on something, formulate your response by speculating based on the chat history that you are using in generating your response. Be conversational. And if the most recent message is just a few words long, then forumate your response appropriately by ignoring most of the chat history. Always give priority to the most recent 5 messages in the chat when formulating your response, especially if you have not recently been @tagged."
In order to get soupy-solr.py up and running, you must first install Solr. I'll leave that up to you to figure out. It's not that hard, I hope.
Once Solr is installed, create a new core with solr create -c soupy
Once installed, Solr must be properly configured with certain fields.
<field name="id" type="string" indexed="true" stored="true" required="true" multiValued="false"/>
<field name="username" type="string" indexed="true" stored="true"/>
<field name="content" type="text_general" indexed="true" stored="true"/>
<field name="timestamp" type="pdate" indexed="true" stored="true"/>
Once all this is complete, it works pretty darn consistently. It is likely that soupy-solr.py won't run at all without solr properly in place.
There are a couple of absolute directory locations. Change it to as you please. Search for scriptlocation.
The first time you successfully run soupy-solr.py after having Solr properly configured, your server's chat history will be indexed into local JSON files by the script. This takes a while. It might take 5 minutes, or over an hour, depending on the number of channels you're indexing, how many days of history you're indexing, and so on. After that initial index, loading is significantly faster.
Pretty big update on 12/17.
Soupy does more stuff now.
- Chat with soupy by @tagging it in any channel that it can access.
- Specify a specific channel where soupy responds to all messages.
- !generate an image based on user input.
- !transform an image based on user input.
- Describe the image attached to a message.
A generated image:
The instruction to transform the image:
The transformed image:
An image description:
Note: These are all from the channel in which soupy responds to all messages. That's an environment variable.
To describe an image, simply attach it to a message and @tag the bot.
Pretty big update on 12/10/2023 -- soupy can now analyze images sent to the channel. It will also follow instructions on how you want the image to be analyzed. For example, "describe this as a 5 year old would describe it" or "give a poetic description of this image."
To use this functionality, just tag the bot in a message with an attached image, and optionally give an instruction about how you would like the image to be analyzed.
Soupy-combined combines the functionality of the chat bot with the functionality of the image generation bot.
Currently, soupy-combined.py uses the chat model specified in the .env, and the image !generate uses a hard-coded model which IS NOT referenced in the .env.
- Install dependencies.
- Create a .env and populate it.
- Have fun.