🏃Training your AI
This section provides tools and features related to training and memory for your personal AI.
Last updated
This section provides tools and features related to training and memory for your personal AI.
Last updated
Video Walkthrough
How to Train your AI using Training Spaces: https://www.loom.com/share/47e0ac7190724f87969c7bfc6e7625ec?sid=4b1691e9-03ac-4542-801e-bfb2c952d227
Training by Saving to AI Memory: https://www.loom.com/share/f3916a8df11b4e05ac48d2512aa47da1?sid=6574d85c-b268-4583-b5a1-42224a2a9828
Training by Editing AI Responses: https://www.loom.com/share/7ea44fbe92e046888e778cd525ee0545?sid=c59d62c5-0f97-4437-807c-c7072d52472a
There are two primary ways to train your AI.
Bulk Training by Uploads. This can be done through the WebApp and Desktop app only.
Continuous Training by Messaging. This can be done through the Mobile App and WebApp.
Bulk Training by Uploads
Available on WebApp and Desktop app
Best for Local Documents (pdfs, docx), Websites, and Cloud Files (Google Drive, OneDrive)
Memory Types and Memory Upload (Step 2 and 3 of Training Spaces) is the most efficient way to plan and upload a large number of documents.
The best uploads are
First author content that represents the AI persona you are building - whether this is video transcripts or written blogs
Pdf must be in text format (not image) - the text must be selectable.
Uploads should have Title and Hashtags to give additional contexts to recalling the memories later.
What to avoid:
Avoid including other speakers' language and items. If it's a question answering session, labeling Question: <user prompt> Answer: <your response> is the best format.
Random texts like headers and footers, text that are irrelevant to the main content on websites.
Upload Use Cases:
Documents (pdf, docx, pptx, txt, md) - pdf must be in text format (selectable text)
Transcript (meetings, audio) - must be in first person and avoid including other speaker's languages unless clearly labeled with speaker names at beginning of each turn
Website Urls - Urls must be singular urls that contains the content needed, overarching websites (e.g. nba.com) will not work. Avoid including headers and footers (this can be cleaned up in Documents view later)
Books/Long form text - these must be broken into reasonable chunks (this is usually around 50 pages)
Folder (in Google Drive or OneDrive) - supported files can be uploaded at the same time via folders. Be careful with this one because it will upload all subfiles and subfolders.
You can either configure integrations in training spaces by pressing + on a memory type or you can go to "integrations," on the My AI. If you are trying to upload a Google Drive or Onedrive file to a specific AI profile.
Custom - Custom type of data such as trial transcripts, medical records, must be broken done via specific parsing algorithms and may require additional professional services. Reach out to our team for this.
If you are looking to upload a smaller amount of files or URLs quickly, you can also click on the AI persona you want to upload the given data to, and click the + icon to view all upload options.
Continuous Training by Messaging - Mobile App/WebApp
We recommend two primary ways to quickly and continuously train on Mobile App/WebApp on the go. One way is to send your AI a message and add it to memory. The other way is to edit a copilot draft (from a DM/Channel) and add your edits to your AI's memory.
Best for Adding Single Memories On-the-Go
Best for Editing Single Responses On-the-Go
Memory must be explicitly added by clicking on Save Memory (+) or Swiping right on Mobile App
Saving Memories by Sending Messages in my AI
Mobile (iOS) is the fastest way to saving memory on the go, simply start in My AI and type your complete thoughts (the more details and context the better), send, and then swipe right to add it to stack. On voiceover, double tap on the message and click Stack.
The best memories are:
Detailed and descriptive about the topic or the subject.
Include context such as people, location, absolute time for AI to reference.
Use hashtags to label and recall the specific memory later.
What to avoid:
Avoid adding single question or prompt to memory.
Avoid adding incomplete memory that lacks context like single words or fragmented sentences.
Avoid adding data that cannot be interpreted by AI, such as a few rows in a form. A better way to stack is to have AI reformat into narrative format before adding to memory.
Message Memory Use Cases:
Links: Adding reference links to documents or urls that you and others access often.
FAQ: When other people engage with your AI and asks a question, edit the response to make it accurate to yourself and then add it to memory. Next time someone asks the question, your AI will be able to respond with the updated information.
Journal or TODOs: Use date hashtag #Jan21 to tag short journals or memories on that day. Recall it later by using the same hashtag in the question or prompt.
Desktop similarly can add memory. Next to each longer form messages (both user and AI), there is an option to Add to Memory.
Adding Memories by Editing responses from Inbound Questions
One of the most effective way of teaching and training your AI on knowledge that is relevant to your use case is to open it up to your internal team and close friends for testing.
Start off with your AI in Copilot mode & share your AI with your team by creating a channel
Your AI will generate responses for you as best it can. From here, you can edit responses based on how you would respond, and stack the message.
Once you are confident with your AI, change it to autopilot. Edit and modify responses when needed. This is a great way to train and test your AI continuously.
Check out Suman's autopilot channel to see this use case in action https://s-suman-autopilot-lounge.personal.ai/messaging/profile
