FAQs: Training your AI

Contents:

• Training success
• Knowledge
• Communication
• Troubleshooting
• Next steps

Training success

Q. How can I create the best AI version of myself?

A. There are a few preparatory steps that will help you make an awesome AI.

1. Define your approach - Think about what style of coach or mentor you'd like your AI to be. Get clear on how you want your AI to help your audience and you. If you're unsure about your AI use case, take our quick quiz to help you decide. Then define your mission!

2. Select your best content - Gather together your most helpful and relevant existing content. Focus on materials that share your wisdom, convey your unique approach, and help solve your clients' specific problems. As you work through the training rooms, you'll see where to add all this brilliant information!
3. Clear some time - It doesn't take long to make your AI. You could be completely set up within 60 minutes. But get intentional. Book a slot of time in your calendar, ready to get it done!

Q. What types of content work best for training my AI?

A. The best content for AI training includes articles, book chapters, website copy, blog posts, social media posts, course materials, podcast episodes, scripted videos and interviews.

Focus on your clearest, most relevant content that shows your expertise and beliefs, and addresses your audience's common questions.

Q. Can I upload books and lengthy content?

A. Yes. Long-form content can be uploaded in sections and is handled well when broken down.

Q. What content should I avoid uploading to train my AI?

A. It's best to avoid outdated or irrelevant material that no longer reflects your current approach, as well as content that doesn't address your audience's common questions. Unedited video or podcast transcripts should be cleaned up before uploading. Steer clear of material that duplicates what you've already uploaded, very complex information, data in tables, or items in unusual formats. Slides with lots of bullet points and images can also be challenging for your AI to interpret.

Q. Can I train my AI on content written by others?

A. Only if you own the rights or have permission to use their content.

Public domain ideas and general concepts are fine. You can also mention people and approaches that have inspired your work - just credit them appropriately and signpost your audience to their resources.

The best training content is your own take on topics. Your experiences, opinions and frameworks are what make your AI sound like you.

Q. How do I know what topics to train my AI on?

A. Focus on your audience's core needs. You could ask them directly about their challenges, run "ask me anything" sessions, and review what early testers discuss with your AI.

Q. The setup sections on my dashboard now have icons when they previously had the numbers 1, 2, 3. How come they've changed?

A. They change when you've trained your AI a certain percentage. This shows you've made great progress with your training :)

Knowledge

Q. What are the different ways I can add training data to my AI?

A. There are three core ways to add training data to your AI:

1. Upload documents, text or web page copy to knowledge upload - This automatically generates prompt and completion pairs from your content. Adding documents is ideal for quickly adding large amounts of information from PDFs, Word documents, or text files. You can also copy in plain text or the URL for a web page. Note: the documents themselves are not stored, just the generated P&Cs.
2. Upload a CSV file to your training data - If you've prepared a spreadsheet of prompts and completions, you can upload this directly to your training data. This can give you more control over your AI's data, as you can edit it first.
3. Manually add to your training data - You can add and edit individual prompt and completion pairs one at a time in the training data room. This is perfect for filling specific gaps or adding particular questions and answers. You can also add P&Cs when reviewing client chats.

You could do all your training very efficiently using method 1, but for best results, use a combination of these methods to build comprehensive training data.

Q. What's the difference between the 'knowledge upload' and 'training data' rooms?

A. Knowledge upload is where you upload documents and text with your content and Coachvox automatically turns these into prompt and completion (P&C) pairs. You can upload PDFs, Word docs, or text files. Training data is more for reference. This is where you see those generated prompts and completions. You can also upload a CSV file of pre-prepared P&Cs here and make manual edits and additions.

Q. How can I ensure my AI formats its answers in the best way?

A. Your AI supports standard text formatting, including bold, italics, bullet points and numbering. You can include examples in your training data demonstrating the formatting you'd like to use. However, this isn't essential as your AI will present all its answers in the best way possible.

Q. Will my AI draw on other knowledge in addition to the data I've uploaded to respond?

A. Yes, its default is to use your training data and to also draw on its wider AI knowledge and communication skills to be as helpful as possible.

Q: Can my AI search the internet?

A: We use AI technology that's trained on data available on the internet and much more. However, it can't specifically search the internet live to gather real-time or very recent news, updates and information. Maybe think of your AI as a supportive mentor, rather than a search engine.

Q. Can I restrict my AI to only use my training data?

A. Yes, you can toggle on "use only my training data" in the engagement training room. Most of the time, it's better to leave this off, but it may be relevant in particular situations. For example, if your AI is trained on highly specific approaches and opinions, particularly if these differ from popular ones. If you wish to select this setting, make sure you've fully trained your AI first. Otherwise, it may not have enough training data to do this effectively.

Your AI will always aim to be helpful in your style, no matter what someone asks. But this setting will keep it focused on your topics!

Q. I'm concerned about my AI answering questions I haven't trained it for. How can I minimise my concerns?

A. Your AI will always try to be helpful no matter what someone asks, and we find our AIs generally handle unexpected questions pretty well! Here are some steps you can take:

• Test your AI yourself in the fine-tuning room with questions you haven't trained it on - have a bit of fun with it and be as challenging as you like!
• Add a disclaimer on your AI's share page or landing page.
• Tell your first users it's in alpha mode and ask for their feedback.
• Review client chats regularly and add training data for any questions your AI struggled to answer.

Q. How much training data do I need for a good AI?

A. There isn't a definitive ideal number; quality matters more than quantity. We can give you a rough guide, as generally the amount relates to the use case. However, it's best to focus on ensuring your AI is trained on your core areas of expertise and the topics most important to your target audience, along with any essential information about you, your approach and services.

• Lead generation – Around 500 prompt and completion pairs should cover your expertise, services, and the questions prospects typically ask.
• Client support – Around 1,000 to 2,000 pairs is likely to cover your core topics and address common client queries.
• Multi-topic coaching – If you cover several areas of expertise, aim for up to 3,000 pairs to ensure good coverage across all topics. More than this is okay, but the main thing is to focus on content that addresses your target audience's needs, rather than uploading everything you've ever created.

This may sound like a lot, but your training data will rapidly increase as you upload documents using our knowledge upload tool.

Q. What are the reasons for not uploading every single piece of content I've created?

A. Reasons not to upload everything include:

• Duplication – when the AI sees the same ideas and answers written in different ways, it may not know which to prioritise.
• Irrelevant content – older material might not reflect your current messaging or approach.
• Dilution – too much information can make your AI less focused or inconsistent.
• Extra maintenance – the more content you upload, the more there is to review or update later if information changes.

Q. How many files can I upload to the knowledge upload room?

A. In a free trial period, this is limited to 30 files. This count includes any that have been deleted. If you want to begin your subscription early to lift the limit, just visit your settings page and click 'Graduate my trial now' under your plan details.

On a paid subscription, the upload limit is set to 100, but this can be increased whenever needed. Just get in contact, and we'll edit this. Overall, we recommend focusing on quality over quantity for the best results.

Q. What's the maximum size for files I can upload?

A. Each file or piece of text content can contain up to 100,000 characters and files can be up to 25MB. If your document is longer than this, split it into smaller sections or chapters before uploading. This will help you with managing and reviewing the training data as you go.

Q. Can I view the documents I've uploaded to my AI?

A. The original files themselves are not stored, so you can't open or view them. The files are immediately turned into training data, and you see a log of the uploads for reference.

Q. How long can prompts and completions be?

A. Each combined prompt and completion pair in your training data can be up to 2,000 characters in total. This is to optimise your AI's performance. Your AI can actually give longer answers within the chat if relevant. Should a prompt and completion pair you're preparing or editing go over the character count, break the information into multiple shorter ones.

Q. Can my AI share files and links?

A. Your AI can't share files, but it can include links to web pages, online videos, file downloads and other resources within its text responses. Make sure to add the full links, including the https:// part, to your AI's training data.

Q. Do I need to generate a label for each prompt?

A. No, labels are optional. They're just for you to keep your data organised and help you find information later. Your AI doesn't infer meaning from them, so there's no need to add labels unless you find them helpful.

Q. Do I need to remove duplicate training data manually?

A. Your AI can handle some duplicate training lines. The main thing to avoid is duplicate prompts with contradictory answers. It's worth doing occasional spot checks when adding new items by searching for key terms to see what you already have on those topics.

Q. By training my model, am I also training other Coachvox AI models?

A. No, your model is completely discrete. The information you provide during training is used only for your model and is not used to instruct or develop any other Coachvox AI models. We may use your training data to help improve our processes and technology infrastructure to continue enhancing Coachvox AI as a product, but it won't be used for other AIs.

Q. What should I do if I have multiple different answers saved for the same starter prompt?

A. Your AI finds the closest match but doesn't prioritise between identical prompts. If you're happy with all the answers, that's fine. However, if some responses are outdated or no longer apply, remove these from your training data to avoid confusion.

Q. How do I train my AI to share multiple meeting booking/calendar links?

A. There are a few methods. As well as, or instead of, selecting to optimise your AI to 'Generate inquiries' in the AI approach section and adding one booking link here, train your AI to share different links in different scenarios. Here's a short video on this topic, with some prompts you can use to make some training data. Make sure the prompts and completions are specific enough for the AI to work out which link to share when.

Alternatively, you could add one booking link that goes to a page with a menu of options.

Q. How do I train my AI not to give guidance on specific topics?

A. There are a few ways to handle this:

1 State this clearly in your coaching methodology - In the engagement training room (under advanced coach mode), explain any boundaries in your approach. For example, you might state that you discuss finance but don’t give financial or legal advice. This sets a clear foundation for how your AI should operate.

2 Train your AI how to respond to out-of-scope topics - Add prompts with likely questions about topics you don’t want to cover, then write completions showing how you’d like your AI to reply. For example:

“I'm not trained to answer questions about [topic], but I can support you with [x, y, z].”

“[Topic] isn’t my area of expertise, but I’ve found this resource helpful: [name/link].”

“I can’t advise on [topic], but tell me more about what this means for you?”

Add these directly into your training data. This keeps your AI helpful and professional while maintaining clear boundaries.

3. Set your AI to use only your content - Toggle on this option in your AI settings. When enabled, your AI relies only on its training data. It will typically say it’s not trained on a topic if it hasn’t been covered, whilst continuing to always be as helpful as possible in your style.

In most cases, the first two steps are enough. It’s often better not to limit your AI’s wider knowledge unless you have a specific reason to do so.

Q. What happens if my AI gives incorrect information?

A. Neither you nor Coachvox are liable for any errors in AI responses. Users are informed that AI can make mistakes. We recommend adding a disclaimer next to your AI chat when placing this on your website.

Q. How do I delete training data?

A. To delete all the prompts and completions created by an uploaded document, head to knowledge upload, hover over the row for that file and click the 'X' button.

To delete just specific prompts and completions, head to training data, hover over a row, then click the 'X' button to delete that P&C pair. Or, select a number of rows with the tick box and then click the 'delete' button to remove them all at once.

Q. When training my AI, am I also training other Coachvox AI models and OpenAI?

A. No, your AI model is completely discrete. The information you provide during training is used only for your AI and is not used to instruct or develop any other Coachvox AI model or OpenAI models.

Q. Can I keep training my AI?

A. Yes! You can update and add to your AI's training data whenever you like.

Add new content as your coaching business grows, check client chat transcripts to spot gaps, and update anything outdated. Learn more in our guide to managing your training data.

Q. What's the best way to anonymise content, such as coaching transcripts, I wish to use?

A. To avoid having to share the unedited data with any third-party tools, we find that opening the transcript document and using "find and replace" to replace any individual names or company names usually works well.

Communication

Q. How do I get my AI to sound more like me?

A. To make your AI sound more authentically like you:

• Update your "personality and tone of voice" field within the style and tone training room with specific details about how you communicate. This has the second biggest influence on your AI's voice after your training data.
• Adjust the style sliders to match your coaching or mentoring approach.
• Add your common phrases, greetings and sign-offs in the engagement training room.
• Add a description of your coaching methodology in the engagement training room.
• Upload content where you're speaking naturally, such as interview transcripts or podcast episodes.
• Use fine-tuning to rate and edit responses, teaching your AI exactly how you'd phrase things.
• Review client chats and add particularly good examples to your training data.

Remember that getting your AI's voice right often takes several rounds of testing and refining.

Q. How do I set my AI to take a non-directive, coaching approach?

A. To create a non-directive coaching AI:

• Set the style slider towards the coaching side in the style and tone training room. You might also want to combine this with setting your AI's responses to be shorter.
• Include questions in your completions rather than just statements.
• Specifically, write that you take a non-directive approach in your "personality and tone of voice" field.
• Upload examples of coaching conversations showing your questioning technique.

Q. Can I train my AI to speak in languages other than English?

A. Yes! Your AI can communicate in more than 50 languages and will automatically respond in the language your client uses to start the conversation. For best results, upload some training materials in the languages you expect your clients to use.

The voice input feature also works in English, German, French and Spanish.

Additionally, you can set your AI to use British or American English spellings and grammar in the style & tone training room.

Q. Can I switch the Coachvox interface into another language?

A. There isn’t a language switch for the chat interface, but Coachvox can absolutely be used in many different languages and we work with coaches around the globe.

You can train your AI in your chosen language, and it will respond in the language that the user starts the conversation in.

You can also embed your AI chat into your own landing page, which allows you to customise the overall experience in your chosen language. Most web browsers will also automatically translate the interface into the user’s preferred language.

Q. Can I train my AI to give a custom welcome or onboarding experience?

A. There are a few ways you can create a tailored welcome experience:

• Prompts & completions - Add the prompt “Hi, my name is [name]” to your training data and write a completion with some initial onboarding questions you want your AI to ask.
• Custom phrases – toggle this on under 'advanced personalisation' in the style and tone room, then add greetings and starter phrases that include your introduction and some onboarding questions.
• Starter prompts – add these in the engagement room. Word them as if the client is saying them, and train your AI to respond with follow-up questions.
• Create a specific landing page for your AI and add guidance on how clients can get the most from their conversations.
• Set up automated follow-up emails for new AI users by connecting your Coachvox account to your CRM/mailing system via Zapier, as explained here.

Fine tuning

Q. What is fine-tuning for?

A. Fine-tuning is where you test, rate and improve your AI's responses. Chat with it as a client would, rate each reply, and edit any that need work - your improved version then gets saved to your AI's training data. Focus on one topic or theme at a time, starting a fresh chat in between. Return to it regularly, especially after adding new training data.

Q. What's the best approach for fine-tuning sessions?

A. Focus on one topic area at a time and ask questions as your clients would, using the language and phrases they use when speaking with you. Start a fresh conversation at any point by clicking the 'new chat' arrow icon to test different scenarios and after any further training.

Q. Is there any difference between rating a completion as 1, 2, 3 or 4 stars?

A. There is no difference in the fine-tuning room itself or for your AI's overall knowledge. Any rating lower than 5 stars will give you the option to edit the completion and save a new version. The difference in rating contributes to the overall training rating in your dashboard, which is just for you to track progress.

Q. How do I check my fine-tuning history?

A. Toggle on fine-tuning history to see the prompts and completions you've added through this method. Look for prompts where you've saved multiple different completions that might contradict each other, and remove any that are no longer relevant.

Q. Can I edit my fine-tuning prompts and completions?

A. Yes, view the fine-tuning history table then click on any prompt or correction to make edits. After editing the text, click anywhere outside of the text box and your changes will be saved straight away.

Q. Can I delete responses I've saved in fine-tuning?

A. Yes, any prompts and completions you add when fine-tuning will appear in your fine-tuning history, and you can delete them at any time.

Q. Can I export my fine-tuning history?

A. Yes, you can download all or selected rows from your fine-tuning history by clicking the download button at the bottom of the table.

Q. How often should I complete fine-tuning?

A. Return regularly to spot check your AI, especially after adding new training data or editing information and settings.

Q. Sometimes my last chat shows in fine-tuning, sometimes it doesn't - why is this?

A. We automatically close conversations after 1 week to prevent too many topics and inactive conversations. If you're not seeing the previous chat in fine-tuning, it's likely that it's been over a week since your last conversation or you've pressed the chat refresh button. When testing your own AI, it's best to start afresh anyway to ensure your AI uses the most up-to-date training data. 

Training troubleshooting

Q. My AI is not responding as I'd like - what do I do?

A. If your AI isn't giving the responses you want, try these steps:

• Check you've clicked 'Save' at the bottom of the training data page - this is essential for updating your AI's knowledge after you've made any edits.
• Search your training data and fine-tuning history for similar phrases to check for conflicting information.
• Make sure you've trained your AI on your ideal answers to your starter prompts.
• Add more training data on the specific topics where your AI is struggling.
• Add or edit a description of your coaching methodology in the engagement training room.
• Use the fine-tuning room to test, rate and edit responses to teach your AI how to improve.
• Review your AI's style and tone settings to make sure they match your desired approach.
• Look at the 'advanced personalisation' section in style and tone to update your conversation phrases.

For specific topics, you can also add examples directly from client chats to your training data.

Q. I can't upload my book, the system says it's too long. What should I do?

A. We recommend uploading one or a few chapters at a time. Not only does this make it possible to upload, but it also makes it easier for you to review the training data produced by each part.

Q. If a document gets stuck in 'processing' in knowledge upload, what should I do?

A. The system will try uploading the document for 1 hour. If it's still not processed after this time, hover over the file and click the 'X' button to delete it, then try again. If it fails on further attempts, contact us, sharing any error messages you see, and we'll look into it. (Note: You will need to wait up to 1 hour until the deletion option appears.)

Q. Why won't my CSV file upload?

A. Check where you are trying to upload the file. CSV files of prepared prompts and completions need uploading directly to your training data room, rather than knowledge upload like other documents.

Q. When I upload a CSV file to my training data, I get a message saying the prompt/completion field is missing.

A. Check your spreadsheet. If there are any rows above the column labels, remove these. Retype each column label (prompt, completion, labels) and resave as a CSV. Then try uploading it again.

Q. I'm seeing incorrect symbols in my CSV file of P&Cs. How do I fix this?

A. If you're seeing odd characters like ô, Äô, or ? instead of punctuation marks, the issue is usually due to character encoding. CSV files need to be saved using UTF-8 encoding to display special characters correctly. Here's how to fix this:

On a Mac: Open the CSV file in Numbers. Go to File > Export To > CSV. Save the file (Numbers automatically saves in UTF-8)

On Windows (Excel): Open Excel and go to Data > Get Data > From File > From Text/CSV. Select your file and choose 65001: Unicode (UTF-8) as the file origin if prompted. Load the data.

This ensures special characters display properly.

Q. My AI is sharing an incorrect link (URL) repeatedly. How do I stop this?

A. If the AI keeps sharing an incorrect link, it indicates that it must be there somewhere in the data. Resolving this is just a case of tracking it down, then editing or deleting it. Search in these two places:

• Training data – Type all or part of the URL into the search bar, or export the data and search the CSV file. If the link appears here, either edit the completion to update it with your preferred link or remove the entire prompt and completion row completely.
• Fine-tuning – Toggle on fine-tuning history and again search for the URL to check if any completions contain the incorrect link. If it appears here, delete or edit the prompt and correction.

This should ensure your AI no longer uses the outdated link and always directs people to the right place 🙂

Next steps

Q. How do I know when my AI is trained and ready to share?

A. Your dashboard includes a readiness score that shows how prepared your AI is to represent your style and knowledge. This is calculated based on the amount of training data you've uploaded and how complete the other training rooms are.

The more you train your AI, the better it gets, though it will never be perfect (just like us humans!). Your dashboard will give you some pointers on how to improve your score.

We suggest aiming for at least 75% as a benchmark, but the right number depends on your use case, and you’ll be the best judge of when it feels client-ready!

Once you've completed the core training, test your AI out in the fine-tuning room with questions and statements your audience is likely to say.

Next, we recommend running an alpha test to gather feedback from a small number of people before officially launching your AI more widely.

Check out our guidance on sharing your AI for more next steps.

We're here to support your AI journey. If you can't find what you need on this page, check out our other FAQs and support resources.