Cleva.Bot - USER GUIDE
Cleva.Bot - Version 2.12
Documentation
Last updated: 30.12.2024
Introduction
Your Cleva.Bot comes pre-trained with knowledge gathered from your website and other data sources to assist your audience effectively. To maximize its potential, periodic training, monitoring, updates, and corrections are essential. These steps help fine-tune your Cleva.Bot and enhance its capabilities.
Your Cleva.Bot is trained by:
- Pointing it to your website or specific URLs
- Uploading files (e.g., PDF, DOC, XLS, CSV, TXT)
- Creating custom Question & Answer pairs
- Adding text-based information
- Correcting inaccurate answers to improve future responses
Actively maintaining your bot will ensure it consistently delivers accurate and meaningful interactions.
About This User Guide
This guide explains how to effectively manage your Cleva.Bot. It is organized as a long-form document that can be explored using the left column menu or by scrolling the page.
Key Concepts & Insights
No technical knowledge or expertise is needed to manage Cleva.Bot. However, you will be able to more confidently evolve your bot’s knowledge base by having a foundational understanding of how your data interacts with AI. This user guide provides step by step instructions and the necessary insights to achieve that aim.
AI Large Language Models:
Large Language Models (LLMs), such as ChatGPT, are designed to generate human-like responses based on the input received and the extensive training data the LLM has been exposed to. It is likely that ChatGPT's default understanding of your business is quite limited, often rudimentary or even non-existent. This is where Cleva.Bot is invaluable.
How Cleva.Bot Works:
Cleva.Bot bridges the gap by transforming your data sources into a format AI can process effectively:
- Vectorization: Your data is converted into numerical representations known as vectors. These vectors are non-human-readable arrays of numbers that capture the meaning behind the information by grouping related data based on semantic similarity.
- Data Handling: Cleva.Bot does not retain or store your original source files. Instead, it maintains a secure vector database. Your original data remains your independent property. It does not exist in Cleva.Bot's vector database in its native form.
What Happens When a User Queries Your Cleva.Bot
- Query Conversion: The user’s question is converted into a vector representation that captures its semantic meaning.
- Semantic Search: Cleva.Bot searches its vector database to find related vectors based on semantic meaning (not keyword matching).
- Contextual Prompting: The retrieved vector data (called an Embedding), is packaged with Cleva.Bot's custom-engineered prompt and sent to ChatGPT as context via OpenAI's API service.
- Response Generation: ChatGPT combines the provided context and uses its vast existing knowledge to generate a detailed and relevant response aligned to the user’s query.
Data Security
Your vectorized data is only shared with ChatGPT during the real-time user interaction to provide the context for the response. The data is not retained by OpenAI or incorporated into ChatGPT’s native training data.
Your vectorized data remains independent, secure and isolated. The non-human-readable nature of vector data is further assurance that any intellectual property contained in the original sources cannot be compromised.
By understanding these principles, you’ll be better equipped to effectively leverage Cleva.Bot for your business needs.
Training on Web Content
Adding URLs for indexing
You have the following options to choose from:
- Scan whole website
- Scan single page/url
- Scan urls from sitemap
Adding new website sources allows your Cleva.Bot to retrieve and utilize information directly from your website. This ensures that your Cleva.Bot can provide accurate and current responses based on the latest content available on your site. Whether it's product details, FAQs, blog posts or general pages, integrating these sources enhances Cleva.Bot's ability to assist users effectively.
Where can I find it?
Sign in to my.cleva.bot and select your Cleva.Bot from the main page.
Select the 'Training' tab then choose 'Websites' from the left column menu.
Click "+ Add new Website" button to add urls.
Scan whole website
In the Address field provide the root url of your website ie. www.yourdomain.com, choose “Whole website” from the right dropdown, then 'Start scanning the website'.
Cleva.Bot will begin scanning while displaying a progress indicator.
Once scanning is completed you will see a nested list of pages/urls (per your website's file structure) including the data-size (character count) for each of the links retrieved.
Review the list and uncheck any items you don’t want included in the training data. Then click “Proceed with Training”
The form will disappear and you will see the list of website sources with their training progress:
Scan single link/url
In the Address field, provide the URL and choose “Single address” from the right dropdown menu, then 'Start scanning the website'
Note: If you neglect to change the dropdown option from "Whole website" to "Single address", the crawler will attempt to gather a list of all your pages. No problem! Simply allow the scan to complete, choose cancel when the list is presented, then start over with "Single address" selected in the dropdown.
The progress bar will appear while scanning and will turn green when completed.
Scan from XML sitemap
In the Sitemap field provide the url of your sitemap and click “Start scanning the website”.
TIP: For most websites your sitemap address will be https://yourdomain.com/sitemap.xml. This will display a list of all your pages. Check that the list is up to date. If your sitemap.xml page only contains links to additional sitemaps (i.e. multiple urls ending in ".xml") then scan each sitemap url by re-running the "Scan from XML sitemap" function for each .xml url. If you do not have a sitemap XML file, or it is not automatically updated, then use the Scan Whole Website function.
Once scanning is completed you will see a nested list of pages/urls (per your website's file structure) including the data-size (character count) for each of the links retrieved.
You can check or uncheck any items. Only checked links will be included for training. Click the button to “Start the Training” to index your selection. The form will disappear and you will see the list of website sources with their training progress/status.
Reindexing/Updating URLs
When you publish new pages/posts on your website (or whenever old pages are removed) remember to update your Cleva.Bot's knowledge base index to ensure it reflects the latest information.
You should also reindex any urls when the contents of the pages have been substantially changed.
To reindex a page, first remove it from the list of saved URLs before adding it again for training.
Where can I find it?
Sign in to my.cleva.bot and select your Cleva.Bot from the main page.
Select the 'Training' tab then choose 'Websites' from the left column menu.
Use the search function to locate the page in the list
Check the page you want to remove and select delete. You can now reindex the page
Click "+Add new Website" button, then follow the Scan Single Page/URL procedure to index the page
Scheduled/Automatic reindexing
If you have urls containing contents that are frequently updated we can implement automated scheduling (daily, weekly or monthly) subject to certain system limits. Contact support@cleva.bot to discuss options.
Train your Cleva.Bot with files
You can upload documents and spreadsheets to add training data to the knowledge base. Accepted file types are .PDF, .TXT, .DOC, .DOCX, CSV, .XLS, .XLSX.
This can enhance your Cleva.Bot's knowledge to provide accurate and relevant responses to user queries.
Where can I find it?
Sign in to my.cleva.bot and select your Cleva.Bot from the main page.
How to add Files for scanning?
Open the Training tab then select "Files" from the left column menu.
Select "+ Add new File" button.
You can include up to 20 files with each upload
How It Works
When you upload a file to your Cleva.Bot's knowledge base, Cleva.Bot processes the document to extract relevant information. This data is then indexed as vector data and made available for your Cleva.Bot to reference when responding to user queries.
Training with Question & Answer Pairs
Question-and-Answer pairs allow you to proactively add hypothetical interactions that will help your Cleva.Bot correctly answer related questions on the topic or in a specified way.
Providing your Cleva.Bot with example interactions is a highly effective, preemptive training method. (To train bot responses using actual chatlog transcripts, see CORRRECTIONS)
Where can I find it?
Sign in to my.cleva.bot and select your Cleva.Bot from the main page.
Select the 'Training' tab then choose 'Questions & Answers' from the left column menu.
Click "+ Add new QA" button to create a new entry.
How to add Questions & Answers?
Type your question in the “Question” field and the answer in the “Answer” field. Cleva.Bot will recognise if the same question was asked in a different way or language and will adjust it’s response.
The “Label” is only visible to you to help manage entries by grouping them in a shared label/topic. Grouping items in a label has no bearing on bot training/knowledge. Using labels, you can organise your Questions & Answers into sets - with a single set having multiple questions and answers under one label. This that can be helpful to navigate through other sets.
What It Is For?
Adding new question-and-answer sources allows your Cleva.Bot to access a broader range of information. This ensures that Cleva.Bot can provide accurate, up-to-date, and contextually relevant answers to user queries. By incorporating various sources, you can enhance Cleva.Bot's ability to understand and respond to a wide array of questions, improving user satisfaction and engagement.
Training with Text entries
Why add text?
Text entries are a convenient way to add information directly into your Cleva.Bot's knowledge base.
This is especially useful for adding information that's not on your website, or to extend deeper knowledge on a certain topic.
Where can I find it?
Sign in to my.cleva.bot and select your Cleva.Bot from the main page.
Select the 'Training' tab then choose 'Text' from the left column menu.
Click "+ Add new Text" button to create a new entry.
How to add text for scanning?
Click on the 'Add new text' button. You will be asked to enter the label and your text data.
How It Works
When you add a text entry to your Cleva.Bot's knowledge base, Cleva.Bot processes the text into vector data based on the semantic information contained. This data is then indexed and made available for Cleva.Bot to reference when responding to user queries.
Markdown for text sources
When Cleva.Bot crawls your website/urls, or imports your uploaded files, all your pages are automatically converted to Markdown syntax. Markdown is a lightweight, versitile and simple-to-write markup language that uses plain-text syntax for formatting. For example, # creates a heading, - makes lists, and text wrapped in ** is bold.
With Markdown you can preserve or create structured documents, including headings, lists, links, images, tables, layouts and code blocks, using easy-to-read and easy-to-write conventions. Markdown is your superpower when creating text sources. This includes "Text Entries", "Question & Answer Pairs"; and "Corrections".
You can learn more about markdown syntax here.
Using markdown is entirely optional. But, here's why it's beneficial, and how you can do it in a snap...
Markdown:
WHY?
"Formatting is to documents what body-language is to human interaction"
Formatting conveys far more than words alone. By presenting context, hierarchy, order, and relationships within a document, formatting enhances the overall fidelity and clarifty of the information and its value as training data.
By using Markdown, your bot will also be able to include clickable links and other Markdown encoded elements in its responses.
Markdown:
HOW?
Markdown is simple to learn. Or, you can ask ChatGPT to do it for you...
Tip/Trick: Format your text using Word, then ask ChatGPT to covert it to Markdown. Easy as that!.
Just COPY/PASTE and you're done!
Knowledge Base Optimizer
Your Cleva.Bot's responses are drawn from the cumulative sum-of-knowledge it has been trained on — or, more accurately, the sum of semantic relationships established in the data.
Pinpointing to a specific training source as being the origin of an answer, or as the definitive reason why a response was formulated in a particular way, is difficult because multiple sources are usually involved.
'Knowledge Base Optimizer' is a powerful tool that allows you to see the relevant sources, weighted by probablility, that are contributing to the particular response or topic.
This allows you to:
- Identify which training data is being utilized in your knowledge base
- Verify that you have the correct and most up-to-date training data
- Make necessary corrections or updates to your training data if discrepancies are found
Using this tool, will assist you to fine-tune your training sources and improve your Cleva.Bots knowledge base to provide the most accurate and relevant information.
Where Can I Find it
From the Training tab select the blue "Knowledge Base Optimizer" button to open the controls.
Enter a question in the message field. The output will list the sources influencing its response.
Analytics
From the top tabs, select Analytics to view user engagement metrics. Select a date range to widen or narrow the reporting scope
Insights
The "Conversation insights" report can be generated weekly to provide deep analysis of user activity. This is AI generated by analysing the conversation history.
Where Can I Find it
From the Analytics page, select "Conversation insights" (top/right).
Once a report has been generated you will be unable to generate a new report for 7 days. This duration is necessary to have sufficient transcript history for meaningful analysis.
The report will list and summarize
- user questions the bot could not answer;
- the most common topics users queried about;
- other key insights and recommendations discovered from the transcript analysis
Chatlogs
Chatlogs provide the full transcript history of conversations between your Cleva.Bot and your users. (If 'Live Chat' is enabled for your Cleva.Bot you will also find the conversation transcripts with your team members included here).
Where can I find it?
Select "Chatlogs" from the main/top tabs. You can browse conversations by date range; search conversations by keyword; or export conversation history as a csv file for further analysis or to import into other systems.
Response corrections
Teaching your bot on its previous errors is an easy and effective way to make the bot better understand your customers and your data.
If your Cleva.Bot gives an incomplete, inadequate or incorrect response, adding a correction will improve the accuracy of future responses on the same topic.
How do I add a correction?
The best way to add a correction is via the transcript in the chatlogs. Find the conversation then click the 
edit icon to the righthand side of the response you want to correct
Enter the response that should have been provided for the query
Tip: Copy/paste the Model response into the correction field to modify it.
How do I manage corrections?
You can manage all your correction from the corrections page.
Where can I find it?
Select "Training" from the main/top tabs then select Corrections from the left side menu
Here you can edit or remove past corrections or add new corrections.
If you add a new correction from scratch, you’ll need to provide a user’s question, an invalid response and a correct response.
(If you are correcting an actual response that was previously provided by your Cleva.Bot, this is best done via the edit button in the Chatlogs transcript
Cleva.Bot UI/User Actions
Your Cleva.Bot has been created to be consistent with your brand and website colours/style. These presets can be modified on request by contacting support@cleva.bot
Cleva.Bot's optional, add-on features have been enabled/disabled based on your use case. Contact support@cleva.bot if you wish to add, modify or remove any of these features.
Deployment
Deploying Cleva.Bot on your website can be done as a widget or an iframe. Widget is the recommended method.
Where can I find it?
From the main/top tabs select "Add to Website" to obtain the required code.
Widget
The widget places the start icon as a fixed object bottom/right of the screen.
To use the widget, paste the code in the head section of the html on all pages that you want the widget to appear. Most management systems will provide a way to deploy this site-wide. If you are unsure how to do this, or you do not have a developer who can assist, then reach out to our support team. We have a depth of experience with most Content Management Sytems and will be happy to help.
Widget example
The widget icon is a fixed object in bottom/right of the screen. The welcome message is also displayed until opened or dismissed.
Clicking the icon opens the conversation window.
iFrame
Use the iframe method if you want to display your Cleva.Bot as part of your webpage content.
Paste the iframe code in your page html wherever you want the iframe to appear. The iframe can be customised to meet the parameters of the content area. Reach out to our support team for assistance and fine-tuning if required.
iFrame example
Integrations
Ecommerce Integrations
Cleva.Bot integrates with both Shopify and with WooCommerce to seamlessly interact with your online store's products and customer data.
These integrations allow users to search for products while chatting with your Cleva.Bot.
With direct access to your product database, Cleva.Bot is able to:
- present your product listings as search results
- provide detailed product information
- advise on availability, colour or style options
- recommend complimentary items
- + other/similar advanced capabilities.
With access to your customer data, users can also query the status of their orders by providing their order identifier.
Talk to our dev team to obtain documentation and instructions if you wish to pursue Shopify or WooCommerce intergation.
(Please note that surcharges may apply for larger scale stores or if custom development or developer assistance is required.)
FAQs
Prompt
Q: How do I access my bot's main prompt?
A: No access is provided to edit the prompt. This is necessary for two reasons:
-
To protect the functional integrity of your bot.
Your prompt has been carefully engineered and tested. Modifications should only be performed by someone experienced in prompt engineering as even seemingly minor modifications can have significant, unintended outcomes. Talk to us if your Cleva.Bot is not performing as intended and we can introduce modifications to remedy those issues.
-
To protect our prompt-engineering intellectual property.
Your Cleva.Bot prompt is a hybrid comprising your use-case, brand and requirements with our proprietary, custom engineered master prompt. The master prompt has been meticulously iterrated and fine-tuned over hundreds of deployments and is a key reason why Cleva.Bot consistently outperforms other, similar chatbot solutions.
Contact Support
