The Checkbox AI Assistant is an intelligent chatbot designed to answer questions from business users and redirect them to relevant apps and documents when appropriate. This guide outlines best practices for setting up sources and instructions to optimize the AI Assistant's performance.
Table of Contents
Preparing Sources
Training Content
When creating training content it’s important to keep it simple and organized. You can have multiple training documents, but each one should only cover a specific subject.
|
Example: Sexual Harassment Policy You want to upload your company's Sexual Harassment policy to the bot so users can seek information on identifying and reporting any sexual harassment at the company. Make your filename as the bot is able to understand filenames and make a connection between the name and the content inside the document. So you can name the file: ACME Sexual Harassment Policy.docx Download our example file at the bottom of this article. |
File Types Supported
In order to create a consistent and easy experience for uploading documents, you can only upload .doc or .docx files (MS Word files). We no longer support PDF files.
You can import PDF files into Word by opening them up in the Word application. MS Word automatically converts PDFs to Word files.
| Supported Elements | Unsupported Elements | Ignored Elements |
| Paragraphs | Formulas | Images |
| Single level ordered and unordered lists | Tables | Footers |
| Document filename | Captions | |
| Headings | Multi column layouts |
Definitions
Supported Elements: explicit support for these and are encourage use.
Unsupported Elements: these are unsupported and should be removed from the doc.
Ignored Elements: these are ignored by the AI and don’t need to be removed.
File Names
Make sure the name of your .doc or .docx file is relevant and clear about the material. The bot takes this into consideration.
✅ Supported:
- ACME Sexual Harassment Policy.docx
- ACME Conflict of Interest Policy.doc
❌ Not Supported:
- ACMEPolicy134-359.docx
- RevisedACMEConflict393.doc
Sections & Structure
Your content should be organized into relevant sections with standard Heading structure (Ex: H1, H2, H3, etc). Avoid having the same or similar content spread out among multiple sections. For example, if you have details about your Conflict of Interest policy, then make sure they’re in the Conflict of Interest headed/sub-headed section and not included in a place in the document that is under the Sexual Harassment policy.
The bot will look at each section header and assume that relevant information will be contained in paragraphs or lists following it.
You can use MS Word’s built in Heading formatting or you can incorporate text that is center-aligned in a single line as a Heading for a section. The bot will treat both instances as “Headings.”
Your document must be in a one column layout without any tables. Remove any tables and convert the information to lists.
Footers are acceptable, but the bot will ignore any information in them. If there is relevant information, then move the content out of the footer and into the main portion of the document.
Hyperlinks and Email Addresses
Any embedded links or email links in Word will be automatically recognized by the bot. Please make sure that you validate the links according to access levels for your users. You don’t want to include links in training documents that may only be accessed by a select group of users. This would be a poor user experience for the users of your bot.
Lists
Unordered (bullet points) and Ordered (A, B, C or 1, 2, 3) Lists are supported normally. However we do not support Nested lists. See example below:
✅ Supported:
According to our Discrimination Policy, you may not discriminate based on:
- Sexual orientation
- Parental status
- National origin
- Pregnancy
- Sexual harassment
- Race, color, and sex
- Reprisal or retaliation
❌ Not Supported:
According to our Discrimination Policy, you may not discriminate based on:
- Appearances
- Race, color, and sex
- Body Type
- Pregnancy
- Status
- Parental status
- National origin
- Sexual orientation
- Sexual harassment
- Reprisal or retaliation
Assistant Instructions
The Assistant Instructions are very sensitive to word choice and language. Avoid telling it what “not” to do. When you give it instructions, you should tell it what it should be doing and how it should handle everything. Giving examples is always helpful.
Here’s an example Instruction set that is focused on making the response visually appealing with design and content:
You are a world-class, insightful AI chatbot responsible for answering questions based on the provided supporting information and instructions.
Respond in a friendly, professional, and conversational tone.
Limitations
While Large Language Models (LLMs) like the one powering the Checkbox AI Assistant are powerful tools, they have certain limitations. Understanding these limitations is crucial for setting realistic expectations and using the AI Assistant effectively.
Areas where LLMs may struggle:
- Mathematical Calculations: LLMs are not designed as calculators. They may make errors in complex or even simple mathematical operations. For accurate calculations, it's best to use dedicated calculators or spreadsheet software.
- Counting and Data Analysis: LLMs don't have the ability to accurately count occurrences of words or phrases across multiple documents. They cannot perform detailed data analysis tasks that require precise counting or statistical calculations.
- Up-to-date Information not in source documents: The knowledge of LLMs is based on their training data, which has a cutoff date. They may not have information about very recent events or changes, unless explicitly provided in source documents.
- Understanding Images or Audio: Checkbox AI Assistant is currently text-based and cannot process or understand images, audio, or video content.
- Remembering Previous Conversations: Each interaction with Checkbox AI Assistant is independent.
FAQ
When to use a separate AI Assistant vs combining everything in one "mega" AI Assistant?
Generally speaking, limited scope AI assistants performs better than broader scope. Broader scope usually means lots of (somewhat unrelated) source documents are uploaded. This makes it harder for us to retrieve the relevant "chunks" of documents to provide to the AI Assistant as background.
Can the AI Assistant ask clarifying questions?
Currently, this behaviour is not built into the AI Assistant. However, the "Assistant Instructions" can be adjusted to promote this behaviour. In addition, the source document should give enough information for the Chatbot to know what information is required. The best case would be if the source document explicitly calls out the datapoints required for a decision.
How can we give the bot a personality?
We've seen "personified" AI Assistants used as a way to drive adoption and improve user experience. You can setup "personified" Ai Assistants by adjusting the "Assistant Instructions". Fun examples include: "speak like Yoda", "pretend you're Santa", "make spooky references" during Halloween, or even just incorporate company values into responses.