Ask YSM: Chatbot Product Overview

Ask YSM is an internally facing chatbot designed to help Yale School of Medicine employees quickly find information from the public pages of the school’s main website.

Overview

• Ask YSM uses all publicly accessible pages on medicine.yale.edu as its knowledge source.

  • The bot references the data as crawled by the Bing search engine to generate responses

  • The chatbot will always include at least one reference at the bottom of each response linking to the page(s) where it pulled the answer from.

  • Content that is protected with CAS or other authentication is not included in AskYSM responses at this time.

• The bot uses a Microsoft platform called Copilot Studio (covered under the University’s MS agreement), which allows configuration for specific topics in addition to general instructions for the chatbot

• Access to Ask YSM requires authentication with a @yale.edu Microsoft account

  • You will not be asked for your credentials if you have previously authenticated with your Microsoft account in the browser where you access the chatbot

Benefits

The Ask YSM chatbot excels at accessing any knowledge that’s available on public pages of the YSM site. It does well with questions like:

• Who is the lead administrator of Orthopaedics?

• Who is performing research using CRISPR?

• How can I get help with data science for my research?

• How do I apply for faculty promotion?

• Where do I find the Powerpoint template?

Ask YSM will respond with a cohesive answer within a few seconds, including references to webpages where further information can be found.

How to use the Ask YSM chatbot

1. Visit MyYSM on the preview website: https://preview.yale.edu/myysm/

2. Open the MyYSM Chatbot (see screenshot below)

3. Sign in with your @yale.edu Microsoft account when prompted.

Open

The Ask YSM icon located at the bottom right of the MyYSM website, click the icon to start interacting with the chatbot

Advice for using the Ask YSM chatbot

Use context with your question

Referencing YSM and putting your question into a full sentence will get you a better answer than a question that is vague or too broad. Consider “Is Heidi the service dog on YSM’s campus today?” instead of “Are there dog events?”

Ask questions about YSM

Referencing the specific area you’d like answers from is most useful. “Does YSM have a policy about branding?” as opposed to “Is there a policy about branding?”

Known issues, limitations, and risks

Ask YSM only includes information on pages visible to the public

Even though a login is required to access Ask YSM, the chatbot only uses information on public webpages when generating responses. This runs the risk of not answering questions from users who are asking about specific content which is only available after logging in with CAS.

Remediation

If we want the chatbot to be able to respond about content that is present only on secured pages of the YSM website, or is not present on our site at all, we can add to the chatbot’s knowledge by adding relevant text that links to the content on question on our site.

For example, Ask YSM does not have access to Yale University policy information. But, if users ask about university policy it will generally provide a link to this page of MyYSM which links to various YSM and Yale University policy resources:
https://medicine.yale.edu/myysm/policies/

Issues with use of Large Language Model AI (LLMs)

There are known issues with relying on an LLM to supply answers for this type of chatbot tool. These include:

Responses are inconsistent

You will find that you can ask the same question repeatedly and get no response, a detailed response, or the same response in different formats.

Time is misunderstood and conceptually mistaken

AI profoundly struggles with the concept of time. It processes information as static, ever-present patterns, therefore it may reference medical information in a news article from three years ago as current medical information.

Context is missing, flattened, or too individualized

LLMs can take information out of context or not provide context with a factual answer (i.e., “the sky is blue” can be a factual answer, but a human knows it is not technically blue, and also does not appear blue at night or during a storm.)

LLMs rely on your previous questions to determine the meaning of your new questions. Therefore, if you ask it questions that aren’t related to the content on YSM web pages, it can assume your new questions are not related to YSM even if they are.

Hallucinations

All LLMs available today have the potential to hallucinate, providing incorrect or nonsensical answers. However, in testing Ask YSM we have found very few hallucinations. The only hallucination problem we encountered in testing was rare, and involved the chatbot returning fake, made-up URLs linked in its responses.

Remediation

We have instructed the chatbot not to include links in the response itself. Instead, it should completely rely on linking to “references”, which are always the correct URLs of pages used to generate the response.

It is still possible those reference URLs could be out of date, e.g. pages recently removed from our site. Refer to the https://ysmweb.atlassian.net/wiki/spaces/HELP/pages/4286119938/Ask+YSM+Chatbot+Product+Overview#Ask-YSM-relies-on-cached-web-data limitation for more info.

Reliance on cached web data provides outdated answers

Like ChatGPT and Claude, Ask YSM relies on cached website data to generate responses. This means if a webpage is updated, that updated information will not immediately be reflected in responses from Ask YSM.

Remediation

1. We are adding IndexNow in our web publishing process. This tool will push changes made on our website to Bing much faster, meaning content used by Ask YSM and for Bing search results will update in 24 hours or less, whereas currently these updates can take weeks. (We expect this functionality in June.)

2. When an immediate fix is needed, we can update the instructions of the bot to prevent discussing topics that result in outdated or otherwise problematic answers. These changes generally take 1-3 hours to apply for all chatbot users on the site (submit a ticket to request this update)

Controversial content and “tricking” the AI

Ask YSM has safety guardrails, but it’s possible that by manipulating the nature of the AI--to be helpful and to provide instructions--it can be tricked into generating harmful, restricted, or forbidden content it is programmed to refuse.

In testing we have not been able to force the bot to make any statement that would be considered problematic, but the possibility exists.

Remediation

We have found that the built-in moderation from Copilot Studio does a thorough job of preventing the bot from making problematic statements. For instance, asking the bot for medical advice may produce some general advice based on YSM news articles or other web content that discusses treatments for diseases and conditions. But it will always include statements such as “Always consult with a healthcare provider before starting any new medication…” or “…it is important to seek medical attention promptly…” depending on the severity of the topic.

Additionally, we have placed a permanent disclaimer on the chatbot interface which reads “This AI tool helps find YSM information but can make errors. Verify answers by reviewing the references.”

If the unwanted response is due to information found on YSM webpages, the most appropriate remediation would be to remove the content from the website, or place the content on a page protected by CAS.

What to do with feedback

If you have feedback to send, use the thumbs up or down graphic at the bottom of the response, provide some detail, and click send:

Open

Alternately, contact us at the YSM Office of Communications Helpdesk.