Why "Ask A Question"?

We’ve spent the last 8 months or so building Ask-a-Question, an AI question-answering service for direct-to-citizen helplines. We decided not to create yet-another-RAG solution. Ethan Mollick summarizes our concerns well. In short, the tech for a cheap, scalable, and guaranteed error-free AI is not there - the kind that you’d want to roll out to citizens in high risk and high trust use-cases like health. Instead, we wanted to lean on our experience building a question-answering for MomConnect to build something that is (a) trustworthy; and (b) provides actionable insights to support continuous learning and improvement.

Trustworthy

By this we mean a hundred percent accurate - zero hallucinations and no misleading responses.

Not everybody needs these things, especially since it’s not completely costless - more on that below. We understand that. And for those people, there are numerous RAG-in-a-box solutions out there on the market. But for users where trust is critical, like governments or orgs providing health or legal advice to citizens in need, accuracy is non-negotiable.

Here’s how we do this.

First, we ask our users to curate their content. Instead of dropping the kitchen sink of documents in a typical RAG, you convert your documents into content cards. No superfluous, irrelevant, or inaccurate content is added to the system. Yes, it is more work than just uploading documents but in return you get finer control over the content that is actually sent to your user.

Second, we argue that searching for the right content is what your users are actually looking for. AAQ finds and returns the top 5 most relevant content cards. By sending back exactly the content in the cards, and not getting an LLM to construct the final answer, we guarantee 100% safety - no more stressing about jailbreaks and prompt injections - and zero hallucinations.

If you think you definitely need an LLM to create a summarized custom answer for each user, and are comfortable with non-zero hallucinations, you can turn on LLM summaries by changing just one parameter. And even in this case, we include guardrails to ensure high confidence in the answers being returned.

Actionable Insights

We believe that orgs who are committed to providing a high-quality direct-to-citizen helpline are continuously looking to improve their content and the accuracy of their service. And for that you need insights into how your users are engaging with your service.

One of the benefits of using content cards is you can get feedback on the individual content pieces. Not only can users tell you if your search quality is poor, they can also tell you if a content card is outdated or confusing. We created a dashboard for you to deeply engage with this user feedback. We use LLMs to summarize feedback from your users, identify popular and emerging topics, and highlight missing content. The dashboard helps you understand what content is most or least popular and how the performance of your solution is changing over time.

MomConnect use case

NDOH and Reach are committed to providing an accurate question-answering service as part of their MomConnect platform in South Africa. Citizens accessing this service are usually new and expecting mothers who may not have access to other reliable healthcare advice. Hallucinations can not only be dangerous and potentially deadly, they would be embarrassing for the government and erode trust in a platform used by millions of women.

NDOH and Reach created content cards for the most frequently asked questions. These were vetted by healthcare professionals and approved by the government. Citizens are now able to ask questions about their and their baby’s health and get the best matching content from this database.

A dashboard shows when these users engage with the service, and what topics are most frequently accessed. You can monitor the performance of each content card - improving popular ones further or refining those that get negative feedback.

Try it out for free

Beta is now available to try for free. Go to app.ask-a-question.com and sign up with your google account to try it out!