Deliver personalized advice based on users input using OpenAI API
We need to be prepared for the future:
AI (Artificial Intelligence)
Over the past year, AI technology has seen an immense boost, with the development of cutting-edge tools such as DALL-E, Midjourney, ChatGPT, and VALL-E. In the years to come, AI will become a core feature of a vast majority of digital products, if not all. Thanks to Natural Language Processing (NLP) APIs, such as OpenAI, it is now possible to bring these innovative experiences to life.
Mental health for Bloom means offering personalized content
We have always dreamt of creating a personalized mental health experience for our users, where they can finally make their wishes come true: an app where the content responds directly to their text input, adapts according to their unique needs and where they track their mental health progress over time.
With Bloom, users benefit from a vast library of content, including single video-based therapy sessions, quick and easy daily exercises, and programs to delve deeply into a single topic. Our users appreciate the wide variety of content Bloom provides.
How can we improve the users experience on the app?
We encountered a problem: although the content library provides plenthy of resources, it has its limitation when it comes to personalizing the experience to the user's individual needs.
We asked ourselves:
How can Bloom provide an immediate solution to a user's unique situation?
How can Bloom offer the most personalized experience tailored to the user's specific needs?
Programs
5 Days structured programs to dive deeper on a specific topic with sessions and activities.
Single sessions
10 min self-guided therapy sessions, from burnout to anxiety, confidence to self-esteem, to inspiration and motivation.
Habits
Daily quick exercises to boost self-care & therapy practices.
AI Use Case suits Bloom: Content Generation
AI has become increasingly advanced, often outperforming humans when it comes to content generation. For this reason, Bloom is utilizing AI to deliver digital therapies to their users with remarkable success.
From a business standpoint, this will not only help us to stay ahead of the competition and establish our product as a leader in AI-driven content generation, but it also provides us with invaluable insight and knowledge that could be incredibly advantageous down the line.
How to use Chat GPT for content creation
Why prompt engineering is a UX and domain experts effort?
Prompt engineering* requires a collaborative effort between UX and domain experts, such as head of AI and therapists (in this specific case) , because it involves using insights from these disciplines in order to create an effective and user-friendly product. These professionals must work together to create a product that meets the needs of the users, explore different case scenario, while meeting design standards.
UX design plays an integral role by identifying what information is valuable to users, exploring options to present the output (e.g., as graphs, text, audio or visuals), and selecting the most effective way of displaying it to users.
Prompt engineering skills
Language
Communication
Creativity
Critical Thinking
Domain expertise
*developing, refining and optimizing AI-generated text prompts to ensure they are accurate, engaging and relevant for various applications.
AI feature ideation & development
for Bloom
To ensure that our feature is valuable to our users and is seamlessly integrated into the app, we need a way to incorporate AI into our daily core flow. This includes data tracking such as mood, emotions, and symptoms, as well as journaling. We have crafted an experience that provides immediate AI-generated insights and AI-generated advice based on users' input data.
Feature 1: AI Daily Insights
The Daily Insights provides users with detailed insights into their mental health by analyzing their journal entries and mapping out the data into intuitive metrics. The scales - which are available for immediate visualization - allow users to quickly gain an understanding of their mindset and pinpoint relevant for self-improvement. This ongoing feedback loop facilitates more conscious decisions in their day-to-day lives.
Structured output (chart data)
We take user text input and instruct OpenAI to analyze it based on criteria related to mental health, delivering the resulting output in JSON format, which is a structured text data.
Feature 2: AI Advice
The Today’s Advice feature provides users with tailored Cognitive Behavioral Therapy (CBT) advice based on users data tracking and journals. When users completes the core flow (data tracking and journaling), they can access the feature and get the AI Advice.
Prompt engineering, Prompt chaining
& Few shot prompting
We crafted a prompt to create a short and digestible piece of advice to find quick relief. Through trial and error with prompt engineering, we were able to come up with a prompt that is flexible and applicable across different scenarios.
The user input (journal entries) is sent to our backend, which embeds the user input into a prompt and sends it to the OpenAI API. OpenAI generates an output that we display in the frontend.
PROMPT CHAINING EXAMPLE IN USE
Prompt chaining allows to orchestrate a series of steps to process user input, generate prompts (using PromptTemplate), and obtain meaningful outputs from an LLM. The LLM chain acts as a pipeline (as shown above), where the output of one step serves as the input to the next step.
Launching the MVP
When User Testing might
not be enough
We finally launch the MVP, that aims to assess user engagement changes from an AB test and evaluate user retention in regards to that feature post-launch.
We encountered a challenge: to quickly get qualitative and quantitative insights from our users on feature that integrates AI technology for users to interact with, as we are in a completely new space. Millions of users come to the app every day seeking relief for their private mental health issues, and we need to quick assess the feature value, consistency and relevance for many different use cases and scenario.
We decide to integrate a feedback screen to help us get a clue about what users think and discover opportunities to improve the features.
Feedback screen Findings
The users' overall opinion of the feature is positive, though it requires some further development in order to meet their expectations. The main issue we have noticed is that advice can often be either too vague or not specific enough.
“The advice did not know the details of my health to make the recommendation, this is why asking questions to go deeper may be necessary for generating a safe and healthy response.”
“It was a little dismissive of the relationship and focused on the finances.”
💡 It is our assumption that the AI is not always generating on-spot insights due to the user's input being too ambiguous. Consequently, it is crucial that users be encouraged to more thoroughly detail their situation and concentrate on one specific topic at the time when journaling. This will help the AI to generate more meaningful insights.
The UX Solution
We enhanced the prompt to become more actionable and compassionate, and designed three different approaches to help users with the journaling, to get the best out of the AI Advice: a Chatbot, an AI-generated follow-up flow, and a Journal Assistant.
Option 1: Chatbot
Option 2: AI-generated follow-up flow
Option 3: AI Assistant
Following internal testing, conducted with the full team's participation, we found the Journal Assistant solution to be easier to implement and deliver a better UX experience. The "Journal Assistant" kicks off the journaling with a generic question, and while the user is in the process of typing, it will scan the content and generate follow-up questions to help the user explore their thoughts further, dig deeper into one single topic at the time and get a relevant AI Advice.
AI feature design
What’s next?
Launching an in-app feature is just the beginning. To ensure it meets user needs, it's essential to collect user feedback and iterate on the product. Valuable insights and suggestions can be gained from users that weren't identified in the development process, while incorporating feedback helps enhance user experience and improve user retention. Incorporating these changes creates opportunities for both user-focused and product-focused improvements that can maximize satisfaction of all parties involved.
