Tiyaro DeepQuery
An enterprise-focused AI-powered conversational search app
Context
Traditional search methods within enterprises often involve navigating through complex folder structures, using keyword-based searches, or relying on static document metadata. This approach is time-consuming, can lead to information silos, and may not provide accurate or comprehensive results.
The goal of this project is to design a conversational search app that leverages the power of natural language processing and conversation to provide employees with a seamless and efficient way to access, search, and retrieve enterprise documents and data.
Role and
Responsibilities
As a sole designer in a 7 member early-stage startup, I had complete ownership and served as a single point of contact for all the design decisions and changes. I collaborated closely with cross-functional teams, including engineering and business stakeholders, to ensure the design aligned with technical feasibility and business objectives.
As a product designer, I was responsible for conceptualizing, designing, and prototyping the user interface for the conversational search app.
The Scenario
Tiyaro had the technology but not the user interface
Team Tiyaro had build a proprietary software that enabled conversational search capability on enterprise data. To put it simply, this means that you can ask questions about a specific dataset and this software will find answers for you. Equivalent to ChatGPT but for enterprise owned data. This is called conversational enterprise search. The technology was ready but the interface was not. This is where I come in, to design that interface.
Given the completely new interface and absence of existing users, the initial UX methods focused on exploration, understanding user needs, and defining the foundational aspects of the conversational search app.
Understanding Conversational Search
Traditionally, search systems use literal keywords taken directly from query text to navigate their indexes and databases. It would typically be a query followed by a list of results that match with the keywords in the query. As opposed to a traditional keyword search, a conversational search system takes complex grammatical sentences and can use context from previous interactions to provide more useful and comprehensive results.
Why do we even need a more efficient information search for enterprises?
Competitive Analysis
-
Researched existing solutions in the enterprise search domain and analyzed their strengths and weaknesses.
-
How conversational search is applied in various contexts to solve enterprise level challenges.
-
Understand how these solutions handle conversation-based interactions.
-
By analyzing similar conversational apps in the market, I gained insights into emerging trends and identified design opportunities.
Designing for conversational search experience
The 8 second attention span is real
With increasing digital mediums our attention span is decreasing. A Microsoft study revealed that we now have an attention span of eight seconds–making our attention span shorter than that of a goldfish (at nine seconds). Users can get frustrated and quit using the tool if they do not get the information they are looking for quickly or within first few attempts.
Users want relevant information, not more information
When users search for information, they are not looking to gather as much information as possible, they just want to see the most relevant information. Going to the next level of relevancy, users expect information personally relevant to the work they do in the organization.
Search can behave differently depending on who the user is and where they are
Different roles (like manager, IT, software developers) with the same query may have different information needs and an enterprise search system can exploit this information. While designing for enterprise search we can utilize contextual information not only about the search query but also that of the user.
Stakeholder and user interviews
As a first step, I conducted interviews and discussions with stakeholders and potential users to understand their current document and data retrieval pain points and expectations.
The questions I sought to answer were:
-
What is a typical process for employees to browse through enterprise data?
-
What are the business requirements for the first version to be rolled out for this app?
-
Gain insights into the challenges they face
-
The types of documents they work with.
-
Their preferred ways of interacting with data.
These insights are captured in the following persona:
How can we make organizational information finding easier for John with conversational search?
Ideation
Having understood conversational search and user pain points, it was time to explore creative solutions that aligned with user needs and Tiyaro's business objectives. Let's first understand how having an enterprise search app would help Tiyaro users.
User story map
Where do we even start? What will we need in such an app? I first got my ideas out on paper with the help of user story maps. I identified 5 tasks that were crucial for the end-to end app experience and then ideated the details of each task. These detailed steps we divided into two phases.
Phase 1: High Priority, to be taken up first
Phase 2: Low priority, to be taken up later
User Happy Path
Lets first understand what are the various scenarios that a user might encounter.
User Flow
Once I had all the feature requirements, I mapped out the user flow. It started with basic conversation flow which I iteratively improved to include various features. I primarily relied on stakeholder and engineering team input to selectively include these sub-flows.
Basic conversation flow
Improved conversation flow after including feedback, clarification of ambiguous questions, Follow-up questions
High Fidelity Mockups
Landing page when you start the application
Conversational flow - turn-taking, accessing conversation history, presenting references
Followup questions
Reference documents
Presenting Information Source
One of the major reported issues when interacting with conversational AI applications is the credibility of answer. Whether the answer we get is trustworthy?
One way we have tried to address this issue is by present the information source. This creates credibility for the information and the user has flexibility to verify the answer right there.
The reference section evolved from option 1 to option 3 with the help of various stakeholder inputs and design feedback. When the user click on any of the resources, a document preview opens up as shown in the fourth image
Option 1
1. Presents all references in one list.
2. Focuses too much attention on the reference section.
3. The document type is represented by the file extension., not ideal.
Option 2
1. Separate section for primary references and other resources.
2. Use of icons to represent the file type.
Option 3
1. Reduced visual weight of references section.
2. Option to open the document in a modal, giving user the preview of information in the source document.
When the user click on the document to preview, it opens up a modal window as shown.
1. Gives information about section and subsection of the exact information.
2. Option to toggle between different resources used for generating the answer.
Dis-ambiguating user question
Often times, users ask generalist questions that are applicable for different contexts. For example, "How do you change network settings?" is applicable for Product A and Product B but the procedures are different for both. What do you do in that case?
I explored 4 variants to handle this use case. Each has its own positives and negatives. And considering it all, we decided to go ahead with the third one.
1
Directly present a list of disambiguation options and prompt the user to select one.
2
Presents answer for frequently accessed or recently used context followed by a list of clickable alternatives. This increases the likelihood of user not having to select from the options and click on one.
3
Presents answer for frequently accessed or recently used context followed by a list of clickable alternatives.These options have some contextual data to make selection easier for the user.
4
Suggests various contexts right when the user is typing the query. However, this approach is computationally heavy and may lead to issues with latency.
Learnings
Keep stakeholders in loop
In an early stage startup like Tiyaro, things move quite quickly. Constraints change and new requirements come up that influence design decisions. Hence, it's crucial to keep everyone updated about your design changes and stay on top of new developments on business and technical fronts.
Testing with even one participant is better than no testing
Since this product was in early Proof of Concept (PoC) stage, access to actual users was limited. Even then, I made sure that in PoC meetings with customers we actively make them use the application by doing certain tasks. I took this opportunity to observe how they use it, what questions they ask, and where they get stuck. This helped me uncover some critical issues that interrupted user flow and fixing them helped the team present a robust prototype during the next meeting.