Completing missing search context through conversation
B2C SaaS AI Start-up
Shipped 2025
Overview
Search Redesign
: Making search less dependent on users knowing what to type
Our first attempt used case-specific examples to show users what a stronger query could look like. But the results were inconsistent. Some users picked up the pattern and wrote detailed descriptions, while others continued with short queries.
I led the next redesign to move more of that work into the product itself, defining what context could improve a search, then designing a conversational flow that asked for it when needed. I took the work from concept to production-ready handoff within 3 weeks.
Team
Product Designer (me!)
AI Engineer (Product Owner)
Back-end developer
Front-end developer
Skills
Product Strategy
AI Interaction Design
Conversation Design
Behavioral Data Analysis
Prototyping
Design Systems
Tools
Figma, Jira, Miro,
Internal User Analytics
Duration
3 weeks
To see the initial approach...
If you’d like to learn more about the initial approach, visit the full story in Guiding better search results with examples ↗ !
Outcome
Filling in missing context led to better search outcomes!
37.5%
66.7%
User-rated satisfaction with search results
solution Preview
Describe your situation naturally
Start with your situation in your own words. If useful context is missing, the search seamlessly shifts into a guided conversation instead of finding unreliable results right away.
Fill in key details through follow-up questions
Answer targeted questions one at a time as the AI picks up on what you’ve already shared and asks for details that could make the search more precise.
Review your updated context & search
See your original situation combined with the details gathered through the conversation, then review the full context and decide when you’re ready to search.
Previous approach
We showed users what a good query looked like.
Case-specific examples gave users something concrete to reference before searching.
USER BEHAVIOR
Some followed the examples, others still typed one line.
When we analyzed de-identified search inputs, the gap in query length and detail was hard to miss. The guidance helped, but only when users chose to do the work.
I bought a used phone on Karrot to use as a second phone, but when I put my SIM in it the phone got suspended. I called customer service and they said the seller still has an eSIM registered on the phone, so because the eSIM and my SIM are under different names they can’t unlock it. They told me to ask the seller to cancel the eSIM but the seller blocked me on Karrot yesterday. I have no way to contact them now.
I’m the victim in a case involving sexual videos being leaked, threats, personal information being posted, false rumors, and stalking. The video was posted on Telegram and Instagram, they demanded 500,000 won and said they would spread it more if I didn’t pay, and they posted my name, address, phone number and workplace online. They also hacked my Instagram, called and messaged me over 100 times late at night, and sent sexual messages to my mom. What kind of sentence could they get and how much would a settlement affect it?
My ex and I are still basically in a common-law relationship, but now my ex is living with a married coworker. They blocked me from contacting my child and aren’t following the visitation agreement. I texted my ex, went to the house twice, and went to their workplace convenience store four times, and now I’ve been reported for stalking. What should I do?
About 5 years ago I borrowed 400 million won from someone I know to open a pub. Then COVID happened, the business kept losing money and I had to shut it down. I haven’t had enough income since then so I still haven’t been able to pay the money back. Could this be considered fraud even though I did actually use the money for the business?
I was indicted for distributing illegally filmed videos for profit three times. There were two unidentified victims and one identified victim, and I already settled with the identified victim and got a letter saying they don’t want punishment. I did earn points from the site but never cashed them out. I’m paying child support, was being treated for depression, and this happened while I was going through a divorce. Afterward I also contacted the website and government agencies to get the videos deleted and blocked. What kind of sentence am I looking at?
It’s a criminal case related to purchasing sex.
I’m in my 20s and have a question about calculating my severance pay.
Can an employee be legally fired for making a mistake that caused the company a major financial loss?
I won a bid, but now I’m being pressured into signing a subcontract. What should I do?
My landlord is refusing to return my security deposit.
I got into a fight at a bar and the other person filed an assault complaint against me.
My company suddenly changed my position and cut my salary without my consent. Is that legal?
Someone keeps contacting me after I told them to stop. Could this be considered stalking?
Detailed queries
One-line queries
The deeper gap
... users still had to figure out what mattered on their own.
A few user interviews helped explain the pattern. The examples showed what a detailed query could look like, but users still had to decide for themselves which details actually mattered.
“I could see the example, but I wasn’t sure which parts of my situation actually mattered.”
“I didn’t know what else I was supposed to add.”
Approach
What if the product could figure out what was missing first?
Instead of expecting users to anticipate every useful detail upfront, we envisioned the product could look at what they had already written, identify potentially useful missing context, and ask for it directly.
Context Detection
Guided Questions
Design Decision
1. Context Detection: How does the system know what’s missing?
I first mapped the context that could narrow almost any search.
I reviewed de-identified queries from the previous flow and mapped the details that consistently made a search more specific. And then grouped those recurring signals into a reusable context model the system could check against each new query.
A. Situation & Parties
Time
Location
Sequence of events
People & roles
B. Outcome & Impact
Injury
Financial loss
Other consequences
C. Supporting Information
Evidence
Records
Witnesses
Then I adapted the context to what each case type actually needed.
I compared the shared signals across different legal situations to see where the baseline stopped being enough. And then I surfaced details that became especially useful within specific case types, such as injury severity for assault or division of assets for divorce.
Assault
Injury severity
Available evidence
Divorce
Grounds for divorce
Division of assets
Debt Recovery
Loan terms or agreements
Default actions
Domestic Violence
Nature of abuse
Frequency and duration
Then I structured those signals into a 3 + 2 context-checking framework
To make context detection responsive to each user’s query, I structured the model around 3 shared context areas + 2 case-specific ones. Starting from the initial query, the system checked what was already present, and surfaced a follow-up for a missing or key detail, then re-evaluated the updated context after each response.
Situation &
Parties
Outcome &
Impact
Supporting
Information
Case-specific
Context 1
Case-specific
Context 2
3 shared areas
2 case-specific contexts
Scaling the case-specific layer across 100 legal situations
To account for the range of situations users actually brought, I used AI to expand the case-specific layer across 100 legal types, each mapped to 2 context areas that could guide more relevant follow-ups.
2. Guided Questions: Now, how should the system ask for it?
I mapped how the conversation should behave from first follow-up to search-ready state.
I structured the conversation around three response stages. Opening the interaction, building context turn by turn, and transitioning to search once enough information had been gathered. At each turn, I also defined how the system should handle off-topic, vague, unanswerable, or contradictory responses before deciding whether more context was needed.
3. Interaction Design: How should this conversation live in the product?
Keep the conversation optional
The conversation was designed to improve search, not to become a required step. When the initial input already provided enough context, the system skipped the guided flow and searched immediately. Also when users simply wanted to move faster, they could bypass the conversation and search with what they had.
Automatic bypass when context is sufficient
User-initiated bypass
to search as is
Keep the evolving context visible
I made the system’s evolving understanding visible throughout the conversation, allowing users to review changes, revisit earlier versions, and search with the version they preferred.
Context updates after each response
Revisit and search from any version
Prevent accidental loss of progress
I surfaced a warning before actions that would reset the conversation, so users could choose to keep refining or leave without accidentally losing their progress.
Make the loss explicit before resetting progress
Takeaway
Designing this AI experience pushed me to think beyond flows and edge cases, toward the system behind them. User inputs were far more varied than I expected, and supporting them well meant designing a structure that could scale in branches, behaviors, and resources, as new patterns emerged.
A few things stood out...
1. Designing for variability means designing for scale
A flexible AI experience needs more than a solid core flow. The system itself has to be built to keep scaling as new input patterns and behaviors emerge.
2. Giving users control within a system-led flow
A system can guide the process, but users should never feel trapped by it. Making the process visible, and giving users clear ways to accept, skip, or override it, became essential to reducing friction.