SKIP TO CONTENT
← ALL WORK
03 — CONVERSATION SEARCH · SAAS · AI · 2025

One search box that finds any conversation, on any channel.

Search for the Tawk.to omni-channel inbox. Ranking and three filters get an agent to the right thread in seconds, with tags, status and assignee still attached when they land on it.

Role
Product designer — interviews, wireframes, hi-fi UI, usability study
Client
Tawk.to · shipped
Timeline
2025
Platform
Web dashboard
Conversation Search — Tawk.to inbox
1
search entry point for every channel
3
filters that matter: assignee, updated, tags
0
context lost: tags, status and priority travel with the results
Live
shipped into the product used by teams worldwide

Problem

Agents worked four inboxes and opened all of them to find one thread. Plain-text search ranked nothing, and the results arrived stripped of tags, status and assignee, so people re-read conversations they had already handled.

SMS and ticket threads were the worst of it. The conversation was there; finding it took longer than answering it.

Research

Ask an agent about a thread and they answer with a name and a day.

I sat with agents while they searched. Most queries were a version of “the one Sarah handled yesterday”: a person and a rough time, rather than the exact words in the message. I built the filters around that.

4
fragmented inboxes agents toggled between
#1
recall cue: who handled it and when
2×
the same query re-run across channels

Process

  1. DISCOVER
    Agent interviews, inbox shadowing, audit of search across channels
  2. DEFINE
    One entry point; ranking first, filters second; metadata carried over
  3. IDEATE
    Wireframed filter models around assignee, last updated and tags
  4. DESIGN
    Hi-fi UI within the Tawk.to design system; ranked result cards
  5. VALIDATE & SHIP
    Usability study, iteration, shipped to the live product

Decisions

01

Rank first, filter second.

Ranking surfaces the most relevant threads; filters narrow what is already there.

TRADE-OFF

A filter-first model would behave more predictably. Agents cared more about reaching the right thread quickly.

02

Three filters that match how agents remember: assignee, updated, tags.

Three controls built around recall, instead of an advanced-search panel nobody opens twice.

TRADE-OFF

Power users lost a few edge-case filters. The search they run all day got much faster.

03

Results carry their context.

Tags, assignee and status appear on every result, so an agent can pick the thread up without opening it first.

TRADE-OFF

The result cards got busier. Agents could answer straight from them.

PrototypeTap the glowing hotspots to move through the app, like the real prototype.

Conversation Search — Search entry point
Conversation Search — Search with assignee filter
Conversation Search — Ranked results
Conversation Search — Sort by last updated
Conversation Search — Filter by tags
Conversation Search — Tagged result detail
Search entry point
One search for every channel, right where agents already work.
01 / 06

Before → After

BEFORE

A separate inbox per channel, plain-text search with no ranking, and results stripped of tags, status and assignee.

AFTER

One search box across SMS, email, chat and social. Ranked results that keep their context, and three filters shaped by how agents remember.

Outcome

  • Agents find the right conversation in seconds and continue it with the context intact.
  • Ranking does the work; assignee, updated and tags give enough control without an advanced-search panel.
  • Shipped into the live Tawk.to product that support teams use worldwide.

Reflection

Powerful search is easy to build and hard to make feel simple. If I picked it up again I would measure how long a query takes to reach its first open, and try natural-language queries like “Sarah’s chats from yesterday” on top of the ranking.