From autocomplete to a search rethink

Inside Figma, the search for existing work had become a genuine time sink. Our own designers were spending too long hunting for source files when all they had was a screenshot—often resulting in Slack threads asking colleagues for help. That problem prompted the launch of AI-powered search at Config 2024, adding two new ways to find work: visual search, which lets you use a screenshot, selected frame, or a rough sketch as your query, and semantic search, which uses AI to interpret context behind text queries even when you don't know exact component names.

These features look obvious in retrospect, but getting them to a state where they truly add value involved months of scoping and testing. The journey started with a three-day AI hackathon in June 2023 where we landed on a working prototype for design autocomplete—an AI assistant that suggests components like a “Get started” button while you work. The momentum was strong enough that we put it on the product roadmap, but as we iterated with internal designer feedback, consistent usage patterns started pointing us in a different direction.

Designers don't start projects from scratch; they riff on existing work, revisit past explorations, and pull in old ideas—but surfacing the right work required knowing organizational structure and file ownership. In fact, 75% of all objects added to the Figma canvas come from other files. This led to a key question: should we ship autocomplete first, or focus on the search capabilities that support it?

We saw two distinct workflows among designers: exploration mode (brainstorming solutions to user problems) and execution mode (getting something on the canvas fast). Since designers oscillate between them, we had to address two core issues:

  1. Design elements aren't discoverable at the right altitude. Existing search forced designers to choose between low-level design system components or entire files, making mid-level frames—combining several components—hard to find without disrupting flow.
  2. Thumbnails don't tell you if a result is relevant. Designers had to open files to check for the latest version or the right design rather than quickly scanning results.

While autocomplete could wait, these search gaps couldn't—so the roadmap shifted to improving search first, which would later power features like autocomplete through Retrieval Augmented Generation (RAG) by finding designs similar to what a designer is already working on.

Scoping and indexing an infinite canvas

Research pointed to three core use cases: frame lookup (finding a specific design to edit), frame variations (seeing different design approaches for a particular UI piece), and broad inspiration (thematically similar concepts to jump from). We started with the most common creation-driven searches like finding exact designs before tackling less frequent ideation use cases, letting us build on existing behavior and learn faster.

Indexing everything was never an option due to cost. Knowing designers work from past designs, we focused on top-level frames that look like UI designs, using heuristics like common UI frame dimensions and considering non-top-level frames when they met certain conditions. Identifying UI designs is especially tricky because designers often riff by duplicating and tweaking, generating pages full of similar designs. We made only one instance of similar designs searchable instead of indexing duplicates, and skipped unaltered copies of entire files as well. Another containment strategy: we held off indexing designs until a file hadn't been edited for four hours, keeping unfinished work and archived designs out of results while reducing load on our systems.

We're still experimenting—one direction is refining "quality signals," like whether a frame is marked ready for development, to rank results.

Measuring quality: evals in two stages

Getting the AI model right required constant performance measurement on tests known as evals. We picked eval queries by talking to internal designers and analyzing file browser search patterns—from simple "checkout screen" queries to descriptive ones like "red website with green squiggly lines" to project-specific searches like "[codename] theme picker."

Having relevant results across similarity levels was crucial. Research showed users prefer starting with a very close or similar match even when ultimately hoping to end up with diverse results. If we couldn't prove the feature could find a needled-in-a-haystack result, designers wouldn't trust it for broader exploration. The capability needed to deliver from exact matches all the way down to somewhat similar options. To grade results at scale, we built a tool via our public plugin API for labeling search results on an infinite canvas, complete with keyboard shortcuts for speed.

Designing for how designers actually work

Rather than assuming a neat linear flow from exploration to execution, we built around the fact that designers go back and forth between modes all the time. Instead of predicting current mode, we offer a range of results and let users choose what fits their need. That meant building one unified interface for refining results regardless of input type, adding filters like “created by” plus metadata showing which file a result came from, who made it, and how recently it was edited.

The user interface itself went through many iterations. One idea called “rabbit holing” let designers click into a result type to explore deeper—but we scrapped it to keep the Actions search experience uncomplicated. When we realized Actions would become the central home for AI features across Figma, it felt increasingly natural for improved search to live there, though limited real estate created interesting puzzles. The design team added details like “peek” previews where results are small enough to scan in the Actions panel, with CMD + Enter letting designers go full-screen for a closer look.

From beta to launch

Months of iterative work, a closed beta, and continuous refinement with an internal eval plugin preceded the wider rollout of AI search in Figma. Four principles guided this effort:

  1. AI for existing workflows: The goal was to streamline things users already do, such as browsing files and copying frames into their current document.
  2. Rapid iteration: Regular updates to staging, paired with insights from the internal beta, drove feature refinement.
  3. Systematic quality checks: Custom evaluation tools were built specifically to monitor and improve result accuracy.
  4. Cross-disciplinary teamwork: Product, content, engineering, and research teams collaborated closely throughout.

What's next

Looking ahead, the roadmap includes plans to extend visual and semantic search to the Figma Community. Anchoring the work in real user problems and using feedback and data to fine-tune the experience has created a foundation for AI-powered design tools that aim to support, not interrupt, the creative process.

The authors thank the many teams at Figma who contributed to these features, along with Spencer de Mars, Augustus Griffin, and Rohun Golkar for their contributions to this article. Figma is also offering the chance to get started and try the new tools for free.