Natural language search component with LLMs.

Search is one of the most fundamental interaction models in digital products. From flight booking to hotel reservations to e-commerce, it shapes how users express intent and how systems retrieve relevant results. Over time, the interface for search has become highly structured—sets of form inputs, dropdowns, filters, and step-by-step flows.

But what if the rigid structure of search interfaces could be replaced by a more fluid, natural language-based model? What if, instead of requiring users to fill in isolated fields, we allowed them to express their needs the way they naturally would—through typed sentences—and let the system handle the structure?

This paper explores that idea: how LLMs can reshape search into an adaptive, interactive layer that blends the flexibility of natural language with the clarity of structured inputs.

The Traditional Search Interface

Most current search interfaces are built around structured inputs. A typical flight booking form includes fields for:

• Departure location (From)

• Destination (To)

• Dates (Departure / Return)

• Number of passengers

• Travel class

Additional options—such as preferred airline, price range, or departure time—are often hidden in a secondary layer of filters, accessible only after an initial search is made.

This interface model is predictable and stable, but also rigid. It assumes users follow a specific mental model: fill in the basics first, then refine later. It offers little flexibility for expressing intent in alternative ways, and it fragments the search experience across multiple steps and views.

The Opportunity: Search as Language

LLMs now allow us to treat search not as form-filling, but as conversation.

Consider the following input typed into a single search field:

“I want a flight from LAX to JFK tomorrow morning under $200.”

In this one sentence, the user communicates not only the core parameters (origin, destination, date), but also filters (time of day, price range). This is more expressive than most form-based searches allow upfront.

The challenge has always been parsing this kind of input. But with modern language models, it becomes not only possible—but increasingly robust.

A New Interaction Pattern: Inline Structured Detection

To take full advantage of LLMs, we propose an interaction model where users enter their query in natural language, and the interface dynamically identifies and highlights the structured components of the input.

For example:

“I want a flight from LAX to JFK tomorrow morning under $200.”

As the user types, the system detects entities such as cities, dates, times, and prices. These elements are visually marked—perhaps with subtle pill-shaped UI tokens or highlights—and transformed into interactive components.

Users can then click on “LAX” to change the origin airport, or on “tomorrow” to adjust the date via a calendar. The natural language input becomes a hybrid interface: both readable and manipulable.

This interaction model offers several advantages:

• Fewer steps: Users can express both intent and preferences in a single action.

• Increased clarity: Visual cues confirm that the system correctly understood key details.

• Editable inputs: Users don’t need to delete and retype—they can adjust values inline.

• Scalability: The model extends to other domains like hotels, shopping, and even food delivery.

Design Considerations

A successful implementation would require:

1. Entity Recognition
   Accurate real-time detection of dates, times, currencies, places, and other domain-specific tokens.

2. Visual Consistency
   A lightweight and consistent visual style for detected tokens to avoid disrupting the reading flow.

3. Fallbacks and Flexibility
   If a user enters an ambiguous term (“next week”), the system should offer clarification rather than fail silently.

4. Editable Tokens
   Clickable tokens must be context-aware—e.g., clicking “tomorrow” should bring up a date selector already set to tomorrow.

5. Accessibility
   For screen readers and keyboard users, the system must expose structured data in a navigable format.

Potential Applications

While the example here focuses on flight booking, the pattern can extend to:

• Hotel booking: “Show me 4-star hotels in Paris next weekend under $150/night.”

• E-commerce: “Looking for wireless headphones under $100 with noise cancellation.”

• Event search: “Find concerts in Berlin next Friday near Kreuzberg.”

Anywhere users need to filter and sort through a large set of options, this model applies.

A Transition, Not a Replacement

This approach doesn’t aim to eliminate traditional search interfaces. Instead, it proposes an additional layer—one that reduces friction, especially for experienced users who already know what they want.

For some, structured inputs will remain important. But for others, especially those seeking speed or flexibility, natural language search with inline structuring may offer a faster, more satisfying experience.

Conclusion

The structure of search interfaces hasn’t changed much in two decades. But with the rise of LLMs, there’s an opportunity to rethink how we capture intent—moving from forms and filters to fluid expressions that still retain structure.

By blending natural language with interactive elements, we can build search interfaces that are not only smarter, but also more human.