AI-Powered Natural Language Flight Search Implementation

alaska_airlines · Tech · 2024 · source Agentic: no · promotional bias: med · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Consumer travel search and flight booking

Alaska Airlines wanted to help travelers who do not already know an exact destination search for trips in natural language, while grounding recommendations in real flight availability, customer context, destination visuals, and prices.

Constraints:

• the experience is customer-facing and must preserve guest trust
• recommendations must be grounded in real available tickets and prices
• the interface should support exploratory trip sparks rather than only known-destination searches
• customer and airline data must be integrated with Gemini's world knowledge
• the product needed to move from prototype to production-quality implementation quickly
• the team rejected a pure chatbot experience in favor of an immersive destination-search interface

Architecture

Alaska Airlines built a natural-language destination search powered by Gemini. A traveler enters a trip spark such as a cabin getaway or cultural destination. Gemini interprets the natural-language intent and combines world knowledge with Alaska Airlines data and customer context. Function calling invokes Alaska's flight API so each recommendation is grounded in ticket availability and price. The UI presents destination recommendations, booking links, generated destination copy, images, and Google Maps aerial/visual integrations. The implementation was accelerated by Alaska's modernized cloud-native microservice platform and existing Google data partnerships.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  destination-search-ui["destination-search-ui"] --> gemini-intent-layer["gemini-intent-layer"]
  gemini-intent-layer --> alaska-customer-context["alaska-customer-context"]
  gemini-intent-layer --> flight-api-function["flight-api-function"]
  flight-api-function --> booking-link["booking-link"]
  booking-link --> destination-search-ui
  destination-search-ui --> google-maps-visuals["google-maps-visuals"]
  gemini-intent-layer --> generated-copy["generated-copy"]
  generated-copy --> destination-search-ui
  cloud-native-platform["cloud-native-platform"] --> destination-search-ui
  cloud-native-platform --> flight-api-function
  cloud-native-platform --> alaska-customer-context

Components

Data flow

A guest enters a natural-language travel spark. Gemini interprets the intent, coordinates world knowledge with Alaska-specific and customer data, then uses function calling to call the flights API. The API confirms ticket availability and returns prices. The destination-search UI displays recommendations, ticket links, map imagery, destination photos, and dynamically generated copy, including multilingual output where Gemini supports the language.

Why this architecture

The system combines Gemini with function calling because travel inspiration alone is not enough: each recommendation needs factual availability and price. Alaska chose an immersive destination-search UI over a chatbot after guest feedback, matching the exploratory nature of trip planning. Google Maps and generated copy make destinations tangible, while the cloud-native microservice platform and existing Google partnership made rapid production-quality integration possible.

Models used: Gemini · Google Maps AI-enhanced imagery

Scale

• Alaska Airlines serves about 45 million customers and guests each year.
  “we serve about 45 million customers and guests every year”transcript, stated
• Alaska flies to 120 plus destinations itself and can reach 1,000 plus destinations through partners.
  “we traveled to uh 120 plus destinations ourselves and then we joined the One World Alliance”transcript, stated

  “in partnership with all the um our uh Global carriers you can go to a thousand plus destinations”transcript, stated
• Alaska described the visible experience as reaching production-level quality in a matter of weeks.
  “we went from inception to that production level quality in matter of weeks”transcript, stated
• The company described an eight-step GenAI process from idea selection to productionization.
  “we have created an eight-step process from all the way from selecting our ideas for which ones to go after all the way to getting them productionize”transcript, stated

Alternatives considered (roads not taken)

Traditional known-destination search only

Rejected because: The team wanted to support travelers at the spark stage who can describe a trip without knowing the destination.

Chatbot-style experience

Rejected because: Guest feedback pushed the team toward an immersive destination-search experience.

Ungrounded destination recommendations

Rejected because: Every recommendation first checks the flights API for available tickets and price.

Patterns

tool-use-orchestration (orchestration)

Gemini recommendations are grounded by function calling into Alaska's flight API for availability and price.

rag-retrieval (retrieval)

The experience grounds Gemini's world knowledge with Alaska-specific data and customer information.

structured-output-contract (integration)

Flight recommendations are coupled to concrete ticket links, availability, and prices rather than free-form destination text only.

eval-driven-development (evaluation)

The team iterated prototypes around guest feedback and changed from chatbot to immersive destination search.

Tradeoffs

Build an immersive search experience.

Chose natural-language destination search with maps and visual context over a chatbot.

Rationale: Guest feedback showed a richer exploration interface was a better fit for travel planning.

Ground recommendations with flight APIs.

Chose Gemini function calling into Alaska APIs over LLM-only destination suggestions.

Rationale: The product needed factual availability, price, and purchasable ticket links.

Use cloud-native modernization as the integration foundation.

Chose microservices and cloud-native platform over monolithic legacy integration.

Rationale: The modernized stack let Alaska integrate new AI capabilities quickly.

Gotchas (what bit them)

Guest trust limits how aggressively to expose customer-facing GenAI.

Root cause: Alaska has a trusted customer relationship and cannot risk a poor or misleading AI booking experience.

Resolution: Start carefully, use factual grounding through function calling, and preserve guest choice.

A chatbot was the wrong product shape for travel inspiration.

Root cause: Guests wanted an immersive exploration experience, not just a conversational box.

Resolution: Build natural-language destination search with maps, imagery, generated copy, and booking links.

Recommendations without availability would be misleading.

Root cause: Travel search must connect inspiration to bookable inventory and price.

Resolution: Use function calling to the flights API before displaying a purchasable recommendation.

Evaluation

Alaska used an eight-step GenAI process, multiple prototypes, and guest feedback to select and reshape the product experience. The team evaluated whether recommendations were grounded by available tickets and prices.

Metrics: guest feedback on prototypes · availability check through flight API · price returned with recommendation · time from inception to production-level quality

Outcomes

• Alaska demonstrated a customer-facing natural-language destination search powered by Gemini
• recommendations include availability, price, and ticket links
• the team moved from inception to production-level quality in weeks
• the experience supports generated multilingual destination copy

Alaska serves about 45 million guests yearly, flies to 120 plus destinations directly, and reaches more than 1,000 destinations through partners.