AI makes it easier for consumers to discover more relevant people, places and things.

Instead of going through menus, choosing options from filters, AI allows us to just describe what we want.

This is unlocking a wave of incredible chat to search features in retailers like JCrew and Rent the Runway.

While AI has little value to add in “lower funnel” search (e.g., “Aesop Reverence Hand Wash”), many expect it to rise in research and discovery or upper and middle funnel search, respectively. Walmart describes this in the forthcoming AI search release.

For example, a parent planning a birthday party for a child that loves unicorns. Instead of multiple searches for unicorn-themed balloons, napkins, streamers, etc., the parent can simply ask the question “Help me plan a unicorn-themed party for my daughter.”

We’re excited by these more expressive interfaces for search – helping us edit our now expansive digital access.

But we see the state of AI much as the internet was in 1995. We’re struck by a 2001 NYT telling of it

At that moment in Web history, every visit to a site was like the first, with no automatic way to record that a visitor had dropped by before. Any commercial transaction would have to be handled from start to finish in one visit, and visitors would have to work their way through the same clicks again and again; it was like visiting a store where the shopkeeper had amnesia.

Despite its more natural interface, our AI-powered interactions on sites have amnesia. AI services don’t remember us. They remember what we do in our limited interactions on one surface, but require that we work our way through the same context over and over just so we can be understood.

The solution called for each Web site's computer to place a small file on each visitor's machine that would track what the visitor's computer did at that site. [Netscape engineer] LouMontulli called his new technology a ''persistent client state object.”

We have a similar approach. We allow users to link their personal AI user agent – Crosshatch! – to their favorite apps in a few taps. Rather than work our way through the same context to each site we go to, we just tap “Continue with Crosshatch” so that our favorite sites can reference this context all automatically.

AI search has many technical implementations, including

• Classical (e.g., Elastic) with AI-generated metadata
• Generating related searches and using existing search infra
• Using retrieval augmented generation with LLM reranking

Re-ranking with large language models like GPT-4o-mini or Gemini 1.5 Flash perform well and at unit economics that can beat purpose-built solutions.

‍

Add personalized re-ranking to your app

In this blog, we implement LLM-powered re-ranking, but allow users to bring their own context, in this case, itemized summaries of the user’s most recent past 25 purchases so that brands can deliver a search that’s personalized to the user, without the user having to repeat themselves (like we did in 1995).

‍

This works just like typical re-ranking – here with the LLM gpt-4o-mini, but in the Crosshatch case we enable brands and users to privately activate linked context in order to get results that adapt to all user context, not just that provided in the search bar.

The output of this call is the re-ranked list, but it operates on the user’s past purchases. This activation is privacy preserving in that the only output was the re-ranked list, even though the AI call had access to the user’s permissioned purchase history.

This type of data activation is convenient, transparent and privacy preserving, in that it requires no infrastructure to set up, proxies popular commercial AI models, and happens privately in user-governed AI infrastructure.

So while we’re excited about chat to shop interfaces, we think their final form is more adaptive, adapting to the full context that users link, not just what they might share in a search box.

To get started

• Create a free sandbox account on Crosshatch
• Check out our docs
• Reach out for the full re-ranking source code (in our case we performed re-ranking on eBay's catalog)