Finetune Your Search with a Single API Call: Objective Finetuning Launches into Beta

We’re kicking off the Summer of Shipping here at Objective HQ with a big one: Objective Finetuning is now in self-service Beta! You can now bring your own feedback data to finetune Objective Search Indexes to your dataset, your business goals, and the way your users search. We’ve been heads-down working on this one and we’re really proud of how it has turned out. But what we’re most proud of are the results that customers are already seeing with their finetuning pipelines on Objective, with some companies seeing a 25% improvement in relevance with finetuned Indexes. With just a single API call.

If you have an account, it’s available on your Text Indexes today, with support for Image & Multimodal Indexes coming soon. Check out the quickstart guide to get started. If you don’t have an account yet, get one today!

Going Beyond Frozen Embeddings.

The legacy world of keyword search doesn’t learn — it relies entirely on you to manually tweak rules and parameters. If you’re running an online clothing store, you would need to manually set up synonyms for ‘dark pants’ with every pant with a dark color in your inventory. And then do it all over again next time you add a new seasonal apparel line. Frozen embedding models, used for semantic search, move in the right direction, using more generalized understanding in retrieval. Left to it’s own devices, searches for “bats” may return animal-themed results rather than the desired sporting equipment. Your search needs to be tuned to your dataset and your business goals, which you can’t do with someone else’s off-the-shelf models. As a result, we’ve talked to a lot of companies that felt that customizing AI-native search was too heavy a lift. And we’ve talked to others who believed they needed to home-grow a solution themselves in order to customize vector search. Objective Finetuning supercharges the benefits you get from vector search by orienting it all around your data, your training, and your users.

Historically, this sort of search power has been reserved for a handful of the largest technology companies in the world, who can hire buildings full of experts to refine models, track user interactions, and operate and maintain the systems that support petabytes of data & multi-million-dollar compute budgets.

With Objective, it’s now as easy as supplying training data to .finetune() on the Text Index of your choice, and in around 30 minutes you’ll have a new, finetuned search engine optimized for your needs & your users.

Making Finetuning Even Finer.

We do a lot under the hood to tune & optimize your Index to your dataset & your user’s needs. And that translates into real world impact on the relevancy of your search. The results speak for themselves — the following chart details how a finetuned Objective Index stacks up against a comparable Elastic Search installation for Liloshop, an Objective customer.

Nadine El Ashkar, CEO at Liloshop said “We saw a huge lift when switching from Elastic to Objective, and were already impressed to see such a jump with out-of-the-box Objective Indexes. Seeing another huge lift in relevance when we released our first finetuned Index was unbelievable. We’re excited to see what happens when we keep finetuning.”

Teach a Model to Fish.

We’re on a mission to make search human. A big part of that is making this kind of powerful toolchain available to every developer and business. Jungmin Hwang, a Product Manager at Pod Foods, had this to say about Objective Finetuning —

“Our users search brand names a lot — when someone searches a brand like ‘Better Sour’, we also want to be able to surface other related sour candies. Before using Finetuning, only 20% of search results for the ‘Better Sour’ query were relevant, and following the direct brand matches the results would drift into other ‘sour’ items like strawberry sparkling water and kombucha. After providing just a few ‘GREAT’ and ‘BAD’ match grades to Finetuning our new model returns 100% relevant results, pulling in sour candies from other brands. This new process of teaching rather than programming the matching algorithm saved me countless development hours.”

Ship a Finetuned Search Index Today, Not ‘Q4ish’.

Kicking off Finetuning on your Text Index is a straightforward API call — to get started building, check out the quickstart! You’ll receive status notifications in your Console while the Finetuning is processing, and an email notification when your new index is ready. All in all, a process that would likely take an expert team days to complete is condensed down into a brisk 30-ish minutes.

Take a coffee break while we do the heavy lifting. When your Finetuned Index is ready, you’ll get a friendly email from the platform letting you know it’s ready to go.

Now more than ever, we really can’t wait to see what you build!