Making Each Search Rewarding: How Ibotta Reworked Provide Discovery With Databricks

At Ibotta, our mission is to Make Each Buy Rewarding. Serving to our customers (whom we name Savers) discover and activate related presents by means of our direct-to-consumer (D2C) app, browser extension, and web site is a vital a part of this mission. Our D2C platform helps tens of millions of customers earn cashback from their on a regular basis purchases—whether or not they’re unlocking grocery offers, incomes bonus rewards, or planning their subsequent journey. By the Ibotta Efficiency Community (IPN), we additionally energy white-label cashback packages for a number of the greatest names in retail, together with Walmart and Greenback Common, serving to over 2,600 manufacturers attain greater than 200 million customers with digital presents throughout associate ecosystems.

Behind the scenes, our Information and Machine Studying groups energy vital experiences like fraud detection, provide advice engines, and search relevance to make the Saver journey personalised and safe. As we proceed to scale, we’d like data-driven, clever programs that help each interplay at each touchpoint.

Throughout D2C and the IPN, search performs a pivotal function in engagement and must hold tempo with our enterprise scale, evolving provide content material, and altering Saver expectations.

On this publish we’ll stroll by means of how we considerably refined our D2C search expertise: from an formidable hackathon undertaking to a sturdy manufacturing characteristic now benefiting tens of millions of Savers.

We believed our search might higher sustain with our Savers

Person search conduct has advanced from easy key phrases to incorporating pure language, misspellings, and conversational phrases. Trendy search programs should bridge the hole between what customers kind and what they really imply, deciphering context and relationships to ship related outcomes even when question phrases don’t precisely match the content material.

At Ibotta, our authentic homegrown search system, at occasions, struggled to maintain tempo with the evolving expectations of our Savers and we acknowledged a chance to refine it.

The important thing areas for alternative we noticed included:

• Bettering semantic relevance: Specializing in understanding Saver intent over actual key phrase matches to attach them with the fitting presents.
• Enhancing understanding: Deciphering the total nuance and context of person queries to offer extra complete and actually related outcomes.
• Rising flexibility: Extra quickly integrating new provide sorts and adapting to altering Saver search patterns to maintain our discovery expertise rewarding.
• Boosting discoverability: We wished extra sturdy instruments to make sure particular sorts of presents or key promotions had been persistently seen throughout a big selection of related search queries.
• Accelerating iteration and optimization: Enabling sooner, impactful enhancements to the search expertise by means of real-time changes and efficiency tuning.

We believed the system might higher hold tempo with altering provide content material, search behaviors, and evolving Saver expectations. We noticed alternatives to extend the worth for each our Savers and our model companions.

From hackathon to manufacturing: reimagining search with Databricks

Addressing the restrictions of our legacy search system required a centered effort. This initiative gained vital momentum throughout an inside hackathon the place a cross-functional staff, together with members from Information, Engineering, Advertising and marketing Analytics, and Machine Studying, got here along with the concept to construct a contemporary, various search system utilizing Databricks Vector Search, which some members had realized about on the Databricks Information + AI Summit.

In simply three days, our staff developed a working proof-of-concept that delivered semantically related search outcomes. Right here’s how we did it:

1. Collected provide content material from a number of sources in our Databricks catalog
2. Created a Vector Search endpoint and index with the Python SDK
3. Used pay-per-token embedding endpoints with 4 totally different fashions (BGE giant, GTE giant, GTE small, a multilingual open-source mannequin, and a Spanish-language-specific mannequin)
4. Related every thing to our web site for a reside demo

The hackathon undertaking gained first place, generated sturdy inside buy-in and momentum to transition the prototype right into a manufacturing system. Over the course of some months, and with shut collaboration from the Databricks staff, we remodeled our prototype into a sturdy full-fledged manufacturing search system.

From proof of idea to manufacturing

Transferring the hackathon proof-of-concept to a production-ready system required cautious iteration and testing. This section was vital not just for technical integration and efficiency tuning, but in addition for evaluating whether or not our anticipated system enhancements would translate into constructive adjustments in Saver conduct and engagement. Given search’s important function and deep integration throughout inside programs, we opted for the next strategy: we modified a key inside service that known as our authentic search system, changing these calls with requests directed to the Databricks Vector Search endpoint, whereas constructing in sturdy, swish fallbacks to the legacy system.

Most of our early work centered on understanding:

Within the first month, we ran a check with a small proportion of our Savers which didn’t obtain the engagement outcomes we had hoped for. Engagement decreased, notably amongst our most lively Savers, indicated by a drop in clicks, unlocks (when Savers categorical curiosity in a proposal), and activations.

Nonetheless, the Vector Search resolution supplied vital advantages together with:

• Sooner response occasions
• An easier psychological mannequin
• Better flexibility in how we listed knowledge
• New talents to regulate thresholds and alter embedding textual content

Happy with the system’s underlying technical efficiency, we noticed its larger flexibility as the important thing benefit wanted to iteratively enhance search end result high quality and overcome the disappointing engagement outcomes.

Constructing a semantic analysis framework

Following our preliminary check outcomes, relying solely on A/B testing for search iterations was clearly inefficient and impractical. The variety of variables influencing search high quality was immense—together with embedding fashions, textual content combos, hybrid search settings, Approximate Nearest Neighbors (ANN) thresholds, reranking choices, and plenty of extra.

To navigate this complexity and speed up our progress, we determined to ascertain a sturdy analysis framework. This framework wanted to be uniquely tailor-made to our particular enterprise wants and able to predicting real-world person engagement from offline efficiency metrics.

Our framework was designed round an artificial analysis setting that tracked over 50 on-line and offline metrics. Offline, we monitored normal info retrieval metrics like Imply Reciprocal Rank (MRR) and precision@okay to measure relevance. Crucially, this was paired with on-line real-world engagement indicators corresponding to provide unlocks and click-through charges. A key determination was implementing an LLM-as-a-judge. This allowed us to label knowledge and assign high quality scores to each on-line query-result pairs and offline outputs. This strategy proved to be vital for fast iteration primarily based on dependable metrics and amassing the labeled knowledge vital for future mannequin fine-tuning.

Alongside the way in which, we leaned into a number of components of the Databricks Information Intelligence Platform, together with:

• Mosaic AI Vector Search: Used to energy high-precision, semantically wealthy search outcomes for analysis checks.
• MLflow patterns and LLM-as-a-judge: Offered the patterns to judge mannequin outputs and implement our knowledge labeling course of.
• Mannequin Serving Endpoints: Environment friendly deployment of fashions instantly from our catalog.
• AI Gateway: To safe and govern our entry to 3rd social gathering fashions by way of API.
• Unity Catalog: Ensured the group, administration, and governance of all datasets used throughout the analysis framework.

This sturdy framework dramatically elevated our iteration pace and confidence. We performed over 30 distinct iterations, systematically testing main variable adjustments in our Vector Search resolution, together with:

• Completely different embedding fashions (foundational, open-weights, and third social gathering by way of API)
• Numerous textual content combos to feed into the fashions
• Completely different question modes (ANN vs Hybrid)
• Testing totally different columns for hybrid textual content search
• Adjusting thresholds for vector similarity
• Experimenting with separate indexes for various provide sorts

The analysis framework remodeled our growth course of, permitting us to make data-driven choices quickly and validate potential enhancements with excessive confidence earlier than exposing them to customers.

The seek for the very best off-the-shelf mannequin

Following the preliminary broad check that confirmed disappointing engagement outcomes, we shifted our focus to exploring the efficiency of particular fashions recognized as promising throughout our offline analysis. We chosen two third-party embedding fashions for manufacturing testing, accessed securely by means of AI Gateway. We performed short-term, iterative checks in manufacturing (lasting a couple of days) with these fashions.

Happy with the preliminary outcomes, we proceeded to run an extended, extra complete manufacturing check evaluating our main third-party mannequin and its optimized configuration in opposition to the legacy system. This check yielded blended outcomes. Whereas we noticed general enhancements in engagement metrics and efficiently eradicated the damaging impacts seen beforehand, these good points had been modest—largely single-digit proportion will increase. These incremental advantages weren’t compelling sufficient to totally justify a whole substitute of our present search expertise.

Extra troubling, nonetheless, was the perception gained from our granular evaluation: whereas efficiency considerably improved for sure search queries, others noticed worse outcomes in comparison with our legacy resolution. This inconsistency offered a major architectural dilemma. We confronted the unappealing alternative of implementing a fancy traffic-splitting system to route queries primarily based on predicted efficiency—an strategy that will require sustaining two distinct search experiences and introduce a brand new, complicated layer of rule-based routing administration—or accepting the restrictions.

This was a vital juncture. Whereas we had seen sufficient promise to maintain going, we would have liked extra vital enhancements to justify absolutely changing our homegrown search system. This led us to start fine-tuning.

High-quality-tuning: customizing mannequin conduct

Whereas the third-party embedding fashions explored beforehand confirmed technical promise and modest enhancements in engagement, additionally they offered vital limitations that had been unacceptable for a long-term resolution at Ibotta. These included:

1. Lack of ability to coach embedding fashions on our proprietary provide catalog
2. Issue evolving fashions alongside enterprise and content material adjustments
3. Uncertainty relating to long-term API availability from exterior suppliers
4. The necessity to set up and handle new exterior enterprise relationships
5. Community calls to those suppliers weren’t as performant as self-hosted fashions

The clear path ahead was to fine-tune a mannequin particularly tailor-made to Ibotta’s knowledge and the wants of our Savers. This was made attainable because of the tens of millions of labeled search interactions we had gathered from actual customers by way of our LLM-as-a-judge course of inside our customized analysis framework. This high-quality manufacturing knowledge grew to become our coaching gold.

We then launched into a methodical fine-tuning course of, leveraging our offline analysis framework extensively.

Key parts had been:

• Infrastructure: We used AI Runtime with A10s in a serverless setting, and Databricks ML Runtime for classy hyperparameter sweeping.
• Mannequin choice: We chosen a BGE household mannequin over GTE, which demonstrated stronger efficiency in our offline evaluations and proved extra environment friendly to coach.
• Dataset engineering: We constructed quite a few coaching datasets, together with producing artificial coaching knowledge, finally deciding on:
  • One constructive end result (a verified good match from actual searches)
  • ~10 damaging examples per constructive, combining:
    • 3-4 “onerous negatives” (LLM labeled, human-verified inappropriate matches)
    • “In-batch negatives” (sampling of outcomes from unrelated search phrases)
• Hyperparameter optimization: We systematically swept issues like studying fee, batch measurement, length, and damaging sampling methods to seek out optimum configurations.

After quite a few iterations and evaluations throughout the framework, our top-performing fine-tuned mannequin beat our greatest third-party baseline by 20% in artificial analysis. These compelling offline outcomes offered the boldness wanted to speed up our subsequent manufacturing check.

Search that drives outcomes—and income

The technical rigor and iterative course of paid off. We engineered a search resolution particularly optimized for Ibotta’s distinctive provide catalog and person conduct patterns, delivering outcomes that exceeded our expectations and supplied the pliability wanted to evolve alongside our enterprise. Based mostly on these sturdy outcomes, we accelerated migration onto Databricks Vector Search as the inspiration for our manufacturing search system.

In our last manufacturing check, utilizing our personal fine-tuned embedding mannequin, we noticed the next enhancements:

• 14.8% extra provide unlocks in search.
  This measures customers deciding on presents from search outcomes, indicating improved end result high quality and relevance. Extra unlocks are a number one indicator of downstream redemptions and income.
• 6% improve in engaged customers.
  This exhibits a larger share of customers discovering worth and taking significant motion throughout the search expertise, contributing to improved conversion, retention and lifelong worth.
• 15% improve in engagement on bonuses.
  This displays improved surfacing of high-value, brand-sponsored content material, translating instantly to higher efficiency and ROI for our model and retail companions.
• 72.6% lower in searches with zero outcomes.
  The numerous discount means fewer irritating experiences and a significant enchancment in semantic search protection.
• 60.9% fewer customers encountering searches returning no outcomes.
  This highlights the breadth of affect, displaying that a big portion of our person base is now persistently discovering outcomes, enhancing the expertise throughout the board.

Past user-facing good points, the brand new system delivered on efficiency. We noticed 60% decrease latency to our search system, attributable to Vector Search question efficiency and the fine-tuned mannequin’s decrease overhead.

Leveraging the pliability of this new basis, we additionally constructed highly effective enhancements like Question Transformation (enriching imprecise queries) and Multi-Search (fanning out generic phrases). The mixture of a extremely related core mannequin, improved system efficiency, and clever question enhancements has resulted in a search expertise that’s smarter, sooner, and finally extra rewarding

Question Transformation

One problem with embedding fashions is their restricted understanding of area of interest key phrases, corresponding to rising manufacturers. To handle this we constructed a question transformation layer that dynamically enriches search phrases in-flight primarily based on predefined guidelines.

For instance, if a person searches for an rising yogurt model the embedding mannequin may not acknowledge, we will remodel the question so as to add “Greek yogurt” alongside the model title earlier than sending it to Vector Search. This offers the embedding mannequin with vital product context whereas preserving the unique textual content for hybrid search.

This functionality additionally works hand-in-hand with our fine-tuning course of. Profitable transformations can be utilized to generate coaching knowledge; as an illustration, together with the unique model title as a question and the related yogurt merchandise as constructive ends in a future coaching run helps the mannequin study these particular associations.

Multi-Search

For broad, generic searches like “child,” Vector Search would possibly initially return a restricted variety of candidates, doubtlessly filtered down additional by focusing on and finances administration. To handle this and improve end result variety, we constructed a multi-search functionality that followers out a single search time period into a number of associated searches.

As an alternative of simply trying to find “child,” our system robotically runs parallel searches for phrases like “child meals,” “child clothes,” “child medication,” “child diapers,” and so forth. Due to the low latency of Vector Search, we will execute a number of searches in parallel with out growing the general response time to the person. This offers a wider and extra numerous set of related outcomes for wide-ranging class searches.

Classes Realized

Following the profitable last manufacturing check and the total rollout of Databricks Vector Search to our person base – delivering constructive engagement outcomes, elevated flexibility, and highly effective search instruments like Question Transformation and Multi-Search – this undertaking journey yielded a number of helpful classes:

1. Begin with a proof of idea: The preliminary hackathon strategy allowed us to rapidly validate the core idea with minimal upfront funding.
2. Measure what issues to you: Our tailor-made 50-metric analysis framework was essential; it gave us confidence that enhancements noticed offline would translate into enterprise affect, enabling us to keep away from repeated reside testing till options had been actually promising.
3. Do not soar straight to fine-tuning: We realized the worth of totally evaluating off-the-shelf fashions and exhausting these choices earlier than investing within the larger effort required for fine-tuning.
4. Accumulate knowledge early: Beginning to label knowledge from our second experiment ensured a wealthy, proprietary dataset was prepared when fine-tuning grew to become vital.
5. Collaboration accelerates progress: Shut partnership with Databricks engineers and researchers, sharing insights on Vector Search, embedding fashions, LLM-as-a-judge patterns, and fine-tuning approaches, considerably accelerated our progress.
6. Acknowledge cumulative affect: Every particular person optimization, even seemingly minor, contributed considerably to the general transformation of our search expertise.

What’s subsequent

With our fine-tuned embedding mannequin now reside throughout all direct-to-consumer (D2C) channels, we subsequent plan to discover scaling this resolution to the Ibotta Efficiency Community (IPN). This is able to deliver improved provide discovery to tens of millions extra customers throughout our writer community. As we proceed to gather labeled knowledge and refine our fashions by means of Databricks, we consider we’re properly positioned to evolve the search expertise alongside the wants of our companions and the expectations of their prospects.

This journey from a hackathon undertaking to a manufacturing system proved that reimagining a core product expertise quickly is achievable with the fitting instruments and help. Databricks was instrumental in serving to us transfer quick, fine-tune successfully, and finally, make each search extra rewarding for our Savers.