At Ibotta, our mission is to Make Each Buy Rewarding. Serving to our customers (whom we name Savers) discover and activate related affords by way of our direct-to-consumer (D2C) app, browser extension, and web site is a essential a part of this mission. Our D2C platform helps tens of millions of consumers 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 means of the Ibotta Efficiency Community (IPN), we additionally energy white-label cashback applications for a few of the greatest names in retail, together with Walmart and Greenback Common, serving to over 2,600 manufacturers attain greater than 200 million shoppers with digital affords throughout associate ecosystems.
Behind the scenes, our Knowledge and Machine Studying groups energy essential experiences like fraud detection, provide suggestion engines, and search relevance to make the Saver journey customized and safe. As we proceed to scale, we want data-driven, clever programs that help each interplay at each touchpoint.
Throughout D2C and the IPN, search performs a pivotal position in engagement and must preserve tempo with our enterprise scale, evolving provide content material, and altering Saver expectations.
On this submit we’ll stroll by way of how we considerably refined our D2C search expertise: from an bold hackathon venture to a strong manufacturing function now benefiting tens of millions of Savers.
We believed our search may higher sustain with our Savers
Consumer search habits has advanced from easy key phrases to incorporating pure language, misspellings, and conversational phrases. Fashionable search programs should bridge the hole between what customers sort and what they really imply, decoding 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 instances, 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 correct affords.
- Enhancing understanding: Decoding the total nuance and context of person queries to offer extra complete and really related outcomes.
- Growing flexibility: Extra quickly integrating new provide varieties and adapting to altering Saver search patterns to maintain our discovery expertise rewarding.
- Boosting discoverability: We needed extra sturdy instruments to make sure particular forms of affords or key promotions had been constantly seen throughout a big selection of related search queries.
- Accelerating iteration and optimization: Enabling sooner, impactful enhancements to the search expertise by way of real-time changes and efficiency tuning.
We believed the system may higher preserve 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 constraints of our legacy search system required a centered effort. This initiative gained vital momentum throughout an inner hackathon the place a cross-functional crew, together with members from Knowledge, Engineering, Advertising 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 discovered about on the Databricks Knowledge + AI Summit.
In simply three days, our crew developed a working proof-of-concept that delivered semantically related search outcomes. Right here’s how we did it:
- Collected provide content material from a number of sources in our Databricks catalog
- Created a Vector Search endpoint and index with the Python SDK
- Used pay-per-token embedding endpoints with 4 completely different fashions (BGE giant, GTE giant, GTE small, a multilingual open-source mannequin, and a Spanish-language-specific mannequin)
- Related all the pieces to our web site for a dwell demo
The hackathon venture gained first place, generated sturdy inner 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 crew, we reworked our prototype into a strong 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 part was essential not just for technical integration and efficiency tuning, but in addition for evaluating whether or not our anticipated system enhancements would translate into optimistic modifications in Saver habits and engagement. Given search’s important position and deep integration throughout inner programs, we opted for the next method: we modified a key inner 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 share of our Savers which didn’t obtain the engagement outcomes we had hoped for. Engagement decreased, notably amongst our most energetic Savers, indicated by a drop in clicks, unlocks (when Savers specific curiosity in a suggestion), and activations.
Nevertheless, the Vector Search answer provided vital advantages together with:
- Sooner response instances
- An easier psychological mannequin
- Larger flexibility in how we listed information
- New skills to regulate thresholds and alter embedding textual content
Happy with the system’s underlying technical efficiency, we noticed its higher flexibility as the important thing benefit wanted to iteratively enhance search outcome 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 lots of extra.
To navigate this complexity and speed up our progress, we determined to ascertain a strong 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 surroundings that tracked over 50 on-line and offline metrics. Offline, we monitored customary data retrieval metrics like Imply Reciprocal Rank (MRR) and precision@ok to measure relevance. Crucially, this was paired with on-line real-world engagement indicators equivalent to provide unlocks and click-through charges. A key choice was implementing an LLM-as-a-judge. This allowed us to label information and assign high quality scores to each on-line query-result pairs and offline outputs. This method proved to be essential for fast iteration primarily based on dependable metrics and accumulating the labeled information crucial for future mannequin fine-tuning.
Alongside the way in which, we leaned into a number of elements of the Databricks Knowledge 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: Supplied the patterns to guage mannequin outputs and implement our information labeling course of.
- Mannequin Serving Endpoints: Environment friendly deployment of fashions straight from our catalog.
- AI Gateway: To safe and govern our entry to 3rd celebration fashions through API.
- Unity Catalog: Ensured the group, administration, and governance of all datasets used inside the analysis framework.
This sturdy framework dramatically elevated our iteration velocity and confidence. We performed over 30 distinct iterations, systematically testing main variable modifications in our Vector Search answer, together with:
- Totally different embedding fashions (foundational, open-weights, and third celebration through API)
- Varied textual content combos to feed into the fashions
- Totally different question modes (ANN vs Hybrid)
- Testing completely different columns for hybrid textual content search
- Adjusting thresholds for vector similarity
- Experimenting with separate indexes for various provide varieties
The analysis framework reworked our improvement course of, permitting us to make data-driven selections quickly and validate potential enhancements with excessive confidence earlier than exposing them to customers.
The seek for the most effective 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 way of AI Gateway. We performed short-term, iterative checks in manufacturing (lasting a number 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 combined outcomes. Whereas we noticed general enhancements in engagement metrics and efficiently eradicated the adverse impacts seen beforehand, these features had been modest—largely single-digit share will increase. These incremental advantages weren’t compelling sufficient to totally justify an entire alternative of our present search expertise.
Extra troubling, nevertheless, 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 answer. This inconsistency offered a big architectural dilemma. We confronted the unappealing alternative of implementing a posh traffic-splitting system to route queries primarily based on predicted efficiency—an method that may require sustaining two distinct search experiences and introduce a brand new, advanced layer of rule-based routing administration—or accepting the constraints.
This was a essential juncture. Whereas we had seen sufficient promise to maintain going, we would have liked extra vital enhancements to justify totally changing our homegrown search system. This led us to start fine-tuning.
Positive-tuning: customizing mannequin habits
Whereas the third-party embedding fashions explored beforehand confirmed technical promise and modest enhancements in engagement, additionally they offered essential limitations that had been unacceptable for a long-term answer at Ibotta. These included:
- Lack of ability to coach embedding fashions on our proprietary provide catalog
- Issue evolving fashions alongside enterprise and content material modifications
- Uncertainty relating to long-term API availability from exterior suppliers
- The necessity to set up and handle new exterior enterprise relationships
- 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 information and the wants of our Savers. This was made potential due to the tens of millions of labeled search interactions we had collected from actual customers through our LLM-as-a-judge course of inside our customized analysis framework. This high-quality manufacturing information turned 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 surroundings, 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 information, finally deciding on:
- One optimistic outcome (a verified good match from actual searches)
- ~10 adverse examples per optimistic, combining:
- 3-4 “arduous 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 charge, batch measurement, length, and adverse sampling methods to search out optimum configurations.
After quite a few iterations and evaluations inside the framework, our top-performing fine-tuned mannequin beat our greatest third-party baseline by 20% in artificial analysis. These compelling offline outcomes supplied 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 answer particularly optimized for Ibotta’s distinctive provide catalog and person habits patterns, delivering outcomes that exceeded our expectations and provided the flexibleness 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 affords from search outcomes, indicating improved outcome high quality and relevance. Extra unlocks are a number one indicator of downstream redemptions and income. - 6% improve in engaged customers.
This reveals a higher share of customers discovering worth and taking significant motion inside 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 straight to raised 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 influence, displaying that a big portion of our person base is now constantly discovering outcomes, enhancing the expertise throughout the board.
Past user-facing features, 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 flexibleness of this new basis, we additionally constructed highly effective enhancements like Question Transformation (enriching imprecise queries) and Multi-Search (fanning out generic phrases). The mix 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, equivalent to rising manufacturers. To deal with 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 rework the question so as to add “Greek yogurt” alongside the model title earlier than sending it to Vector Search. This gives the embedding mannequin with crucial 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 information; as an example, together with the unique model title as a question and the related yogurt merchandise as optimistic ends in a future coaching run helps the mannequin study these particular associations.
Multi-Search
For broad, generic searches like “child,” Vector Search may initially return a restricted variety of candidates, doubtlessly filtered down additional by focusing on and finances administration. To deal with this and improve outcome range, we constructed a multi-search functionality that followers out a single search time period into a number of associated searches.
As a substitute of simply looking for “child,” our system robotically runs parallel searches for phrases like “child meals,” “child clothes,” “child drugs,” “child diapers,” and so forth. Due to the low latency of Vector Search, we will execute a number of searches in parallel with out rising the general response time to the person. This gives a wider and extra numerous set of related outcomes for wide-ranging class searches.
Classes Discovered
Following the profitable last manufacturing check and the total rollout of Databricks Vector Search to our person base – delivering optimistic engagement outcomes, elevated flexibility, and highly effective search instruments like Question Transformation and Multi-Search – this venture journey yielded a number of beneficial classes:
- Begin with a proof of idea: The preliminary hackathon method allowed us to shortly validate the core idea with minimal upfront funding.
- 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 influence, enabling us to keep away from repeated dwell testing till options had been really promising.
- Do not leap straight to fine-tuning: We discovered the worth of totally evaluating off-the-shelf fashions and exhausting these choices earlier than investing within the higher effort required for fine-tuning.
- Accumulate information early: Beginning to label information from our second experiment ensured a wealthy, proprietary dataset was prepared when fine-tuning turned crucial.
- 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.
- Acknowledge cumulative influence: 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 dwell throughout all direct-to-consumer (D2C) channels, we subsequent plan to discover scaling this answer to the Ibotta Efficiency Community (IPN). This may convey improved provide discovery to tens of millions extra consumers throughout our writer community. As we proceed to gather labeled information and refine our fashions by way of Databricks, we consider we’re effectively positioned to evolve the search expertise alongside the wants of our companions and the expectations of their clients.
This journey from a hackathon venture to a manufacturing system proved that reimagining a core product expertise quickly is achievable with the correct 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.