The Zero Results Problem, Solved by Vespa.ai

The Zero Results Problem

The Zero Results Problem, Solved by Vespa.ai

Shoppers don’t search the way product catalogs are written. Misspellings, synonyms, slang, and natural language queries routinely break traditional keyword-based search, resulting in zero-result pages that end sessions and drive customers elsewhere.

In this webinar, you’ll learn:

– Why zero results happen and how lexical, exact-match search fails to capture user intent
– The hidden revenue impact of empty results pages on conversion and trust
– How hybrid lexical + vector search works to combine precision with semantic understanding
– How vector embeddings interpret context and meaning, even for unfamiliar or conversational queries
– What full-coverage retrieval enables, from fewer dead ends to higher relevance and conversion rates

You’ll walk away understanding how hybrid search reduces zero-result queries and lays the foundation for scalable, AI-powered product discovery across your eCommerce experience.

Hello everyone. Thanks so much for joining us today. My name is Bonnie, and I lead product marketing at Vespa, and I’m joined by Zohar, who is our lead Solution Architect. Now today, we’ll be talking about how to solve the zero results problem in E commerce. This is the first of a series of webinars, so if you haven’t registered for the full series, be sure to check that out in the Resources tab on the right side of your screen.

Now, before we get started, a few housekeeping notes. This session is being recorded, and you will get a copy of the recording and the slides within 48 hours after the webinar ends, we’d love to get your thoughts and questions so you’ll see a Q and A option on the right side of your screen as well. Drop your questions here, and we’ll address these at the end of the session. With that, let’s go ahead and get started now.

Today, I’ll do a quick overview of the zero results problem. Zohar will talk through ways that you can solve that problem, and then we’ll do a quick demo of Vespa solution and the E commerce experience there, and of course, at the end, we reserve some time to answer questions, so don’t forget to drop your questions in that tab on the right side of your screen. All right.

Problem Overview

Bonnie Chase: So let’s start with something every one of us has experienced as a shopper. You type in exactly what you’re looking for, and you’re met with nothing, no products, no guidance, just a dead end. And that’s what we mean by the zero results problem. Customer query returns no items, even though the catalog does contain things that could be relevant. And this hurts on multiple levels.

From the user’s perspective, it’s frustrating. It breaks trust. The shopper feels like this store doesn’t understand me, and they often leave and you know, possibly to a competitor. From the seller’s perspective, it’s just as painful, zero results, revenue, stranded inventory and lost opportunities to connect and manage.

Now, the key nuance here and this matters for everything that follows is that this isn’t just a buyer problem or a seller problem. It’s a shared relevance problem where both sides have equally critical needs that the system has to balance.

So let’s talk about why zero results happen so often. Well, modern shoppers don’t search in simple terms anymore. They come in with very specific intent. They’re looking for particular shapes, colors, styles, brands, sizes. They filter by technical specs, compatibility, price ranges and availability. And of course, they expect personalization. They don’t want a generic list. They want results that feel like they were chosen for them, and this is where traditional keyword matching breaks down.

If the system only understands exact strings or treats filters as hard constraints, it’s incredibly easy to over constrain a query and end up with nothing. But the shopper didn’t mean show me nothing. They meant, show me the closest, most relevant alternatives that respect my intent.

Of course, relevance doesn’t stop with the buyer. Merchants also have their own set of priorities that are just as real and just as dynamic. They need to balance stock levels, pricing strategies and promotions they may want to push, push certain items for business reasons, clearing excess inventory, promoting a current collection before a new launch, or spotlighting high margin products.

And then there are external drivers. No one can ignore seasonality, cultural moments, weather shifts, even regional trends. The challenge is that many commerce systems force a trade off. You either optimize strictly for buyer intent or you override relevance with merchandising rules.

In reality, the best outcomes happen when the system can blend both without hacks, manual overrides or fragile pipelines. So now I’m going to hand it over to Zohar to talk a little bit more about solving the zero results problem.

Zohar Nissare-Houssen: Thank you, Bonnie, so now let’s double click on why the zero result problem happened in the first place. And you may be familiar with, you know, some of the challenges.

First of all, shoppers don’t speak in catalog language, so your search system might have indexed an item as athletic performance running footwear, but customers are actually searching for running shoes. So traditional keyword first search system depends on exact or near exact overlap and any mismatch between whole products are indexed and how users are searching could eventually create either non relevant results or zero result event.

Secondly, keyword fixed on scale. I mean, you can add synonym rules like sneakers equal trainers, hoodie equal sweatshirt, but this can quickly become unmanageable as query volume, language variation and catalog size grow, manual rules fall behind and long tail query inevitability will break through the cracks.

And finally, filters and constraints can eliminate results through overly rigid combination, or might actually end up in non relevant results. So as a shopper searches then applies for a brand’s size, color and filter combination, they can suddenly end up with either zero result or non relevant results.

This happens because the combination of filters become too restrictive, like items that partially match user intent are excluded entirely, instead of being degraded gracefully sample they could be ranked lower or surfaced as near match. And as this hard constraint stack, the system will end up returning non relevant reasons.

In a traditional setup, what we find with our E commerce customer is that they have separate components. They have a search query processing component, they have a product reprimanded system, and they have a re ranker, and each one of them is operating kind of independently, passing, you know, data to each other.

This siloed architecture creates real problem when you take, for example, ranking, when it’s separate from retrieval, you cannot incorporate real time signals like user preference, click History or firm browsing session context, you are ranking based on static scores computed earlier, not what’s actually happening right now.

The same goes for business rules. So if you have some business rules that you’d like to apply in real time, like the actual inventory level, and you want to boost the product that have excess inventory, it becomes quite difficult to coordinate between three different system which are isolated from each other.

What Vespa Enables

So now let’s talk about Vespa. First of all, Vespa is one unified engine, combining retrieval, ranking and serving in real time.

Let’s discuss a little bit what this enables. First of all, it enables unified retrieval. Unified retrieval means a semantic search, lexical search and structure filtering, all happening in a single query pipeline.

Secondly, Vespa is capable of doing real time ranking on incorporating all real time signals like we discussed, just, you know, inventory in real time or the user context.

Finally, Vespa is a production grade skill. Vespa can handle billions of documents and hundreds of 1000s of query per second with millisecond response time.

So what are tensors? Tensors in Vespa are numerical structure.

Secondly, on top of these, tensors, which are, you know, can represent complex data modality you can, you can do complex math, right?

So this really unlocks a lot of possibility when it comes to e commerce and addressing the zeros result problem. And Bonnie is going to discuss this more in the next section. Thank you.

Bonnie Chase: Thank you, Zohar. And if you have any questions about tensors, again, drop questions in the Q and A portion of the screen.

One of the things that Zohar mentioned was Vespa being real time.

And you know, if we were to sum it up with what you gain with Vespa, I mean, you get that seamless user experience you get the fewer results, and the greater, the fewer zero results, and the greater satisfaction you get.

Now, so we can show better than we can tell, we’d like to go ahead and share a short demo so that you can see what a solution like this looks like.

Demo

Zohar Nissare-Houssen: Okay, thank you, Bonnie, so without further ado, let’s jump into the demo.

All right, so I hope this was useful in demonstrating one particular concept about vspa, in terms of whole tensors, allows to represent very nuanced level of information and allows you to represent very accurate results and surface items that are the most relevant to what a user is looking for. Thank you for attending this thing.

Conclusion

Bonnie Chase: Now the zero results problem is real and it’s damaging, but it’s absolutely solvable.

The core issue isn’t missing data, it’s missing flexibility and intelligence and how relevance is computed.

Now, with that, we would love to take some questions. We do have a couple in the Q and A ready to go. But if you haven’t asked your question yet, there’s still time, so go ahead and drop your questions in now.

Questions and answers from the end of the session.

Are the tensor representations based on the image or text?
The tensor is generated directly from the image. During the data engineering process, a vision-language model (VLM) is applied to each product image and prompted to identify relevant attributes (such as style, color, shape, or material). The model is guided with a predefined attribute space and value ranges.

The output is a structured tensor, essentially a set of key–value pairs, that captures visual semantics in a form Vespa can efficiently retrieve and rank.

This enables search and personalization based on what the product looks like, not just how it’s described in text.

How does Vespa differ from Lucene-based solutions?
The key difference is advanced ranking capabilities.

– Blends business logic and traditional search relevance into a single ranking system
– Uses a logical vertical data architecture instead of traditional sharding
– Supports large-scale vector search (e.g., HNSW graphs) using full node memory efficiently
– Designed to avoid heavy reindexing and scaling limitations seen in some Lucene-based systems

What scale can Vespa support?
Proven to handle billions of products and offers.

– – Supports billion-scale vector search and ranking efficiently
Offers optimization strategies for memory and performance

What measurable improvements have customers seen?
Reported improvements include:

– Up to 19% increase in conversion rate (CVR), especially with personalization
– 2.5–5% increase in average page views
– 5–10% increase in category and catalog views
– Personalization is especially impactful for catalogs with similar items (e.g., fashion, cars)

Can Vespa be used for sponsored offers and ads?
Yes.

– Offline scoring models can be imported as attributes and used in ranking
– Custom models can be deployed directly (e.g., via ONNX format)
– Supports tensor operations for advanced model integration
– Used in large-scale ad-serving environments
– Enables integrating user profiles into ad ranking for better relevance

Why does staged ranking matter in vector search?
Staged ranking improves both performance and cost efficiency.

Example with 1 billion items:

– Stage 1: Lightweight model (e.g., binarized vectors) for fast, memory-efficient candidate selection
– Stage 2: Heavier model (e.g., full-precision FP32 vectors) applied only to top candidates

Benefits:

– Faster response times
– Lower memory usage
– Reduced infrastructure cost
– High precision without scanning the full dataset

How is Vespa different from other vector databases?
Vespa is a full AI Search platform, not just a vector database.

It combines:

– Traditional lexical search
– Vector search
– Ranking framework
– Model inference
– Custom processing modules
– HTTP handlers and extensibility

Unlike many vector databases:

– No need for separate systems for lexical search, ranking, or inference
– Everything runs in one integrated system close to the data
– Reduced latency and system complexity