Why Uber Picked Amazon OpenSearch for Semantic Search — Capturing User Intent Better
Intro
Uber moved. They chose Amazon OpenSearch to power semantic search. That matters. It shows the shift from keyword matching to understanding intent. It shows where search systems are going.
What this means, in one line: Uber wants search that thinks like a person. They want results that match intent, not just words.
Who I am
I’m Ermenildo Valdez Castro, Jr. — Ernie Castro to many. I’ve built search and ranking systems at Google, Microsoft, Amazon, and Zulily. I’ve fixed slow queries. I’ve tuned click models. I’ve shipped features that helped real people find what they needed. I’ll tell you what this Uber news means for engineers and product leaders. 😊
What is semantic search?
Semantic search uses vectors and embeddings. It maps meaning into numbers. Then it finds close vectors instead of just matching text. It understands synonyms and context. It captures user intent.
Keywords: semantic search, vector search, embeddings, user intent, OpenSearch.
Why Uber chose Amazon OpenSearch
- Native vector search. OpenSearch supports k-NN and vector indexes. That cuts integration work.
- Scale and ops. OpenSearch scales on AWS well. Uber runs huge traffic. AWS integration matters.
- Faster experimentation. Teams can test new ranking models quickly.
- Cost and control. OpenSearch gives predictable ops costs compared to black-box SaaS.
I remember building a ranking layer at an e-commerce site. We tried a separate vector store and a search engine. The glue code killed us. When everything lived in one cluster, iteration sped up. That’s what teams at Uber likely valued.
Technical benefits: What OpenSearch offers
- k-NN vector indexes for semantic embeddings.
- Hybrid search: text + vector scoring in one query.
- Plugin ecosystem and AWS-managed options.
- Familiar APIs for Elasticsearch users.
This combo matters. Semantic match finds intent. Keyword signals keep precision. Hybrid search gives both.
Trade-offs and risks
- Model choice: Which embedding model? Open-source, AWS models, or custom? Each choice affects relevance and cost.
- Operational load: Running many vectors at scale needs tuning (shards, memory, indexes).
- Vendor lock-in: Moving deep into OpenSearch-specific features can make future migrations harder.
- Privacy and data governance: Where do embeddings live? Who can query them? These matter for regulated data.
From my Amazon days I learned to ask two questions before adopting any search tech: Will it scale? Will it let us experiment fast? OpenSearch checks both boxes for many teams.
How to approach semantic search (practical tips)
- Start with a hybrid approach. Combine BM25 with vector scores.
- Choose a baseline embedding model. Test with a small labeled set.
- Run A/B tests focused on intent metrics: task completion, not just CTR.
- Monitor latency and cost. Vector joins can spike memory use.
- Build offline evals and relevance funnels. Keep a human in the loop.
A quick anecdote
At Zulily we ran an experiment that seemed small. We changed the way synonyms were weighted. CTR rose 3% but returns dropped. We thought we had a win. We didn’t. We hadn’t measured task completion. The lesson: a better-looking result isn’t always a better result for the user. Semantic search helps the user complete tasks. But you must measure the right things.
What product leaders should know
- Semantic search improves discovery. It makes vague queries useful.
- It can reduce no-results queries.
- It requires investment: models, infra, and evaluation.
- Expect faster iteration when search and vector store are integrated.
A small word of wisdom
Search is about truth in signals. Tune the system to serve the user, not to chase vanity metrics. That line sounds like something my grandfather would say in church. It fits here too.
Conclusion
Uber’s move to Amazon OpenSearch is practical. It’s about faster experiments, integrated vector search, and scale. For teams building search, the lesson is simple: marry vectors with good evaluation. Put the user first.
If you build search, start hybrid. Measure the right outcomes. Iterate quickly.
I’m Ernie Castro. I write about tech and life. Follow for more short lessons from real systems. 🚀