Your first decision is whether you'll run the database yourself or hand it to a managed service. Open source engines give you control and portability; managed services take operations off your plate. After that, test how each stores and queries your own embeddings as the data grows, because behaviour at scale matters more than a demo.
Chroma pitches itself simply as open source search infrastructure for AI, and plenty of developers start there for a first retrieval prototype. When a project grows, the questions change: how data is stored and queried at scale, and who keeps the system running at three in the morning. Below are seven alternatives, in no particular order, from open source engines to managed platforms.
Two questions to settle before comparing features
How will it behave at scale?
A vector store that feels instant with a few thousand embeddings can behave very differently with many millions. Index type, memory use, filtering on metadata and how updates are handled all start to matter. The only honest test is to load a realistic slice of your own data and run the queries your application will run, with the filters it will use.
Who runs it?
Running an open source engine yourself gives you control over where data lives and how it's tuned, but someone owns upgrades, backups and scaling. A managed service takes that away, and ties you more closely to one vendor's way of working. Many teams prototype on one and move to the other, so check how easily data moves between the two.
- Can you self-host, use a managed service, or switch between them?
- How does filtering on metadata affect query speed?
- Does it combine keyword and vector search, or only vectors?
- What happens to queries while data is being updated or re-indexed?
Seven options at a glance
| Tool | Kind of engine | Good for |
|---|---|---|
| LanceDB | Multimodal lakehouse | Training data curation alongside retrieval |
| Marqo | Ecommerce AI search | Product search and recommendations |
| Qdrant | Open source vector engine | Self-hosted vector search with a clean API |
| Typesense | Open source search engine | Keyword and vector search in one engine |
| Vespa | AI search platform | Large search and recommendation systems |
| Weaviate | AI-native database | Building full AI features, not just storage |
| Zilliz | Managed Milvus service | Managed vector search on major clouds |
Each option, A to Z
In no particular order. Each line reflects how the vendor describes its own product.
LanceDB
Best for: ML teams curating large multimodal datasets and engineering features for model training, which the vendor frames as a lakehouse for AI.
Marqo
Best for: Ecommerce teams wanting AI-native search, recommendations and merchandising trained on their own catalogue, per the vendor.
Qdrant
Best for: Developers wanting an open source vector search engine written in Rust, with fast similarity search through a simple API, as the vendor describes it.
Typesense
Best for: Engineers needing fast, typo-tolerant site and app search with vector search added, open source and self-hostable, per the vendor.
Vespa
Best for: Teams building AI search, agents, personalisation and recommendations on one platform, as the vendor positions it; a free trial is named.
Weaviate
Best for: Builders shipping complete AI applications who care about hallucination, data leakage and vendor lock-in, per the vendor; a free plan is named.
Zilliz
Best for: Teams wanting a fully managed service powered by Milvus, combining real-time vector search with lake-scale data across major cloud regions, per the vendor.
Grouped by the job you're doing
Open source engines you can host
Qdrant is an open source vector search engine written in Rust with a straightforward API. Typesense is an open source search engine for site and app search that adds vector search and self-hosting to typo-tolerant keyword search, which suits teams who need both kinds of query.
Managed platforms for production
Zilliz offers a fully managed service built on the open source Milvus project, available on AWS, Google Cloud and Azure. Weaviate focuses on complete AI applications and, according to the vendor, less hallucination, data leakage and lock-in.
Search and retrieval beyond plain vectors
Vespa describes itself as an AI search platform covering search, agents, personalisation, recommendations and retrieval. Marqo trains search and recommendations on your product catalogue, which suits ecommerce teams more than general developers.
Want AI doing the heavy lifting in your marketing?
I build the systems that handle the boring 80 percent, so you get your week back. Done properly, with the human kept in.
Data for model training
LanceDB frames itself as a multimodal lakehouse for curating training data and engineering features, so it suits teams whose work goes beyond retrieval into building models. For more options, see my guide to vector databases and my list of RAG tools.
Practical advice on picking an engine
If you're still prototyping, don't overthink it. Pick an open source engine you can run on a laptop, keep your source data safe, and move on to the part of the product users will see. The time to compare engines hard is when you know your data size, your query patterns and how fast results need to come back.
When that day comes, run your own test set against two or three candidates rather than trusting benchmarks. Benchmarks use someone else's data and someone else's queries. If your app leans on metadata filters, or mixes keyword and meaning-based search, that's what you should measure. And if nobody on the team wants to be on call for a database, that's a strong reason to go managed.
From prototype to production
Moving off a prototype store is a good moment to tidy up how your retrieval works. Before you switch, work through this list:
- Keep the raw text and metadata alongside your embeddings, so you can re-embed rather than export vectors
- Write down the embedding model and settings you used, since mixing models breaks search quality
- Build a small set of test questions with known good answers and run it against every candidate
- Measure query speed with the filters your app really uses, not with plain similarity search
- Plan how you'll back up and restore the data, whether you host it or a vendor does
Once the new engine is live, keep the old one running read-only for a short while. If answers look worse, you'll want to compare the two side by side before anyone blames the model.
Do you make vector database or AI search software? Tool makers can ask to be considered for a featured place on this list.
Want your tool on this list? Featured spots and reviews here are paid partnerships. Here's how to partner with me.
Frequently asked questions
What is a vector database?
It stores embeddings, which are numeric representations of text, images or other data, and finds the items closest in meaning to a query. It's the retrieval layer behind most AI search and chat features.
Do I need a dedicated vector database?
Not always. Some search engines and databases now add vector search to what they already do. A dedicated engine makes more sense as data and query volume grow.
Should I self-host or use a managed service?
Self-host if you need control over where data lives and have someone to run it. Choose managed if you'd rather not handle upgrades, scaling and backups.
Can I move my embeddings to another database?
Usually, but it's easier to re-embed from the original text with the same model. Keep your source data and metadata so you're never stuck.
What is hybrid search?
It combines keyword search with vector search, so exact terms such as product codes still match while meaning-based results fill the gaps.
