Zen v17 AI: The Same Database, Now With Opinions About Your Vectors

A quick disclosure before we start: if you know my writing, you know it. I usually write about Informix, the database equivalent of that reliable relative who never misses a family dinner. Zen is, by comparison, the cousin I am only now getting to know properly. Credit, or blame, goes to Desmond Tan, VP of Engineering for Zen and VectorAI DB, who talked me into it over what I recall was an unreasonably convincing amount of enthusiasm. For the record, I love all our databases equally. I simply write about some of them more than others, which is not the same thing, and if you ask Informix, it will agree.

Here’s a scenario that will feel familiar if you build software that ships to other people’s hardware: your customer wants AI features: semantic search, a recommendation engine, a support agent that actually knows the product. And your first instinct, because it is everyone’s first instinct, is to reach for a cloud vector database.

Then reality shows up, uninvited, as it does. Your application runs in a factory with no reliable internet. Or a hospital, where patient data is legally not allowed to leave the building, let alone phone a cloud provider about it. Or a point-of-sale terminal that needs to work whether or not the network is in the mood. Suddenly “just call an API” is not an answer:  it is a compliance risk.

This is, not coincidentally, exactly the world Zen has lived in for decades: embedded, small-footprint, zero-DBA, running quietly inside other people’s products (Sage, Global Shop Solutions, and a long list of OEMs who would rather not think about their database at all, which is the whole point). Zen 17 keeps that promise and reputation intact and adds the one thing that the world has been missing: native vector support.

Not a Bolt-on. The Real Thing.

Zen v17 AI vector capability is not a wrapper glued onto somebody else’s engine at the last minute. It is built on the same technology behind Actian VectorAI DB, Actian’s portable, local-first vector database. Which means Zen inherits, natively, assembled with the kind of precision Desmond and I both love in our ramen (Desmond, you know the place: we are due for a visit):

• HNSW (Hierarchical Navigable Small World) indexing for approximate nearest neighbor search, the standard approach for fast, high-recall retrieval at scale
• Model-agnostic embeddings: bring vectors from OpenAI, Anthropic, Cohere, Hugging Face, or your own fine-tuned model; Zen does not care where they came from; it just stores them.
• Multimodal support: text, images, audio, video, all storable and searchable as vectors, not just strings pretending to be clever.
• The same deployment targets Zen has always shipped to: edge servers, embedded devices, disconnected and air-gapped environments, right down to a Raspberry Pi, if that is your idea of a good time.

In Actian’s own benchmark testing (VectorDBBench methodology, 10 million vectors, identical self-hosted hardware), this engine delivers up to 22x the throughput of open-source alternatives like Milvus and Qdrant Local at that scale, with sub-15ms local query latency and 99% recall holding steady as the dataset grows. Your mileage will vary with your hardware and your data, as you’d expect, but the facts are real: this is not a toy vector store bolted on with a clumsy pipeline for the sake of a feature checkbox.

Why This Matters if You are an OEM

If you have been embedding Zen into your product for years, you already know the pitch: small footprint, no DBA required, runs wherever your application runs, asks for nothing in return. It is, in its own quiet way, doing what Informix has spent decades doing for larger deployments, minus the DBA team and the server room. Zen 17 extends that exact value proposition to AI:

• Manufacturing: Predictive maintenance and quality inspection on the plant floor, no cloud round-trip required, no signal, no problem, no drama when your ISP is down.
• Healthcare: Clinical decision support and record search that keeps patient data on-premises, because “HIPAA-compliant AI” is not supposed to be hosted on a cloud you know nothing about.
• Retail and point of sale: Semantic search and recommendations that work whether the store’s internet is having its moment or not, all that with a crazy fast low-latency for a real conversation with the customer.
• Anything disconnected, regulated, or latency-sensitive: Which, if you build embedded software, is most of what you build anyway

You do not need a new architecture, a new vendor relationship, or a new line item labeled “the AI database.” You need the database you already trust to do one more thing well. That is the whole idea, and it always has been.

One More Thing (for Later)

There is a quieter implication here worth flagging before I let you go: a database that can store vectors locally, right alongside your existing data, is also a database that can start doing something more interesting with metadata. Active, queryable metadata, sitting exactly where your application already lives, is a piece of a puzzle considerably bigger than vector search. I am leaving that thread dangling on purpose: it deserves its own post rather than a rushed footnote here. More on that soon.

For now: Zen v17 AI is available, native vector support included, no cloud dependency required. Go build something that does not need to phone home to work.

And Desmond, since you will inevitably ask: yes, this took longer to write than expected; no, it is not because I liked Informix better; I was simply meeting the new cousin for the first time. Next round is on you.

About Author

About Jean-Georges Perrin

Jean-Georges Perrin has been part of the Informix community for nearly 30 years. Elected to the Board of Directors of the International Informix Users Group (IIUG) in 2002, he served for 15 years and was the first non-US citizen elected to that board. A Lifetime IBM Champion, a distinction he earned through 16 consecutive years of recognition starting in 2009, when he became the first French citizen to receive the title, Jean-Georges has authored two ebooks on Informix, presented at IIUG conferences on topics ranging from 4GL modernization to proving that Informix is not just for legacy applications, and contributed to the Informix community across three continents. He serves in Actian’s CTO Office, focused on data standards and AI strategy. But when HCL Informix does something worth writing about, old habits kick in. He chairs the Linux Foundation's Bitol project, where he leads the development of open standards for data contracts (ODCS) and data products (ODPS). He is the author of several books, including Implementing Data Mesh (O'Reilly) and Spark in Action, 2nd edition (Manning), and is currently writing Building Data Products for O'Reilly. Working on databases in college made him hate them. Then karma introduced him to Informix. He never left.