Actian VectorAI DB Brings Vector Search to the Edge With 22x Claim

What Actian shipped

Actian, the data and AI division of HCLSoftware, introduced VectorAI DB on April 28, 2026, at the AI Dev 26 x SF event. It is a portable vector database for production AI in regulated, disconnected and edge environments where cloud-native tools fall short.1 The headline claim is speed. Actian says VectorAI DB beat leading open-source vector databases by more than 22 times on query throughput, measured with the VDBBench methodology on 10 million vectors running on identical self-hosted hardware.1

This is Actian's first vector database.5 It puts a company best known for Ingres, its relational database, into one of the most crowded parts of AI infrastructure.8 The more useful story is less about the 22x figure and more about the market Actian is choosing. It isn't trying to beat Pinecone or the hyperscalers at managed cloud scale. It is aiming at plant floors, hospital data centers and air-gapped government sites, where cloud vector services often can't operate.

Unpacking the 22x number

The raw figure is 745.2 queries per second on 10 million vectors at 768 dimensions.1 Actian's footnote says the comparison was against Milvus and Qdrant Local on the same self-hosted hardware.1 The company ran the tests itself on April 3, 2026. It reports that VectorAI DB kept 72% of its throughput when scaling from 1 million to 10 million vectors, while the competitors fell to about 12% of theirs.11

That retention figure matters more than the headline multiple. Throughput collapsing as an index grows is a familiar way production retrieval systems fail. Reporters picked up on this: the problem is no longer getting vector search to work at all, but keeping it stable under real production loads.2 Actian's product page adds a 13-millisecond p99 latency at 99% recall on self-hosted hardware.12

The caveats are real, and most coverage passed over them. At launch, the Actian team said the tests ran "with zero vendor optimizations applied to any database." That sounds fair, but it also means Milvus and Qdrant were measured without the tuning their operators would normally apply. One independent analysis of the launch says plainly that the 22x figure was benchmarked by the vendor, has not been independently verified, and should be treated as marketing until someone reproduces it.6

The comparisons also shift between Actian's own materials. The press release names Milvus and Qdrant Local as the baselines.1 The product page's comparison table lists Pinecone at roughly 33 QPS and Qdrant at roughly 34 QPS, alongside VectorAI DB's 745.12 Pinecone is cloud-only, as Actian's own FAQ notes, so a like-for-like test on self-hosted hardware is hard to picture.12 Buyers should ask exactly which configuration produced each number.

Write performance is another open question. Asked on Product Hunt about frequent embedding updates, a team member said loading 10 million vectors took 27,170 seconds, about 2,000 seconds faster than Qdrant Local and more than 12,000 seconds faster than Milvus. That is a relative advantage. Still, a load time of roughly seven and a half hours is a reminder that ingestion, not just query speed, will decide whether this works for streaming sensor data at the edge. Teams with high-velocity embedding pipelines should test that themselves before relying on read benchmarks.

The real pitch: retrieval that stays put

Coverage agrees on why the product exists, even where it differs on the details. Actian argues that most vector databases assume always-on connectivity and distributed cloud infrastructure, and so struggle with latency and data-residency requirements as AI moves from RAG toward autonomous systems and robotics.1 CTO Emma McGrattan put it as bringing AI to the data rather than moving data out of governed systems.1 McGrattan also wrote O'Reilly's Vector Databases for Enterprise AI, which was released alongside the product.14

The supported hardware runs from NVIDIA Jetson and Raspberry Pi boards up to on-premises servers and the cloud, with the same architecture and APIs on each.1 Constellation Research added industrial edge servers to that list and noted support for multimodal data and multiple models.5 Actian's product materials say the database keeps serving queries during network outages and syncs back to the data center when connectivity returns.12 That makes it closer to an occasionally connected data system than a traditional cloud service. Integration is the architectural point BigDATAwire highlighted. VectorAI DB can be embedded directly in an application rather than reached through a pipeline to an external store, so retrieval runs locally.2 TechTarget made the same distinction, calling the database portable rather than cloud-native.13

The target industries are what you would expect: manufacturing, healthcare, financial services and government.1 Security features include AES-256 encryption and customer-managed API keys, with deployment configurations aligned to HIPAA, GDPR, ISO 27001 and SOC 2 Type II.1 One product review points out that these are described as compliance-ready configurations, not audited vendor certifications, so the deploying organization carries the certification burden.16 Regulated buyers will notice.

Under the hood

Later technical coverage filled in what the launch materials left vague. VectorAI DB supports HNSW and IVF indexes, both of which can be combined with vector quantization.3 Actian recommends IVF when memory is tight or data changes constantly, because HNSW's layered graph costs more memory and is expensive to repair on every insert and delete.3 For edge hardware it offers Float16 storage, Int8 scalar quantization and rescoring that re-checks top candidates against the original vectors.8 It also supports hybrid dense-plus-keyword search with Reciprocal Rank Fusion or Distribution-Based Score Fusion, named vectors and a filter language.8

Developers get Python and JavaScript SDKs, integrations with LangChain, LlamaIndex and Hugging Face, and a single Docker container that works with Kubernetes, Helm and Terraform on ARM and x86. The client uses gRPC and REST ports by default.8

The coverage diverges on one point. One knowledge-base entry describes VectorAI DB as a multi-cloud managed service on AWS, Azure and GCP, built on FAISS with on-disk IVF indexes.6 Actian's own comparison table lists its cloud option as "self-managed," not fully managed.12 A separate review counts the lack of a fully managed service as a drawback.16 The managed-service description looks wrong. VectorAI DB is software you run yourself, although Actian does say it is available through AWS Marketplace.12

Pricing, and the Postgres question

The launch offered a free 30-day trial and a free Community Edition.11 The Community tier is capped at 5,000 vectors, which is enough for a prototype and not much more.8 Published list pricing starts at $417 a month for 1 million vectors (Starter) and $1,250 a month for 5 million (Growth). Enterprise and Edge tiers are quoted individually.16 One review argues that pricing by vector count means cost grows with corpus size whatever the traffic.16 It adds that Starter costs noticeably more than running Qdrant, Milvus or pgvector yourself for free.16 A case study on the developer launch notes that Actian later added a monthly pricing option to make evaluation easier.

The pgvector comparison is the one many teams will make. For organizations already running Postgres, adding vector search as an extension is the default path because it adds no new system. Actian's bet is that pgvector and the open-source vector stores don't serve deployments on Jetson boards, in disconnected facilities or under strict sovereignty rules well enough. On that point, analysts quoted by TechTarget mostly agreed. Matt Aslett said putting vector database capabilities on embedded and edge devices is a differentiator, and that a cloud-only product could be seen as incompatible with a sovereign AI strategy.13 Stephen Catanzano said Actian appears to be doing more than catching up, because of its focus on portability, compliance and edge performance.13

The reading

The coverage gives a consistent picture. Constellation Research, TechTarget and BigDATAwire all describe VectorAI DB as a governance and deployment-location product first, with speed as a supporting point.5132 That framing seems right. The vector database market has split into fully managed services such as Pinecone and Weaviate, open-source systems such as Milvus and Qdrant, and vector search added to existing AWS and Google Cloud data services.2 Actian's experience is in on-premises and edge software, and its own team says VectorAI DB grew out of customer requests in exactly those environments.

The 22x claim got Actian attention, but it won't decide the product's fate. Until independent benchmarks with tuned competitors appear, the figure is a vendor number.6 What could last is the category Actian is staking out: vector retrieval that runs where data is created and stays under its owner's control, from a Raspberry Pi to a data center. If regulated buyers find that ingestion holds up, edge sync works, and the compliance configurations survive audit, Actian will have a real niche among the cloud-first vendors. If not, the free open-source options, pgvector among them, will remain good enough for most of the market.