Several graph databases now run vector search next to graph queries, including TigerGraph, ArangoDB, Memgraph, FalkorDB and Amazon Neptune Analytics, which AWS offers in London. For a UK team building a GraphRAG or agent project, the best shortlist depends on how fresh the data must be, what the plan includes and where the data sits. A raw speed ranking will not settle it.
Which graph databases support vector search?
Three things are easy to confuse. A database can store embeddings, it can index and search them, and it can use graph relationships to build an answer. The first two are database features. The third is a retrieval design you build on top.
TigerGraph documents vector operations inside its GSQL graph queries, with HNSW approximate search, cosine, L2 and inner product similarity, and automatic, incremental indexing. Memgraph supports vector indexes on both nodes and edges. ArangoDB's vector indexes use the Faiss library. FalkorDB offers vector indexing for GraphRAG in its documentation.
On AWS, Neptune Analytics lets you run similarity searches on vectors stored along with your graph, and AWS announced it in the London Region on 14 March 2024.
What do the caveats mean for your build?
Each product has a documented catch, and the catches are useful because they tell you what to test.
TigerGraph says vectors that have not yet been indexed will not be included in vector search results. If an agent answers questions about a record changed a minute ago, test how long a new vector takes to become searchable.
Memgraph runs its vector indexes at the READ_UNCOMMITTED isolation level, while the main database can run at any isolation level. Memgraph states that the rest of the database keeps its ACID properties. Ask whether that difference matters for your read and write mix.
ArangoDB says vector index training cannot be done incrementally without affecting quality. Documents added later are assigned to existing clusters, which it says can be suboptimal. A team ingesting new subject areas every week should plan index rebuilds.
How should an agentic AI application choose?
Pick the system of record first. Then decide whether graph and vector retrieval live inside it or in a separate context layer. A simple rule of thumb helps.
If the agent mostly retrieves passages, start with a vector and keyword baseline. If it must explain connections, provenance or dependencies, add graph expansion. If it updates business records, keep the transactional database as the authority and treat retrieval as supporting context only.
GraphRAG is not automatically an upgrade. Compare four arms on the same authorised corpus, with the same embeddings and answer model: vector only, keyword plus vector, vector plus graph expansion, and graph summaries. Measure answer correctness, evidence, latency, freshness and cost together. Buy the simplest arrangement that meets all of them.
Can one application hold relational, JSON, vector and graph data?
There are two sound routes. You can consolidate models in one platform, or keep a transactional source of truth and add a separate retrieval layer. Consolidation reduces synchronisation work. Separation lets each part change on its own schedule.
SurrealDB says embeddings, documents and graph relationships share a single source of truth in one database, as described in its vector search documentation. Spanner Graph is the relational-first example, because Google documents graph queries and vector search together for Spanner. Cosmos DB for NoSQL, covered below, covers the document and vector pairing.
None of these claims means every JSON operation, graph query and vector index works together without limits. Test your actual schema, transaction boundaries and query plans.
How do you set up a graph database with vector search?
Treat the setup as a sequence of checks, not a one-off install.
- Define the entities, relationships, source documents and the permissions that must limit retrieval.
- Choose an embedding model and record its name and version with the vectors.
- Create the vector index and match its dimensions and similarity setting to the embedding model.
- Load representative data and check that updates, deletions and permission changes show up in results.
- Build a vector-only baseline, then add graph expansion and measure what it adds.
- Test the whole path, including access filtering, graph traversal, network calls and evidence assembly, not only the nearest-neighbour lookup.
Which vendors let you compare graph RAG and vector RAG side by side?
None of the documentation we read describes a turnkey side-by-side comparison dashboard. What the products do offer are the parts for both arms. TigerGraph places vector functions inside graph queries, Memgraph indexes nodes and edges, ArangoDB and FalkorDB add vector indexes to graph engines, and Neptune Analytics stores vectors with graph data. Any of these can host a vector-only arm and a graph-expansion arm on the same corpus, which is the comparison that matters. Ask each vendor to run both arms on your data in a live demonstration.
Is there a fastest graph database for AI retrieval?
No source we checked supports a universal answer. The best-known head-to-head is a University of California, Merced report that uses the LDBC Social Network Benchmark. It tests Neo4j 3.5.0 and TigerGraph 2.3.1 on a single node, which are old releases and not a modern LLM retrieval workload. Treat it as history. For your own decision, run your queries, at your concurrency, with your filters, on your hardware.
How should you evaluate Microsoft for vector databases and embeddings?
Microsoft is a credible shortlist supplier for vector workloads, but "Microsoft's vector database" is too vague to buy. The company has several services with different roles, so name the exact one in your specification.
For document-style operational data, Cosmos DB for NoSQL stores vectors alongside the documents. Its page, last updated on 5 August 2026, lists three index types. The flat index handles at most 505 dimensions, while quantizedFlat and DiskANN handle up to 4,096. The latter two need at least 1,000 vectors, otherwise a full scan runs. Match your embedding model to the index before you commit.
The database stores embeddings but does not create them. You still pay for and operate the embedding model separately.
Which vector databases run in the London region on AWS, Azure and Google Cloud?
Region labels matter more than product names here. We found three confirmed London answers, one per cloud.
On AWS, Neptune Analytics is in Europe (London), as above. On Azure, Azure AI Search lists UK South in its regional support table, and the prices below are for that region. On Google Cloud, AlloyDB lists europe-west2 as London. Google also documents vector search with Spanner Graph, but we did not confirm a London configuration for it, so check that before you plan around it.
Do not accept "available on this cloud" as an answer for a UK data residency requirement. Ask for a location schedule covering primary data, replicas, backups, model calls, logs and support access. A UK database region does not settle where your embedding calls go.
What does it cost in pounds?
Azure publishes a retail price API, so one sterling baseline is checkable. We queried it on 10 November 2026 for Azure AI Search in UK South. The chart shows the hourly unit rate returned for each tier. These are individual meters, not a monthly quote, so replicas, partitions, storage and extras come on top. Check the Azure retail price API for the live figures.
Azure AI Search hourly unit rates in UK South
Retail GBP price per unit per hour for each tier, from the Azure retail price API on 10 November 2026.
View the data
| Category | £ per unit per hour |
|---|---|
| Basic | £0.08 |
| Standard S1 | £0.25 |
| Standard S2 | £1.01 |
| Standard S3 | £2.02 |
Graph vendors are harder to compare. FalkorDB's published plans start at $73 per 1GB a month for Startup and $350 per 8GB a month for Pro, quoted in dollars. We have not converted them to pounds, so ask for a sterling quote.
The cheapest plan is not the same service. FalkorDB's Startup plan lists no VPC, cluster deployment or high availability. Pro adds cluster deployment and high availability, while VPC is still marked as not included. If you need private networking and resilience, compare on that control set before comparing price.
What UK rules apply?
Hosting in the UK does not settle data protection. The ICO's guide to international transfers is the place to start when a model API, a support team or a backup sits outside the UK. Put lawful basis, controller and processor roles, retention, deletion and overseas access into the contract.
Test deletion and permission changes as acceptance criteria. A record that has been deleted or restricted should disappear from vector and graph retrieval, not just from the main table.
Questions to ask each vendor
- Are vectors and relationships queried in one engine, or synchronised through a connector?
- When do inserted, updated and deleted records become visible to vector and graph retrieval?
- Are permissions applied before candidate selection, during search or afterwards?
- Where are data, backups, logs, model calls and support access located?
- Which charges cover replicas, re-embedding, index rebuilds, egress and support?
- Can we export relationships, vectors and metadata in usable formats?
Sources
- TigerGraph, Vector Database Operations, GSQL Language Reference 4.3
- Memgraph, Vector search documentation
- ArangoDB, Vector indexes documentation
- FalkorDB, Vector indexing documentation
- FalkorDB, Pricing
- AWS, Amazon Neptune Analytics is now available in the AWS Europe (London) Region, March 2024
- Microsoft Learn, Integrated Vector Store for Azure Cosmos DB for NoSQL
- Microsoft, Azure retail prices API for Azure AI Search, UK South, queried November 2026
- ICO, A guide to international transfers
- Microsoft Learn, Supported regions for Azure AI Search
- Google Cloud, AlloyDB locations
- Google Cloud, Use vector search with Spanner Graph
- SurrealDB, Vector search model
- University of California, Merced, LDBC SNB graph database benchmarking, 2019
