Vector Database Comparison: Choosing the Right One

A RAG system needs a way to retrieve relevant content. That may be a specialist vector service or an existing database with vector search. Choose using retrieval quality, filters, operational requirements, and measured workload rather than an unsupported vendor ranking.

Product details checked on 15 September 2026. The selection criteria below are recommendations, not comparative benchmark findings.

What to Compare

Vector search retrieves records whose embeddings are close under a chosen distance measure. Business retrieval also needs access controls, metadata filters, keyword matching, and a way to connect results to their original documents.

Test those features together. A fast unfiltered search over public examples says little about an application that must isolate each customer’s private documents.

Managed and Self-Hosted Options

Pinecone provides managed vector retrieval. Its serverless costs include read units, write units, storage, and egress; paid plan minimums can apply. Vector count alone is insufficient to calculate the bill. See Pinecone’s cost documentation.

Weaviate supports hybrid retrieval that combines keyword and vector search. This is a documented capability, not evidence that it has the best filtering or relevance for every corpus. Test your queries and filters. See Weaviate hybrid search.

Qdrant offers a managed cloud deployment as well as a self-hosted engine. Using the same engine can reduce application changes when moving deployment, but data transfer, configuration, security, and operational features still need testing. See Qdrant Managed Cloud.

Chroma is not restricted to prototypes. Its documentation describes single-node and distributed architectures, with Chroma Cloud providing a managed distributed deployment. Assess the deployment you would actually run; limitations of a local prototype do not establish the capabilities of its cloud service. See Chroma architecture.

Managed hosting transfers some infrastructure work to a supplier. It does not remove responsibility for application permissions, source quality, retention, retrieval evaluation, or recovery planning. Confirm backup and SLA terms for the exact plan.

The pgvector Option

pgvector adds vector similarity search to PostgreSQL, including exact search and approximate indexes such as HNSW and IVFFlat. Its documentation also describes combining vector results with PostgreSQL full-text search for hybrid retrieval.

If the application already uses PostgreSQL, this can reduce the number of systems to operate. Embeddings can live alongside business records and their metadata.

There is no universal vector-count threshold at which pgvector becomes unsuitable. Dimensions, index settings, filter selectivity, memory, update rate, and query concurrency all affect results. Measure whether retrieval interferes with transactional workloads and whether the index meets the application’s recall and latency targets.

Moving an index to a specialist store is justified when a measured requirement cannot be met reasonably in the existing database. It is not automatically required because the corpus passes one million vectors.

How to Estimate Cost

A meaningful quote needs more than “one million vectors”. Specify:

Vector dimensions and numeric representation.
Metadata and source-text storage per record.
Number of records and expected growth.
Reads, writes, deletions, and re-indexing frequency.
Filters, query concurrency, and response size.
Replicas, backups, region, transfer, and service commitments.

For self-hosting, add infrastructure and operational labour. For managed hosting, check minimum spend, included allowances, and the unit charged for each operation. Keep embedding generation and the answer model’s fees separate.

Do not compare monthly GBP ranges without publishing these inputs. Use the vendor calculator or a pilot’s metered usage and record the date, currency, and configuration.

A Practical Evaluation

Prepare representative questions with the documents a correct answer requires. Include exact identifiers, paraphrases, no-answer cases, stale documents, and queries that would otherwise retrieve another tenant’s data.

Check	Evidence to record
Retrieval quality	Relevant documents found at the chosen result count
Access isolation	Attempts to retrieve documents outside the user’s permissions
Performance	Latency and errors at the expected concurrency
Freshness	Time until additions, changes, and deletions affect results
Recovery	A completed backup restore or rebuild from source
Cost	Metered reads, writes, storage, compute, and operational time

Use the same corpus and comparable settings on the shortlisted stores. Approximate search trades recall against speed and resource use, so compare the operating point your application needs.

Decision Framework

Start by evaluating pgvector if PostgreSQL already holds your data. Include managed specialist services when their operating model or retrieval features meet requirements you do not want to implement yourself. Consider self-hosting when the team can maintain the service and there is a concrete control or cost requirement.

Keep source documents, chunk identifiers, metadata, and embedding configuration exportable. Moving the database does not necessarily require new embeddings if the destination supports the existing representation. Changing the embedding model generally requires a separately evaluated index rebuild. Keep this choice separate from the AI vendor assessment.

Where Fernside Fits

We design and build the AI systems around this choice. chunking, embedding, retrieval, and the eval set that tells you the database is doing its job. If the open question is still “managed vs Postgres vs specialist,” that is advisory work before anyone provisions a cluster.

Need a RAG architecture recommendation against your corpus and team? Book a discovery call and we will shortlist on scale, residency, and ops load, not the loudest benchmark chart.