What is the best vector database in 2026?
The best vector database is often the one you already run. When you do need a new one, the monthly floor matters more than the benchmark.
You're building a RAG pipeline, an AI agent that answers from your docs, or semantic search. You need somewhere to store embeddings and pull back the nearest ones in milliseconds. That's the job of a vector database.
The short answer. Qdrant is the best pick for most teams: a free tier that never expires, open source, native hybrid search and usage billing with no published minimum. Supabase with pgvector wins if your app already lives in Postgres. Weaviate ships hybrid search on every plan. Pinecone is still the fastest way to get an index running, but the bill tracks your read traffic. Zilliz Cloud is built for billions of vectors, and Turbopuffer is the cheap object-storage option at scale.
What most rankings skip: the real floor once you leave the free tier ($45 on Weaviate Flex, $50 on Pinecone Standard, $16 on Turbopuffer Launch), compliance per plan, and the cost of switching later, which usually means re-embedding everything.
The ranking is editorial: a sponsored spot is always labeled on its card, and no affiliate link changes a score.
The best vector database for your project
Free forever, then pay as you go
Best if you're on Postgrespgvector, $25/mo on Pro
Easiest hybrid searchHybrid on every plan, free included
Easiest to get startedBuilder at a flat $20/mo
Best for massive scaleManaged Milvus, 5 GB free
Cheapest at scaleFrom $16/mo, object storage
6 vector databases compared
This ranking is for teams shipping RAG, AI agents or semantic search, from a few thousand to hundreds of millions of vectors. Every ranked database publishes a price on its official page.
| Best for | What sets it apart | Free plan | Team size | Visit | ||||
|---|---|---|---|---|---|---|---|---|
| 1 | Qdrant | Best overall | No published minimum on Standard | 4.4/5 | Free, then usage | ✓ | Startups, SMBs, product teams | Visit → |
| 2 | Supabase (pgvector) | Best if you're on Postgres | Vectors next to your app data | 4.3/5 | $0, then $25/mo | ✓ | Teams already on Postgres | Visit → |
| 3 | Weaviate | Easiest hybrid search | Hybrid search on all plans | 4.2/5 | $0, then $45/mo | ✓ | SMBs and product teams | Visit → |
| 4 | Pinecone | Easiest to get started | Nothing to operate, named EU regions | 4.1/5 | Free, then $20/mo | ✓ | Teams without ops, SaaS | Visit → |
| 5 | Zilliz Cloud (Milvus) | Best for massive scale | Built for billions of vectors | 4.0/5 | $0, then usage | ✓ | Scale-ups, large catalogs | Visit → |
| 6 | Turbopuffer | Cheapest at scale | Object storage pricing, EU regions | 3.9/5 | From $16/mo | — | SaaS with many tenants | Visit → |
Official list prices in USD, checked in 2026 on each vendor's pricing page. Pinecone and Supabase scores come from our full reviews.
Is your tool missing from this page? Email us: hello@hackceleration.com
How we compared and scored
We compared these databases on what makes a RAG project work or fail: retrieving the right chunk, filtering by tenant or document, and paying a predictable bill. Every price comes from the vendor's own pricing page. A database without public pricing (LanceDB, for example) stays out of the ranking, because we can't compare a price we can't see.
The core test has two parts. The database has to support hybrid search (keywords plus vectors) and reliable metadata filtering. All six ranked databases do, each in its own way. Then comes real cost, including the monthly floor, and whether you can self-host or pin data to a region.
Ease of use, value for money and features each weigh 25%, because getting any one of them wrong means a migration. Integrations (LangChain, LlamaIndex, n8n) weigh 15%. Support weighs 10%: you mostly run these tools from their docs. Pinecone and Supabase scores are the ones from our full reviews.
Then we read developer feedback, starting with the complaints. On Hacker News, one developer sums up a common view: "Pinecone is just generally thought as something extremely overpriced". Another writes "they recently increased the cost floor for serverless" and looks at pgvector instead. Qdrant keeps coming up as a solid production choice, with a team people describe as very responsive.
- Ease of useTime to a first working index, API clarity, docs, a free tier you can actually use.25 %
- Value for moneyMonthly floor, usage cost as traffic grows, what the free tier really allows.25 %
- Features and depthHybrid search, metadata filtering, index types, self-hosting, compliance and regions.25 %
- Customer support and assistanceSupport by plan, team responsiveness, quality of answers in community feedback.10 %
- Available integrationsLangChain, LlamaIndex, n8n, Make, official SDKs, fit with the database you already run.15 %
No affiliate link changes a score or a rank.
Vector database pricing moves fast. Get the changes in your inbox
Qdrant
Qdrant is the vector database developers recommend most often for production. On Hacker News, several teams report running it for over a year, one at 50 to 100 million rows, and call the team very responsive. It's open source, so you can run it on your own servers or use Qdrant Cloud.
The Free Tier never expires: one node, 0.5 vCPU, 1 GB of RAM and 4 GB of disk, with no high availability. That covers a prototype or a small internal assistant. Past that, Standard bills by the hour for CPU, memory and storage, and Qdrant publishes no minimum. You pay for what your cluster uses.
For retrieval, Qdrant combines dense and sparse vectors, including BM25 full-text search, then fuses the results with RRF or DBSF. You filter on the payload, the metadata attached to each point. For sensitive data, Hybrid Cloud runs clusters on your own infrastructure, managed from Qdrant Cloud. The vendor shows SOC 2 and HIPAA badges.
The catch: Qdrant Cloud doesn't list its regions publicly. If data residency is a contract requirement, ask before you sign, or self-host.
- Free Tier that never expires
- Hybrid search with RRF or DBSF fusion
- Open source, self-hostable
- ✓No published minimum on Standard
- ✓Strong production feedback
- ✓Hybrid Cloud keeps data in your infrastructure
- ✗Cloud regions not listed publicly
- ✗Free Tier has no high availability
- ✗You estimate Standard costs yourself
The default pick for a dedicated vector database. Start on the Free Tier and keep self-hosting as your exit door.
Supabase (pgvector)
Supabase is hosted Postgres, and pgvector is the extension that adds vectors to it. Supabase's own AI docs put it bluntly: "the best vector database is the database you already have". Your embeddings sit next to your users, orders and permissions, and one SQL query filters and searches.
Free costs $0: a 500 MB database per project and 2 active projects. A free project pauses after 7 days without activity, so don't use it for a client demo. Pro is $25 per month per organization, with $10 of compute credit included. Team jumps to $599 per month, and that's the first plan with SOC 2. HIPAA is a paid add-on on Team and Enterprise.
pgvector offers HNSW and IVFFlat indexes. Hybrid search means combining Postgres full-text search with vectors and fusing the results (RRF or a cross-encoder). Two limits to know. Filters used to run after the approximate index scan, so selective filters returned fewer rows than asked; iterative index scans fix this since 0.8.0. And an index tops out at 2,000 dimensions with vector, 4,000 with halfvec.
- Vectors and app data in one database
- HNSW and IVFFlat indexes
- Hybrid search with Postgres full-text
- ✓No extra database to run
- ✓Filtering in plain SQL
- ✓Affordable Pro at $25/mo
- ✗Free projects pause after 7 days
- ✗SOC 2 only from Team ($599)
- ✗2,000-dimension index limit with vector
If your app already lives in Postgres, start here. Move to a dedicated store only when filtering or scale forces you to.
Weaviate
Weaviate includes hybrid search on every plan, the free one too. One query blends keyword and vector search and returns a single ranked list. For RAG over technical docs, where product names and error codes matter as much as meaning, that saves real work.
Free costs $0: one cluster, 100,000 objects, 1 GB of RAM, 10 GB of disk and a single collection, on AWS in 2 regions. Tight, but enough to validate an idea. Flex starts at a $45 monthly minimum, then bills on usage: from $0.00465 per million vector dimensions and from $0.12 per GiB of storage. Premium starts at $400 per month on a prepaid contract.
Weaviate is open source and runs on GCP and AWS across 7 regions, though the list isn't published. It's SOC 2 Type II audited and offers HIPAA on Enterprise Cloud on AWS.
- Hybrid search included, even on Free
- Open source
- SOC 2 Type II audited
- ✓Hybrid search out of the box
- ✓Readable Flex billing
- ✓Self-hosting available
- ✗$45/mo floor once you leave Free
- ✗Free capped at 1 collection and 100,000 objects
- ✗Region list not published
The pick when hybrid search is the core requirement. Budget the $45 floor from day one of production.
Pinecone
Pinecone is the best-known managed vector database and the quickest to start: create an index, upsert vectors, query. There's no server to size. Dense, sparse and full-text indexes are available on every plan. For hybrid, you combine dense and sparse in one index (with a client-side weight) or merge two indexes. Metadata filters apply before ranking.
Starter is free: 2 GB, 2 million write units and 1 million read units per month, but only on AWS us-east-1. Builder is a flat $20 per month for 10 GB. Standard starts at a $50 monthly minimum plus usage: $16 to $18 per million read units, $4 to $4.50 for writes, $0.33 per GB per month for storage. Enterprise starts at $500 per month.
Compliance is strong: SOC 2 on every plan, GDPR and ISO 27001 from Builder, HIPAA as a $190 per month add-on on Standard. EU regions exist (AWS Frankfurt and Ireland, GCP europe-west4), just not on Starter.
The downside: the bill tracks reads. An assistant that takes off can multiply your cost, and several developers say that's why they left. It's proprietary and cloud only (or in your own cloud account, on quote).
- Nothing to operate
- Flat $20/mo Builder plan
- EU regions and GDPR from Builder
- ✓Fastest start in this ranking
- ✓Full compliance from Builder
- ✓Filters applied before ranking
- ✗Bill climbs with read traffic
- ✗Starter limited to us-east-1
- ✗Proprietary, no self-hosting
The right call if nobody on your team wants to run a database. Watch read units as soon as traffic grows.
Zilliz Cloud (Milvus)
Zilliz Cloud is managed Milvus, the open-source (Apache 2.0) database built for scale. Milvus runs three ways: Lite as a Python library, Standalone as one Docker image, and Distributed on Kubernetes for billions of vectors. It covers multi-field hybrid search, BM25 full-text, sparse vectors and scalar filtering.
Free costs $0: 5 GB, 2.5 million vCUs per month and 5 collections. Serverless starts at $0 and bills on usage. Enterprise Dedicated starts at $197 per month with a 99.95% SLA, SSO and audit logs. Zilliz states SOC 2 Type II and ISO 27001 compliance.
The weak spot is self-hosted Milvus. Distributed mode means running Kubernetes, etcd, object storage and a message queue. Reviewers mention higher resource use than Qdrant and docs that aren't beginner-friendly. Zilliz Cloud takes that work off your plate.
- Managed Milvus, open source underneath
- Multi-field hybrid and BM25
- 99.95% SLA from Enterprise Dedicated
- ✓Generous free tier (5 GB)
- ✓Scales to billions of vectors
- ✓Lite mode for local prototyping
- ✗Distributed Milvus is heavy to self-host
- ✗Serverless unit prices hard to read
- ✗Overkill for a small RAG
Pick it when scale is the problem. For a few million vectors, Qdrant or pgvector will do.
Turbopuffer
Turbopuffer stores vectors on object storage, which keeps the per-GB cost low. Usage tops out at $0.33 per GB stored, $2 per GB written and $1 per PB queried. The official example: one million documents (about 3 GB) costs around $7 in usage, so in practice you pay the $16 Launch floor.
There's no free plan. Launch is $16 per month minimum with SOC 2 and a GDPR DPA. Scale is $256 per month minimum and adds a HIPAA BAA and SSO. Enterprise starts at $4,096 per month plus 35% on usage, with BYOC, customer-managed keys and a 99.95% SLA. Full-text search and filters are included.
Regions include AWS and GCP in Frankfurt, plus Belgium, Dublin and London. The limits: it's proprietary and cloud only, and the jump from $16 to $256 for HIPAA is steep.
- Object-storage pricing
- Full-text and filters included
- Frankfurt, Belgium, Dublin and London regions
- ✓Low storage cost at volume
- ✓SOC 2 from the first plan
- ✓Clear official cost example
- ✗No free plan
- ✗Big jump to $256 for HIPAA
- ✗Proprietary, cloud only
A strong pick for multi-tenant SaaS at volume. For a single small RAG, a free tier elsewhere does the job.
What a vector database really costs per month
Pricing pages lead with the free tier. What matters is the floor you hit in production, and what drives the bill after that. Official USD prices, before tax.
For a prototype
$0 on Qdrant (1 GB RAM), Weaviate (100,000 objects), Pinecone (2 GB), Zilliz (5 GB) and Supabase (500 MB). Chroma gives $5 in credits. Watch out: Supabase pauses a free project after 7 days without activity.
Your first paid floor
Turbopuffer Launch: $16 minimum. Pinecone Builder: $20 flat. Supabase Pro: $25 per organization. Weaviate Flex: $45 minimum. Pinecone Standard: $50 minimum. Qdrant Standard and Zilliz Serverless: usage only, no published minimum.
When compliance shows up
SOC 2: included on every Pinecone plan and on Turbopuffer Launch, but you need Team on Supabase ($599) and on Chroma ($250). HIPAA: a $190/month add-on on Pinecone Standard, $256/month minimum on Turbopuffer Scale.
- Does your plan include hybrid search?
- Are filters applied before or after the vector search?
- What's the monthly floor once you leave the free tier?
- Is your required region available on that plan?
- Does the price follow reads, storage or memory?
- Can you self-host if pricing changes?
Other options worth knowing
Before adding a database, check whether yours already does vectors. These come up often, but target a different case or don't publish clear vector pricing.
Chroma: the fastest way to prototype
LocalChroma installs in one line and runs locally, which is why so many tutorials use it. Chroma Cloud's Starter is $0 plus usage with $5 in credits; Team is $250 per month with $100 of usage included and SOC 2. Usage: $2.50 per GiB written, $0.33 per GiB stored per month. Great for learning, but compare that Team floor before production.
- Price
- $0 + usage, Team $250/mo
- Best for
- Prototypes, solo developers
Neon: serverless Postgres with pgvector
PostgresNeon includes pgvector on every plan. Free is $0 (100 compute hours per project, 1 GB per project), then Launch at $0.106 and Scale at $0.222 per compute hour, with storage at $0.35 per GB-month. Neon lists ISO 27001, ISO 27701, GDPR, and SOC 2 and HIPAA on Scale.
- Price
- $0, then $0.106 per compute hour
- Best for
- Postgres teams
MongoDB Atlas: if your data is already there
Already in placeAtlas Vector Search does hybrid search (vector plus full-text) and pre-filtering on indexed fields. Atlas starts with a free M0 cluster (512 MB), but the pricing page doesn't say whether vector search is included there, so check first. Flex costs $0.011 per hour capped at $30 per month, and a dedicated M10 starts at $56.94 per month. It also works self-hosted with Community Edition.
- Price
- Flex capped at $30/mo, M10 from $56.94
- Best for
- Apps already on MongoDB
Elasticsearch: if your search already runs on it
Already in placeElastic Serverless starts at $0.14 per VCU-hour for ingest and $0.09 for search, with storage from $0.047 per GB per month. The monthly minimum isn't shown. If your search already runs on Elastic, adding vectors saves you a database.
- Price
- From $0.09 per search VCU-hour
- Best for
- Teams already on Elastic
Amazon OpenSearch Serverless: for AWS shops
AWSExpect $0.24 per OCU-hour. Classic collections need at least 2 OCUs (one indexing, one search), and vector collections don't share OCUs. NextGen collections have no minimum and scale to zero after 10 idle minutes. There's no free tier.
- Price
- $0.24 per OCU-hour
- Watch out
- 2-OCU minimum on Classic
Redis: if you already use it for caching
Already in placeRedis Cloud starts free (30 MB), then Essentials from $5 per month and Pro from $200. Its pricing page doesn't break down vector search by plan, so confirm before committing. Handy when Redis is already in your setup and your index stays small.
- Price
- $0, Essentials from $5/mo
- Best for
- Small indexes, Redis in place
LanceDB: the embedded option
LocalLanceDB embeds in your app (Python, TypeScript, Rust) with no server. It handles vector, full-text, hybrid, reranking and filters, with SQL. Enterprise has no public pricing, which keeps it out of the ranking.
- Price
- Open-source version free, Enterprise on quote
- Best for
- Local apps, data on disk
What derails a vector database project
These come up again and again in developer threads. None of them show on a pricing page.
Filters that empty your results
QualityWith pgvector before 0.8.0, filters run after the approximate index scan. You ask for 10 chunks from one customer and get 2. Turn on iterative index scans, or pick a store that filters before ranking, like Pinecone.
- Check
- Pre-filtering or post-filtering
An embedding model too big for the index
TechnicalOpenAI's text-embedding-3-large outputs 3,072 dimensions. pgvector indexes up to 2,000 with vector and 4,000 with halfvec. Switch to halfvec or shorten the vectors before loading everything.
- pgvector limit
- 2,000 dims with vector
A bill that grows with your success
PricingOn Pinecone Standard, every query burns read units at $16 to $18 per million. An assistant going from 100 to 10,000 questions a day lands in a different price bracket. Model your traffic six months out, not your demo.
- Model
- Traffic six months out
Assuming you can switch over a weekend
MigrationSwitching stores mostly means re-embedding and rebuilding indexes. Keep your source text and the exact embedding model name so you can reload anywhere.
- Keep
- Source text and model name
Forgetting where your data lives
CompliancePinecone Starter runs only in us-east-1. Qdrant, Weaviate and Chroma don't list their regions publicly. If your documents hold personal or health data, lock the region before you load anything.
- Named EU regions
- Pinecone, Turbopuffer
Get more from your vector store without paying more
A few settings that change answer quality in your RAG, whatever database you pick.
Turn on hybrid for names and codes
QualityPure vector search misses SKUs, error codes and customer names. Add BM25 full-text and fuse with RRF: Qdrant, Weaviate, Milvus and pgvector all support it.
- Fusion
- RRF, easiest to tune
Set your metadata on the first load
StructureStore customer, source, language and date with every chunk. You can then filter to one customer's contracts without reloading. Adding a field later often means touching every record.
- Useful fields
- Customer, source, language, date
Keep free tiers for development
PricingA Qdrant Free Tier or a Pinecone Starter index covers your dev work. Pay only for production, and wake your free Supabase project before each demo or it'll be paused.
- Saving
- One paid plan instead of two
Feed your store with n8n or Make
Integrationn8n vector store nodes (Qdrant, Pinecone, Supabase, Weaviate, Milvus) chunk documents, compute embeddings and load them with no code. Drop a file in Google Drive and your index follows.
- Tools
- n8n, Make
Building an AI agent on your own docs? Get our next guides
Where to go next
These pages pick up where this one leaves off, on tools and setup.
Our full review.
You're going with PostgresSupabase reviewOur full review.
You're comparing bothPinecone vs SupabaseCriterion by criterion.
You want Pinecone's pricing in detailPinecone pricingPlans and usage.
You're wiring Qdrant into n8nQdrant in n8nThe node, step by step.
You want us to build your agentAI agent agencyWe build it with you.
Your questions about vector databases
What is a vector database?
It stores embeddings, lists of numbers that capture the meaning of text or images, and finds the ones closest to a query. It's the memory behind a RAG pipeline or semantic search.
What is the best free vector database?
Qdrant for a dedicated store: its Free Tier never expires (1 GB RAM, 4 GB disk). Zilliz gives 5 GB, Pinecone 2 GB. Skip Supabase Free for demos, since it pauses after 7 idle days.
pgvector vs Pinecone: which should you use?
pgvector if your data is already in Postgres: Supabase Pro is $25 a month and there's nothing new to run. Pinecone if you want a managed store with filters applied before ranking and can live with read-based billing.
What is the best vector database for RAG?
For most teams, Qdrant or pgvector. RAG quality depends more on hybrid search and metadata filtering than on raw speed, and both handle that. Pick Weaviate if you want hybrid with zero setup.
What is the best open-source vector database?
Qdrant for most projects, then Weaviate. Milvus targets billions of vectors but its distributed mode needs Kubernetes, etcd and object storage. Chroma is the simplest locally.
When should you move off pgvector?
When selective filters still hurt recall after 0.8.0, when HNSW indexes no longer fit in memory, or when your model goes past 4,000 dimensions. Budget time to re-embed your data.
Which vector databases are HIPAA-ready?
Pinecone offers HIPAA as a $190/month add-on on Standard. Turbopuffer includes a BAA from Scale ($256/month). Weaviate offers it on Enterprise Cloud on AWS, Supabase as a paid add-on on Team, Neon on Scale.
How much does a vector database cost per month?
From $0 to start, then $16 to $50 for a first paid floor: Turbopuffer $16, Pinecone Builder $20, Supabase Pro $25, Weaviate Flex $45, Pinecone Standard $50. Qdrant Standard bills usage only.