AI Vector Databases Comparison 2026: Pinecone, Weaviate, Qdrant and 7 More (Pricing & Free Plans)

Choosing an AI vector database is harder than it looks. The tools differ in architecture, in how they charge, and in how much a free plan really covers. Some are managed services, some are open source you can self-host, and some add vector search to a database or search engine you may already run. For teams working in Japanese, region availability and tokenizer support add another layer to check.

This article compares ten tools: Pinecone, Weaviate, Qdrant, Zilliz Cloud, Vespa, MongoDB Vector Search, Elasticsearch, Redis, LanceDB, and turbopuffer. For each one, we look at pricing, free plans, Japanese support, supported platforms, strengths, and limitations. We compared them using each vendor's official website, checked in November 2026.

After the individual reviews, you will find recommendations by goal and practical tips for choosing without regrets. Prices and features change often, so check the official website for the latest pricing and features before you decide.

What Are AI Vector Databases? What They Do and What to Know Before Choosing

An AI vector database stores embeddings, which are numeric representations of text, images, or other data. It finds items that are close in meaning rather than only items that match exact keywords. Most tools in this article also store metadata, filter results by it, and combine vector search with keyword search (BM25) in one query. Some add hosted embedding models, reranking, or agent features on top.

• Semantic search: find documents by meaning instead of exact wording.
• RAG: retrieve relevant passages to give a language model before it answers.
• Recommendations: surface items similar to what a user has viewed.
• Hybrid search: merge keyword and vector results, for example with RRF.
• Agent search and memory: give AI agents a store they can query, such as Pinecone Assistant or Weaviate's Query Agent.

Before choosing, check what the free plan or trial actually covers. Pinecone's Starter plan is limited to AWS us-east-1. Zilliz Cloud's Free and Serverless options are offered only in Frankfurt and Oregon. Qdrant's free cluster is suspended after 1 week of inactivity. Vespa offers $300 in trial credits, and turbopuffer has no free plan.

Japanese support varies widely. Some tools offer a Japanese website or Tokyo regions, while others rely on Japanese tokenizers in their documentation. Pinecone and turbopuffer do not list a Tokyo region on their official websites. Also check whether the keyword side can segment Japanese text, since vector search alone depends on your embedding model.

Finally, consider scale and pricing model. Plans include flat monthly fees, usage-based billing, hourly billing for allocated resources, and minimum spends. Minimums range from $16/month for turbopuffer's launch plan to $20,000/month for Vespa Premium. Self-hosting options also differ: Weaviate, Qdrant, and Vespa publish open-source editions, while Pinecone's database cannot be self-hosted.

Comparison Table of Top 10 Tools (Pricing, Japanese Support, Platforms)

Tool	Main plans	Japanese support	Platforms
Pinecone	Builder $20/month / Standard minimum usage $50/month / Enterprise minimum usage $500/month	Partial	Web, API, Python SDK, TypeScript SDK and more
Weaviate	Free $0/month / Flex from $45/month	Partial	Web, API, Python SDK, TypeScript SDK and more
Qdrant	Free Tier free / Contact sales	Partial	Web, API, Python SDK, TypeScript SDK and more
Zilliz Cloud	Free $0 / Standard: Serverless from $0/month, Dedicated from $126/month. 30-day free trial available	Supported	Web, API, Python SDK, TypeScript SDK and more
Vespa	Free trial / Startup: vCPU $0.05/hour / Cloud Basic: vCPU $0.10/hour	Partial	Web, API, Python SDK, Java SDK and more
MongoDB Vector Search	MongoDB Atlas Free $0/hour / Flex from $0.011/hour / Dedicated from $0.08/hour	Supported	Web, API, Python SDK, TypeScript SDK and more
Elasticsearch	14-day free trial / Elasticsearch Serverless: Ingest from $0.14/VCU-hour, Search from $0.09/VCU-hour, Machine Learning from $0.07/VCU-hour	Supported	Web, API, Python SDK, TypeScript SDK and more
Redis	Essentials from $0.007/hour / Pro from $0.014/hour	Partial	Web, API, Python SDK, Go SDK and more
LanceDB	Contact sales	Partial	Web, API, Python SDK, TypeScript SDK and more
turbopuffer	Launch: minimum spend $16/month / Scale: minimum spend $256/month / Enterprise: from $4,096/month	Not stated	Web, API, Python SDK, TypeScript SDK and more

AIpedia has not rated these tools, because they do not have their own pages on AIpedia yet. Pricing and plans change; we checked each vendor's official website in November 2026. Check the latest information on the official website before you sign up.

Features, Pricing, and Best For: 10 Tools

1. Pinecone

Pinecone is a fully managed, serverless vector database. You store embedding vectors and metadata in serverless indexes without building or operating servers. It is used as the search part of semantic search, recommendations, and RAG. Document indexes, generally available since September 2, 2026, combine BM25 full-text search, dense vectors, and sparse vectors in a single index.

• Strengths: There is a gradual path from the free Starter plan to the flat-rate $20 Builder plan and then the usage-based Standard plan. Hosted embedding and reranking models and Pinecone Assistant let you build RAG without separate inference infrastructure.
• Strengths: Hybrid search can be built in one service. Pinecone supports SOC 2, GDPR, ISO 27001, and HIPAA (a HIPAA add-on is available for Standard). Enterprise adds BYOC, CMEK, audit logs, and SCIM.
• Strengths: The official documentation provides integration guides for LangChain, LlamaIndex, and Amazon Bedrock.
• Things to note: There is no Tokyo region for serverless indexes. The only Asian option is AWS Singapore (ap-southeast-1), and Starter is limited to AWS us-east-1. Japanese is not in the language list for full-text search (BM25). The hosted model llama-text-embed-v2 supports 26 languages including Japanese.
• Things to note: The database cannot be self-hosted. The local emulator Pinecone Local was discontinued on October 7, 2026, and running in your own environment is limited to Enterprise BYOC. Unit prices vary by cloud and region, and egress became billable on September 1, 2026, so estimate costs in advance.

Pricing: Starter is free (AWS us-east-1 only, 5 indexes, 2GB storage, 2 million WU/month writes and 1 million RU/month reads, 1GB/month egress, 2 users). Builder is $20/month (flat rate, no overage charges; operations are blocked when the limit is reached. 10 indexes/project, 10GB storage, 5 million WU/month writes and 2 million RU/month reads, all regions selectable, card payment only). Standard has a minimum usage of $50/month, with usage-based billing above that: storage $0.33/GB/month, writes $4 to $4.50/1 million WU, reads $16 to $18/1 million RU (varies by cloud and region), egress $0.10/GB (up to 100GB/month included), import from object storage $0.25/GB, backup $0.10/GB/month, and a HIPAA add-on at $190/month. Standard includes a 3-week, $300 credit free trial. Enterprise has a minimum usage of $500/month (writes $6 to $6.75/1 million WU, reads $24 to $27/1 million RU, 99.95% uptime SLA, BYOC, Pro support included). Inference on Standard and above: llama-text-embed-v2 $0.16/1 million tokens, multilingual-e5-large $0.08/1 million tokens, rerank $2/1,000 requests. Assistant on Standard and above: storage $3/GB/month, input $8 and output $15/1 million tokens. BYOC is contact sales.

Best for: Teams that want a managed service with no server operations and a clear step from free to paid. It suits projects that can accept no Tokyo region and will pick a multilingual embedding model for Japanese documents.

2. Weaviate

Weaviate is an open-source vector database that handles objects, vectors, and inverted indexes in one system. It includes hybrid search, which combines vector search and BM25 keyword search, as standard. You can self-host the open-source edition with Docker or Kubernetes, or use the managed Weaviate Cloud. On Weaviate Cloud you also get Weaviate Embeddings and Query Agent.

• Strengths: The BSD-3-Clause open-source edition can be self-hosted, and the same client can connect to Weaviate Cloud. Hybrid search, multi-tenancy, and RBAC are available on all plans, including the free plan.
• Strengths: Japanese tokenizers such as kagome_ja, gse, and trigram are covered in the official documentation. This makes Japanese keyword search and hybrid search easier to set up. The website also has a Japanese landing page and quickstart (weaviate.io/ja).
• Strengths: Hosted embeddings and Query Agent let you configure everything from embeddings to agent-style search without external API keys. Official clients exist for Python, TypeScript, Go, Java, and C#.
• Things to note: The list of Weaviate Cloud regions is not published, so you must check whether Tokyo is available. The free cluster is on AWS, and Flex is limited to 7 regions across GCP and AWS. Japanese tokenizers are not loaded by default and must be enabled with environment variables.
• Things to note: Pricing depends on vector dimensions, storage, and backup, and unit prices vary by compression method, cloud, and region. The official site announces that Transformation Agent and Personalization Agent are scheduled to be discontinued, and that Engram will end on December 31, 2026.

Pricing: Free is $0/month (free forever, 1 cluster per user, 100,000 objects, 1GB memory and 10GB disk, up to 1 collection and 3 tenants, Embeddings 2,000 requests/day, Query Agent 1,000 requests/month, AWS only). Flex is from $45/month (usage-based, no contract, shared cluster, 99.5% uptime SLA, vector dimensions from $0.00465/1 million, storage from $0.12/GiB, backup from $0.03036/GiB). Premium is contact sales (prepaid contract, minimum from $400/month). Premium Shared: vector dimensions from $0.003875/1 million, storage from $0.10/GiB, 99.9% uptime SLA. Premium Dedicated: vector dimensions from $0.002718/1 million, storage from $0.33/GiB, 99.95% uptime SLA. Weaviate Embeddings: Snowflake Arctic-Embed-M-v1.5 $0.025/1 million tokens, Snowflake Arctic-Embed-M-v2.0 $0.040/1 million tokens, ModernVBERT colmodernvbert $0.065/1 million tokens. Query Agent is $30/organization/month (4,000 requests included, usage-based beyond that). The open-source edition (Community Edition, BSD-3-Clause) is free. Weaviate Assurance, paid support for self-hosting, is a fixed annual fee, contact sales.

Best for: Teams that want an open-source option with a managed cloud alongside it. It is a good fit when Japanese keyword and hybrid search matter and you can verify region availability first.

3. Qdrant

Qdrant is an open-source (Apache 2.0) vector search engine. You can self-host it, use the managed Qdrant Cloud, or run Hybrid Cloud on your own Kubernetes. Private Cloud covers isolated environments. It supports dense vectors, sparse vectors, multivectors, quantization, and payload filtering for RAG, recommendations, and advanced search.

• Strengths: The Apache 2.0 edition is free to self-host, and you can choose Managed Cloud, Hybrid Cloud, or Private Cloud. A free cluster (0.5 vCPU, 1GB RAM, 4GB disk) can be tried without a credit card.
• Strengths: Search features are broad: sparse vectors, multivectors, BM25 scoring on full-text indexes (Qdrant 1.20, released October 22, 2026), and quantization. This gives fine-grained control over how hybrid search is built.
• Strengths: The official documentation states that the multilingual tokenizer for full-text indexes supports Japanese, using vaporetto. Official clients cover Python, JavaScript/TypeScript, Rust, Go, .NET, and Java.
• Things to note: Standard Tier unit prices are not on the pricing page, so you must estimate with the pricing calculator. The Qdrant Cloud region list is not published, so check whether a Tokyo region exists. There is no Japanese official website.
• Things to note: For clusters in non-EU regions, Cloud Inference runs in the US, and the free models are hosted only in the US. The free cluster is suspended after 1 week of inactivity and deleted after 4 weeks.

Pricing: Free Tier is free (free forever, no credit card required). It is a single node with 0.5 vCPU, 1GB RAM, and 4GB disk (about 1 million vectors at 768 dimensions), with free inference models available. It is suspended after 1 week of inactivity and deleted after 4 weeks. Standard Tier is usage-based with dedicated resources, billed hourly for vCPU, memory, disk, backup capacity, and tokens for paid inference models, with a 99.5% uptime SLA. Unit prices are not listed on the pricing page and are estimated with the pricing calculator, so check the official page. Premium Tier is contact sales (committed contract with an annual minimum usage, SSO, private VPC connectivity, 99.9% uptime SLA, 99.95% for multi-AZ). Hybrid Cloud and Private Cloud are contact sales. The open-source edition is free (Apache 2.0). Payment is available by credit card or through the AWS, GCP, and Azure marketplaces.

Best for: Engineering teams that want an open-source engine with flexible deployment and detailed control over hybrid search. Confirm the region and cost estimate before committing to Qdrant Cloud.

4. Zilliz Cloud

Zilliz Cloud is a fully managed service built by the team behind the open-source vector database Milvus. You can choose Serverless, dedicated clusters (Dedicated), or BYOC that runs in your own cloud account. It uses the same SDK as Milvus and supports semantic search and RAG.

• Strengths: Dedicated clusters and BYOC are available in Tokyo on AWS, Google Cloud, and Azure. There are also a Japanese official website and Japanese documentation.
• Strengths: Because it shares an SDK with open-source Milvus, moving between self-hosted Milvus and the managed version is easier. You can start on the Free plan or the 30-day Standard free trial.
• Strengths: Full-text search (BM25) and hybrid search are available. The lindera tokenizer supports Japanese full-text search, and the Function feature can call external embedding and reranking models.
• Things to note: Free and Serverless are available only in AWS eu-central-1 (Frankfurt) and GCP us-west1 (Oregon). To keep data in Tokyo, you need a dedicated cluster or above.
• Things to note: Billing has many items, including CU, vCU, storage, and data transfer, and unit prices vary by plan and region. Vector Lakebase is a public preview, and Business Critical and BYOC prices are not published.

Pricing: Free is $0 (5GB storage, 2.5 million vCU/month included, up to 5 collections, community support). Standard Serverless is from $0/month (reads and writes $4/1 million vCU, storage $0.3/GB/month; available only in AWS eu-central-1 and GCP us-west1). Standard Dedicated is from $126/month ($0.175/CU/hour in AWS us-east-1; Tokyo (AWS ap-northeast-1) is $0.209/CU/hour), with a 30-day free trial. Enterprise Dedicated is from $197/month ($0.273/CU/hour in AWS us-east-1; Tokyo is $0.327/CU/hour; 99.95% uptime SLA, SSO, audit logs, private endpoints). Business Critical and BYOC are contact sales. Dedicated cluster storage is $0.025/GB/month, with data transfer free for the first 100GB. Open-source Milvus is free (Apache 2.0). Check the official website for the latest pricing and features.

Best for: Teams that want a managed service based on Milvus and need Japanese documentation, a Tokyo region, or Japanese full-text search. It suits those who can use a dedicated cluster or above, and those who want an easy path between self-hosted Milvus and a managed service.

5. Vespa

Vespa is an open-source (Apache 2.0) search platform. It runs vector search, keyword search, structured filtering, ranking, and machine learning inference in one distributed engine. It handles tensors natively, so one document can hold multiple embeddings. It is also offered as the managed Vespa Cloud.

• Strengths: Search, ranking, and inference live in one engine, so you do not need to combine a separate vector database, full-text engine, and reranker. Multi-phase ranking can run ONNX or XGBoost models, and ColBERT- and ColPali-style search is supported.
• Strengths: Vespa Cloud has a production zone in Tokyo (AWS ap-northeast-1). Lucene Linguistics provides Japanese analyzers out of the box.
• Strengths: Billing is based on allocated resources rather than query count, and unit prices are published. The Apache 2.0 edition can be self-hosted for free.
• Things to note: The default language processing (OpenNLP) does not support Japanese, so Japanese full-text search requires switching to Lucene Linguistics. There is no Japanese official website or documentation.
• Things to note: Configuration lives in an application package with services.xml and schemas, and changes mean editing files and redeploying. Azure zones on Vespa Cloud are development only, and multi-region failover and SSO/SAML require Premium, with a minimum of $20,000/month.

Pricing: The free trial on Vespa Cloud gives $300 in credits with no credit card required. Clusters are deleted automatically when the credits run out. Startup, for early-stage companies, is vCPU $0.05/hour, memory $0.005/GB/hour, disk $0.0002/GB/hour, GPU $0.03/GB/hour (community support only, no SLA). Cloud Basic is vCPU $0.10/hour, memory $0.01/GB/hour, disk $0.0004/GB/hour, GPU $0.07/GB/hour (next-business-day response). Cloud Professional is vCPU $0.145/hour, memory $0.0145/GB/hour, disk $0.0005/GB/hour, GPU $0.10/GB/hour (24/7 support, backup and disaster recovery). Premium is a custom quote with a minimum of $20,000/month. Enclave, which runs in your own cloud account, is Basic vCPU $0.08/hour and Professional vCPU $0.12/hour, among others. Enterprise, an annual license for self-hosting, is contact sales. The open-source edition is free (Apache 2.0). Unit prices drop by up to 50% (83% for Enclave) depending on the total allocated resources.

Best for: Engineering-led teams building large search or recommendation systems that need custom ranking and inference in one engine. It is less suited to teams wanting a quick setup in Japanese.

6. MongoDB Vector Search

MongoDB Vector Search is a vector search feature built into the document database MongoDB. You store business data, metadata, and embeddings in the same document and query them with the $vectorSearch stage. This removes the need to sync data to a separate vector database. Official pages also still use the older term "Atlas Vector Search."

• Strengths: Data and vectors sit together, and pre-filtering on dates, numbers, and strings works in one query. Hybrid search with full-text search is available through $rankFusion and $scoreFusion.
• Strengths: You can start on the Free cluster or Flex with a $30 monthly cap, then move to Dedicated and dedicated Search Nodes. AWS, Google Cloud, Azure, and self-managed editions are all options.
• Strengths: There is a Japanese official website and documentation. Dedicated Search Nodes are available in AWS Tokyo and Osaka, and Japanese analyzers (lucene.kuromoji, lucene.japanese) are provided for full-text search.
• Things to note: Vector search indexes are held in memory, so large data sets require planning for the hourly cost of dedicated Search Nodes. The Free cluster allows 3 full-text and vector search indexes in total, and Flex allows 10.
• Things to note: Automated Embedding is a preview, and the official documentation does not recommend it for production. Native reranking works only on Atlas with MongoDB 8.3 or later, not on self-managed deployments.

Pricing: Vector search is included in MongoDB Atlas cluster pricing. Atlas Free is $0/hour (512MB storage, shared RAM/vCPU; up to 3 indexes in total for full-text search and vector search). Flex is from $0.011/hour (monthly cap $30, 5GB storage, up to 10 indexes). Dedicated is from $0.08/hour (M10, from $56.94/month, 10GB to 4TB storage). Dedicated Vector Search Nodes are available on M10 and above. On AWS, High CPU runs from S20 at $0.12/hour to S80 at $3.26/hour. Low CPU runs from S30 at $0.11/hour to S110 at $5.87/hour. Storage-Optimized runs from S40-S at $0.25/hour to S90-S at $3.76/hour. Automated Embedding (preview) is priced per 1 million tokens of Voyage AI models: voyage-4-lite $0.02, voyage-4 $0.06, voyage-4-large $0.12. Community Edition is free (SSPL). Enterprise Advanced (commercial license) is contact sales.

Best for: Teams already using MongoDB that want vectors beside their application data. It also suits Japan-based teams that value a Japanese website and Tokyo or Osaka Search Nodes.

7. Elasticsearch

Elasticsearch is a search and analytics platform long used for full-text search, and it now also works as a vector database. kNN search on dense_vector fields (HNSW), BM25 full-text search, metadata filters, and aggregations can run in the same index. Hybrid search can be combined in one API call with RRF and others. The semantic_text field can automate chunking and embedding generation.

• Strengths: Full-text search, vector search, filters, and aggregations are handled in one engine, with hybrid search built in. semantic_text and the Elastic Inference Service let you start semantic search without running embedding model servers.
• Strengths: You can choose Hosted, Serverless, or self-managed. Tokyo regions are available: Hosted supports AWS, GCP, and Azure in Tokyo, and Serverless supports AWS and GCP in Tokyo.
• Strengths: There is a Japanese official website, Kibana can be switched to Japanese, and the kuromoji plugin supports Japanese full-text search. RBAC and audit logs are available for security.
• Things to note: Billing differs among Hosted (resource-based), Serverless (VCU-based), and Vector Database (storage and search capacity), so cost estimates are complicated. Unit prices for Elasticsearch Vector Database could not be confirmed when this guide was written (November 7, 2026).
• Things to note: On Hosted, ELSER and E5 on ML nodes are listed as Platinum-or-higher features. EIS GPU inference, Elastic Rerank, and DiskBBQ are listed as Enterprise features. ELSER is for English, and the official documentation directs you to E5 models for multilingual semantic search.

Pricing: Elasticsearch Serverless is Ingest from $0.14/VCU-hour, Search from $0.09/VCU-hour, Machine Learning from $0.07/VCU-hour, storage from $0.047/GB/month, and Egress from $0.05/GB. All are shown as "As low as" on the pricing page. Elastic Inference Service is from $0.08/1 million tokens. Elasticsearch Vector Database (Serverless) is billed on four factors: storage, search capacity (Search Power), ingestion, and infrastructure cost. Unit prices appear only in the script-rendered Cloud Pricing Table, so check the official page. Elastic Cloud Hosted is Standard from $99/month, Gold from $114/month, Platinum from $131/month, and Enterprise from $184/month (for 120GB storage and a 2-zone configuration). A 14-day free trial is available. Self-managed is free and open (source available under your choice of AGPLv3, SSPL 1.0, or ELv2; the default distribution is ELv2). Platinum and Enterprise self-managed are contact sales.

Best for: Teams that already use Elasticsearch or need strong keyword search alongside vectors, especially for Japanese data. Plan extra time to compare deployment options and pricing.

8. Redis

Redis is an in-memory data store widely used for caching and session management. It also offers vector search through the Redis Query Engine, with FLAT, HNSW, and SVS-VAMANA indexes. KNN and vector range search can be combined with filters on text, numeric, geospatial, and tag fields. Redis 8.4 added the FT.HYBRID command for hybrid search.

• Strengths: The in-memory design enables low-latency vector search. You can run it on the same foundation as existing Redis uses such as caching and session management.
• Strengths: Redis Cloud can start on the free plan (30MB) or Essentials from $5/month. From Redis 8 onward, Redis Open Source includes vector search by default, and AGPLv3 is one of the license choices.
• Strengths: AI components include RedisVL, LangCache, and FT.HYBRID. Redis Cloud is available in the Tokyo regions of AWS and Google Cloud.
• Things to note: Because it is in-memory, you must secure memory capacity that matches your data volume. Redis Flex, which uses both RAM and SSD, is available for large capacities.
• Things to note: The official website has no Japanese version, and Japanese is not among the stemming languages for full-text search. Pro starts at a minimum of $200/month, and LangCache is a public preview on Redis Cloud, so test Japanese keyword search in advance.

Pricing: Redis Cloud Free is $0 (30MB, single DB, best-effort SLA, community support). Essentials is from $0.007/hour (from $5/month, 250MB to 100GB, single DB, shared). Redis Flex, which uses both RAM and SSD, can be selected. Pro is from $0.014/hour (minimum $200/month, with "First $200 free" shown on the pricing page). Pro offers dedicated resources, multiple DBs, Active-Active, private connectivity, and up to 99.999% uptime. Redis Software, the commercial on-premises edition, offers a free trial and sales inquiries, so check the official page for pricing. Redis Open Source is free from Redis 8, with your choice of RSALv2, SSPLv1, or AGPLv3.

Best for: Teams that already run Redis and want to add low-latency vector search to the same stack. It is less ideal if Japanese keyword search is central, unless testing shows it works for you.

9. LanceDB

LanceDB is a vector database built on Lance, an open-source columnar format. Its core, LanceDB Embedded, is a library that runs inside your application's process. It can search tables on local disk or object storage such as S3 without a server. For large-scale use, LanceDB Enterprise (Managed or BYOC) is offered.

• Strengths: LanceDB Embedded is open source (Apache License 2.0) and free. You can start local development or small-scale RAG without setting up a server.
• Strengths: Vector search, full-text search, hybrid search, and SQL filters run on the same table. Images, video, and other multimodal data can be managed in that table too.
• Strengths: Specifying lindera/ipadic as the full-text tokenizer lets you segment Japanese text. Enterprise supports SOC 2 Type II, HIPAA, and GDPR.
• Things to note: Enterprise pricing is not public, and the pricing page redirects to a contact page. In the embedded version, index creation and compaction run on the application side, while distributed search and automatic maintenance are Enterprise-only.
• Things to note: A Japanese official website or documentation could not be confirmed, and no Japan region is stated. SDKs cover Python, TypeScript, and Rust, and model-based rerankers in TypeScript and Rust require your own implementation.

Pricing: LanceDB Embedded (open source, Apache License 2.0) is free. LanceDB Enterprise comes in three forms: Managed, BYOC, and custom. It supports AWS, GCP, and Azure, and pricing is contact sales. The pricing page (/pricing) on the official website redirects to a contact page, and there is no public price list. The former serverless "LanceDB Cloud" is not mentioned in the current official documentation, so check the official website for its availability.

Best for: Developers who want an embedded, serverless-style library for local development, small RAG projects, or multimodal data. Larger production teams should contact the company about Enterprise and deployment location.

10. turbopuffer

turbopuffer is a search database for vector search and BM25 full-text search. It uses object storage such as S3 as its only storage, accelerated by memory and NVMe SSD caches. It is a usage-based SaaS with a minimum spend, and single-tenant and BYOC setups are available for enterprises.

• Strengths: The object-storage design is meant to store and search large amounts of data at low cost. Vector, BM25, hybrid, and regular expression search are available through a single API.
• Strengths: Native embedding (GA in August 2026) handles vectorization at write and query time when you specify a model in the schema. According to the FAQ, database features are the same across plans, except a few such as CMEK.
• Strengths: Clients exist for Python, TypeScript, Go, Java, C#, Ruby, and Rust (Rust in beta since October 2026).
• Things to note: There is no free plan, and a minimum spend applies from $16/month. Unused amounts do not carry over. Write latency is relatively high (p90=248ms for a 512KB upsert), and queries on uncached data are slower (p90=1214ms for 1 million records).
• Things to note: No Japan region is among the public regions, and no Japanese morphological analysis tokenizer is mentioned. Azure is a private preview in eastus2 only, and native reranking is in private beta.

Pricing: Launch has a minimum spend of $16/month. Scale has a minimum spend of $256/month (HIPAA-compliant BAA, SSO, audit logs, IP allowlist, etc.). Enterprise is from $4,096/month, with a 35% markup on usage-based charges (single-tenant, BYOC, private networking, namespace-level CMEK, 24-hour support, 99.95% uptime SLA). Usage-based rates shown by the calculator are storage ≤$0.33/GB, writes ≤$2.00/GB, and queries ≤$1.00/PB. Queries are calculated using the namespace size or 1.28GB, whichever is larger, with a marginal discount of 80% for the 32–128GB portion and 96% above 128GB. Returned data is charged separately at a flat per-GB rate. Audit log stream is +$128/month. Native embedding and reranking are priced per 1 million tokens by model, for example voyage/voyage-4 at $0.06. There is no free plan.

Best for: Teams with large data volumes who value low storage cost and can tolerate higher latency on uncached queries. Japan-based users should discuss a dedicated cluster or BYOC with the company first.

Recommendations by Goal

Goal	Recommended	Why
Start with a free plan	Qdrant, Weaviate	Qdrant's Free Tier needs no credit card, and Weaviate's Free plan is free forever for up to 100,000 objects.
Japanese website, documentation, and Tokyo region	Zilliz Cloud, MongoDB Vector Search	Both have a Japanese official website and documentation, with Tokyo options on dedicated clusters or Search Nodes.
Japanese keyword and hybrid search	Elasticsearch, Weaviate	Elasticsearch offers the kuromoji plugin, and Weaviate documents the kagome_ja, gse, and trigram tokenizers.
Add vector search to an existing system	MongoDB Vector Search, Redis	MongoDB stores vectors in the same document as business data, and Redis runs on the same foundation as caching.
Fully managed with no server operations	Pinecone	It is a serverless service with a free Starter plan and a flat $20/month Builder plan.
Self-hosting an open-source engine	Qdrant, Weaviate	Qdrant is Apache 2.0 and Weaviate is BSD-3-Clause, and both also offer managed cloud options.
Enterprise-scale search with custom ranking	Vespa, Pinecone	Vespa runs search, ranking, and inference in one engine, while Pinecone Enterprise offers a 99.95% uptime SLA and BYOC.
Embedded use or low-cost storage of large data	LanceDB, turbopuffer	LanceDB Embedded is free and runs in-process, while turbopuffer uses object storage as its only storage.

These picks reflect the data we reviewed, not a ranking. Your own data, region needs, and budget should drive the final choice, so test with a small proof of concept.

How to Choose Without Regrets

Check what the free plan really covers

Free plans differ in size, region, and lifespan. Pinecone's Starter is limited to AWS us-east-1, and Zilliz Cloud's Free and Serverless are available only in Frankfurt and Oregon. Qdrant's free cluster is suspended after 1 week of inactivity.

Vespa offers $300 in credits instead, and turbopuffer has no free plan at all. Make sure the free tier matches the region and data size you plan to test.

Match the pricing model to your workload

Pricing models vary a lot. Pinecone Builder is a flat $20/month, while Weaviate Flex bills by vector dimensions, storage, and backup. Vespa bills hourly for allocated resources rather than queries, and turbopuffer applies a minimum spend with usage on top.

Unit prices often vary by cloud, region, or compression method. Estimate costs with your own data volume before choosing. For Qdrant Standard and Elasticsearch Vector Database, unit prices could not be confirmed on the official pages, so check them before you decide.

Test Japanese handling on both sides

Japanese support has two parts: the platform and the search quality. Platform support covers a Japanese website, documentation, and Tokyo regions. Zilliz Cloud and MongoDB Vector Search offer these, and Elasticsearch has a Japanese website and Tokyo regions.

Search quality depends on tokenizers and embedding models. Weaviate, Qdrant, Vespa, and LanceDB document Japanese tokenization options. Pinecone and Redis do not list Japanese for full-text search, so test keyword search with your own documents.

Decide between a new database and your existing stack

If you already run MongoDB, Redis, or Elasticsearch, the built-in vector features may avoid a second system. MongoDB keeps vectors in the same document as business data. Redis is in-memory, so memory capacity must match data volume.

A dedicated tool such as Pinecone, Qdrant, or Zilliz Cloud may fit better when vector search is the main workload. Weigh the extra system against the benefit of focused features.

Consider self-hosting, scale, and support needs

Weaviate, Qdrant, and Vespa offer open-source editions you can run yourself. Pinecone's database cannot be self-hosted, and BYOC is limited to Enterprise. At larger scale, check the minimums: Pinecone Enterprise is $500/month, turbopuffer enterprise starts at $4,096/month, and Vespa Premium starts at $20,000/month.

Also check SLAs, compliance, and support levels on each plan. Confirm the final decision with a qualified professional if regulation applies to your data.

FAQ

Q. Which vector databases have a free plan?

A. Pinecone has a free Starter plan, and Weaviate Cloud has a Free plan that is free forever. Qdrant has a free-forever Free Tier, and Zilliz Cloud has a Free plan with 5GB storage and 2.5 million vCU/month. MongoDB Atlas Free is $0 (512MB), and Redis Cloud Free is $0 (30MB). turbopuffer has no free plan.

Q. Which tools are easiest to use in Japanese?

A. Zilliz Cloud and MongoDB Vector Search have a Japanese official website and Japanese documentation. Elasticsearch has a Japanese website, a Japanese Kibana UI, and the kuromoji plugin. Weaviate, Qdrant, Vespa, and LanceDB document ways to handle Japanese text. Pinecone, Redis, and turbopuffer have limits on Japanese support, so test them first.

Q. How do the pricing models differ?

A. Pinecone has a flat Builder plan at $20/month and usage-based Standard with a minimum usage of $50/month. Weaviate Flex starts from $45/month and is usage-based. Vespa Cloud Basic is vCPU $0.10/hour, billed on allocated resources. turbopuffer's launch plan has a minimum spend of $16/month.

Q. Can I run these tools on my own servers?

A. Weaviate (BSD-3-Clause), Qdrant (Apache 2.0), and Vespa (Apache 2.0) have open-source editions you can self-host. MongoDB has a free Community Edition (SSPL), and Redis Open Source is free from Redis 8. Pinecone's database cannot be self-hosted, and running it in your own environment is limited to Enterprise BYOC.

Q. What does enterprise-level use cost?

A. Pinecone Enterprise has a minimum usage of $500/month. Zilliz Cloud Enterprise starts at $197/month for dedicated clusters, and turbopuffer enterprise starts at $4,096/month. Vespa Premium is a custom quote with a minimum of $20,000/month. Prices for Qdrant Premium and LanceDB Enterprise are contact sales. Check the official website for the latest pricing and features.

Summary

The ten tools differ most in deployment, pricing model, and Japanese support. Some are managed services, some are open source, and some add vector search to an existing database or search engine. Free plans vary in size and region, and several tools publish only partial pricing.

For Japanese use, check both platform support and search quality. Tokyo regions, a Japanese website, and Japanese tokenizers are not available everywhere. Run a small test with your own data and confirm costs before committing.

• Fully managed with a free start → Pinecone (free Starter plan and flat $20/month Builder plan)
• Japanese website, documentation, and Tokyo region → Zilliz Cloud (dedicated clusters available in Tokyo)
• Vectors beside existing application data → MongoDB Vector Search (same document as business data)

View all AI code assistance tools (18 items)