GigaOm Radar for Vector Databases v3

Vector database platforms provide search and information retrieval for generative AI applications, as well as semantic search in non-AI contexts. They serve as backend systems providing the storage, indexing, computing, and search of data optimized for AI. By encoding all types of data as vectors, tensors, matrices, scalars, and other numeric forms, vector database platforms allow individual users, agents, and AI applications to retrieve relevant audio, video, image, text, and metadata in a single query. For AI applications, this is ideal.

Typically equipped with natural language interfaces, these engines make all of an organization’s data available for AI, supporting semantic search, similarity search, natural language Q&A, retrieval augmented generation (RAG), anomaly detection, and other AI use cases. Organizations can profit from the advances in large language model (LLM) technology to generate and analyze content while using vector database platforms to contextualize models’ outputs with approved, vetted, and relevant enterprise data.

Vector database platforms support other types of search that are complementary to semantic and vector similarity search, including lexical and keyword search. Hybrids of semantic/similarity and lexical search, using a combination of sparse and dense vectors, are also possible. The foundational nature of search to organizational processes, workflows, and applications makes vector database systems pertinent across many industries and use cases.

Moreover, LLM deployments are no longer restricted to textual applications but also include analysis and summaries of charts, images, business intelligence (BI) dashboards, and data patterns of tremendous scale for real-time recommendation systems and more. Contextualizing this information with the organization’s proprietary data stored in a vector database is a best practice for a holistic understanding of these additional data sources, so the models can produce analysis that’s truly meaningful for specific use cases.

This report evaluates open source and commercial vector database platforms, occasionally considering combinations of multiple such databases offered by the same vendor. It details both purpose-built vector database platforms and more general databases that have recently added vector embedding, storage, and search capabilities.

This is our third year evaluating the vector database space, previously as Sonar reports, and our first year doing so in Key Criteria and Radar reports. This report builds on that analysise and considers how the market has evolved over the last year.

This GigaOm Radar report examines 17 of the top vector database solutions and compares offerings against the capabilities (table stakes, key features, and emerging features) and nonfunctional requirements (business criteria) outlined in the companion Key Criteria report. Together, these reports provide an overview of the market, identify leading vector database offerings, and help decision-makers evaluate these solutions so they can make a more informed investment decision.

GIGAOM KEY CRITERIA AND RADAR REPORTS

The GigaOm Key Criteria report provides a detailed decision framework for IT and executive leadership assessing enterprise technologies. Each report defines relevant functional and nonfunctional aspects of solutions in a sector. The Key Criteria report informs the GigaOm Radar report, which provides a forward-looking assessment of vendor solutions in the sector.

To help prospective customers find the best fit for their use case and business requirements, we assess how well vector database solutions are designed to serve specific target markets and deployment models (Table 1).

For this report, we recognize the following market segments:

• Small-to-medium business (SMB): In this category, we assess solutions on their ability to meet the needs of organizations ranging from small businesses to medium-sized companies. Also assessed are departmental use cases in large enterprises for which ease of use and deployment are more important than extensive management functionality, data mobility, and feature set.

• Large enterprise: Here, offerings are assessed on their ability to support large and business-critical projects. Optimal solutions in this category have a strong focus on flexibility, performance, data services, and features to improve security and data protection. Scalability is another big differentiator, as is the ability to deploy the same service in different environments.

• Specialized: Optimal solutions are designed for specific workloads and use cases, such as big data analytics and high-performance computing (HPC). 

In addition, we recognize the following deployment models:

• SaaS: This deployment model is a fully managed option hosted by the vendor. Subscribers to the service pay only for what they use. Vendors are tasked with the significant overhead costs of maintaining resources necessary for organizations to use it, and other managerial concerns such as updates. One of the advantages of this deployment model is that large IT teams are not needed to implement it. 

• Self managed: The self-managed option places the burden of managing the deployment on the organizations accessing it through the cloud. Those organizations are responsible for provisioning resources for it and managing compute, storage, memory, and other concerns. One of the benefits of this method is that organizations have much more control over the service than they do with the SaaS paradigm, which can help with security and regulatory compliance concerns. Typically, IT personnel are required to make this model effective.

• Public cloud image: With this deployment model, a third-party cloud provider (such as AWS, GCP, or Microsoft Azure) hosts the service. Organizations typically access the cloud solution via containers or, in some cases, virtual machines, and can leverage the cloud provider’s underlying compute, memory, storage, and scalability resources. 

Table 1. Vendor Positioning: Target Market and Deployment Model

Table 1 components are evaluated in a binary yes/no manner and do not factor into a vendor’s designation as a Leader, Challenger, or Entrant on the Radar chart (Figure 1). 

“Target market” reflects which use cases each solution is recommended for, not simply whether that group can use it. For example, if an SMB could use a solution but doing so would be cost-prohibitive, that solution would be rated “no” for SMBs.

All solutions included in this Radar report meet the following table stakes—capabilities widely adopted and well implemented in the sector:

• Dense vector similarity search

• Sparse vector search

• Basic hybrid search

• Metadata filtering

• Basic embedding capabilities

Tables 2, 3, and 4 summarize how each vendor in this research performs in the areas we consider differentiating and critical in this sector. The objective is to give the reader a snapshot of the technical capabilities of available solutions, define the perimeter of the relevant market space, and gauge the potential impact on the business.

• Key features differentiate solutions, highlighting the primary criteria to be considered when evaluating a vector database solution.

• Emerging features show how well each vendor implements capabilities that are not yet mainstream but are expected to become more widespread and compelling within the next 12 to 18 months. 

• Business criteria provide insight into the nonfunctional requirements that factor into a purchase decision and determine a solution’s impact on an organization.

These decision criteria are summarized below. More detailed descriptions can be found in the corresponding report, “GigaOm Key Criteria for Evaluating Vector Database Solutions.”

Key Features

• Multimodality support: This metric evaluates how accomplished AI computing systems are at enabling organizations to search through, and profit from, their multimodal data. It’s critical to the overall utility of these search engines, since their capital value proposition is the ability to extend search to all types of content.

• Results optimization: This criterion examines how effectively solutions optimize the results of their information retrieval processes for organizations. Optimized results present the best answers from preferred sources, allowing organizations to minimize the number of searches while increasing the value of their outputs.

• Embedding flexibility: This metric assesses how useful vector computing systems are for concerns related to embedding content. These embeddings are the foundation of vector similarity search, without which this information retrieval paradigm would not work.

• Indexing: This criterion assesses the indexing utility that vector similarity search engines provide. A range of indexes is necessary to provide a strong query performance experience for users.

• Search variety: This metric evaluates the breadth and depth of the search mechanisms vector computing platforms have. Solutions with expanded capabilities in this area can service more use cases than those with just modest capabilities can.

• Complex data structure support: This criterion evaluates how well a solution can store and process complex data structures. Doing so extends the utility of the AI computational platform beyond vectors to include additional structures for increased performance for advanced computational use cases.

• Generative feedback loop: This construct allows the results of searches to be input into vector search engines, which enables them to add to the enterprise information they contain. This approach improves the accuracy of future results by allowing the system to learn which current ones provided the best yield for an organization.

Table 2. Key Features Comparison 

Emerging Features

• Advanced access controls: As adoption of vector search engines increases, we envision these solutions incorporating attribute-based access control (ABAC), which is much easier to scale and doesn’t require devising new roles every time something changes in an organization, like the location where a user is requesting access from.

• Quantum computing functionality: Hybrid approaches combining classic and quantum computing models have long been projected to considerably impact the speed and efficacy of vector computing platforms. In some instances, that hybridization entails using classic computing methods to perform traditional vector embeddings, while the retrieval phase is performed with quantum hardware.  

Table 3. Emerging Features Comparison 

Business Criteria

• Security: This criterion assesses the worthiness of the security measures used to fortify a particular vector computing engine and its data. Doing so is imperative for preventing data breaches, data loss, and regulatory complications arising from these issues.

• Scalability: This feature examines how well solutions in this space are able to scale up and down. Doing so is paramount for expanding workloads and applications supported by these retrieval systems to enterprise levels.

• Ease of use: This business feature describes the relative amount of effort required to manipulate vector computing platforms. The easier it is to use these platforms, the greater their adoption rates and ROIs become.

• Customization: This feature details the degree to which organizations can customize vector retrieval systems to fit the unique needs of their use cases. Customizing solutions is necessary to maximize the yield they deliver while increasing the competitive advantage derived from them.

• Cost reductions: This metric evaluates the ability of vector databases to reduce customer costs. The ability to pare down costs is integral to the long-term deployment of these systems.

Table 4. Business Criteria Comparison

The GigaOm Radar plots vendor solutions across a series of concentric rings with those set closer to the center judged as having the most complete solutions. The chart characterizes each vendor on two axes—balancing Maturity versus Innovation and Feature Play versus Platform Play—while providing an arrowhead that projects each solution’s expected evolution over the coming 12 to 18 months.

Figure 1. GigaOm Radar for Vector Databases

As you can see in Figure 1, the vendors are nearly evenly split between the Platform Play and the Feature Play sides of the Radar. This fact indicates that vector databases are no longer a niche technology, with only pure play platforms offering its semantic search capabilities. Moreover, many of the vendors in the Feature Play half of the Radar provide more general-purpose database and data management solutions. This is a result of the expansion of vector computing retrieval functionality into offerings with broader database capabilities, hinting at the popularity and the efficacy of vector-based similarity search.

However, it’s equally notable that the majority of the Leaders, as well as the Outperformers, are pure Platform Plays that specialize in vector database functionality. This reality highlights the effectiveness of the advanced functionality they offer, which is lacking in many general-purpose databases that have adopted vector search capabilities. In particular, support for complex data structures and some cost-saving measures pertaining to compression and quantization are responsible for their positioning in the Leaders circle, and for the Outperformer status of a few of the Platform Play vendors.

Yet, as the entry of more than one general-purpose database vendor into the Leaders circle of the Radar implies, the gap between the two types of vendors is closing. 

In reviewing solutions, it’s important to keep in mind that there are no universal “best” or “worst” offerings; every solution has aspects that might make it a better or worse fit for specific customer requirements. Prospective customers should consider their current and future needs when comparing solutions and vendor roadmaps.

INSIDE THE GIGAOM RADAR

To create the GigaOm Radar graphic, key features, emerging features, and business criteria are scored and weighted. Key features and business criteria receive the highest weighting and have the most impact on vendor positioning on the Radar graphic. Emerging features receive a lower weighting and have a lower impact on vendor positioning on the Radar graphic. The resulting chart is a forward-looking perspective on all the vendors in this report, based on their products’ technical capabilities and roadmaps.

Note that the Radar is technology-focused, and business considerations such as vendor market share, customer share, spend, recency or longevity in the market, and so on are not considered in our evaluations. As such, these factors do not impact scoring and positioning on the Radar graphic.

For more information, please visit our Methodology.

Activeloop: Deep Lake

Solution Overview
Activeloop’s Deep Lake provides a serverless vector computational engine predicated on a columnar storage architecture. Deep Lake provides data ingestion via Python SDK, a CLI, and connectors. The platform offloads data to object storage, enabling organizations to search on hyperscalers’ infrastructure (like S3) or on-premises object storage options (like MinIO) via integrations with these providers.

In addition to information retrieval, the platform is designed to support the data science staples of training and fine-tuning machine learning models, often on the same datasets employed for search. Other data science capabilities are enabled by integrations with PyTorch and TensorFlow, which are helpful for real-time model training predicated on streaming data. The solution also integrates with other generative AI and data science frameworks, including MMDetection, LlamaIndex, Weights & Biases, and LangChain.

Deep Lake exposes search capabilities via Python SDK and REST endpoints. Outside of the coding realm, a Chat-to-TQL (Tensor Query Language, which functions much like SQL) feature provides its own natural language querying (NLQ) interface. In addition to delivering various forms of quantization to reduce costs, Deep Lake also provides RBAC capabilities and integrates with Auth0 to further fortify security. Deep Lake supplies a wide range of data visualizations in addition to version control and streaming data functionality.

Activeloop is positioned as a Challenger and Fast Mover in the Innovation/Platform Play quadrant of the vector databases Radar chart.

Strengths
Activeloop scored well on a number of decision criteria, including:

• Generative feedback loop: Deep Lake’s generative feedback capabilities are some of the strongest available for vector retrieval systems. Deep Memory, which has a suite of optimization capabilities for vector search, was devised to learn from successful searches and prior queries to boost retrieval accuracy for future ones. Deep Memory relies on a shared embedding space that leverages commonalities between query vectors and data vectors. By analyzing the results of searches, it effectively learns to return more accurate results, establishing an intelligent feedback loop. Deep Memory uses a similar approach for increasing the efficacy of the system’s indexing.

• Embedding flexibility: Organizations can choose from a variety of options to generate vector embeddings in Deep Lake. The most readily accessible is the use of the system’s own agent, which can embed content upon ingestion. With this option, users select an embedding model from those the platform furnishes or supports, which includes smaller models from open source offerings, as well as larger ones from OpenAI and Cohere. Visual, drag-and-drop constructs can also be employed to render embeddings. Additionally, organizations can generate embeddings via third-party API calls and user-defined functions (UDFs).

• Multimodality support: Deep Lake natively supports the tensor format, which naturally lends itself to multimodality use cases involving audio, textual, image, structured, and video data, as well as the accompanying metadata. Features include the ability to search multiple vector fields across cloud environments in the same query. Organizations can also run different computations in the same query to contrast the results.  

Opportunities
Activeloop has room for improvement in a few decision criteria, including:

• Results optimization: Adding to the platform’s roster of reranking models would increase the accuracy of its results. Additionally, implementing a multiphase reranking procedure would help to refine the fidelity of information retrieval results. 

• Advanced access controls: Adding native capabilities for ABAC would considerably enhance the security posture Activeloop affords its customers.  

Purchase Considerations
Deep Lake is available via a free tier offering in which storage and input tokens are limited to 100 MB. Output tokens are limited to three queries per day. There’s also a $40-per-month Pro tier that offers 10 GB of storage, with additional storage available at $0.99 per GB. This option includes 5 million input tokens and 1.67 million output tokens. There’s also an enterprise custom tier for which pricing is negotiated. 

Use Cases
Deep Lake’s data lake qualities make it an excellent resource for data scientists’ work and other applications involving high-accuracy analytics. Examples include those pertaining to drug discovery and life sciences, and digital twin use cases and workflows for computer vision.  

AWS: Amazon Kendra and Amazon Bedrock Knowledge Bases

Solution Overview
Amazon Kendra is an ML-powered, intelligent enterprise search service that enables users to refine search results based on factors such as user behavior, content attributes, and freshness. It contains valuable search features such as autocomplete, which automatically finishes a user’s query or prompt. The system is designed to expedite and simplify the steps required for RAG. There’s a Kendra Retriever API, which relies on a semantic ranking model to locate and retrieve content relevant to a user’s query. These chunks and the query are sent to an LLM, which devises a response based upon them. In addition to facilitating the chunking step for organizations, the Kendra Retriever API boosts content predicated on metadata, dates, and underlying repositories.

Amazon Bedrock Knowledge Bases is another RAG-oriented service from AWS. The vector search engine for Amazon Bedrock Knowledge Bases can interface with Amazon Kendra, Pinecone, Amazon Aurora, MongoDB, Amazon OpenSearch Serverless, Redis Enterprise Cloud, or Amazon Neptune Analytics. Amazon OpenSearch Service is AWS's recommended vector database solution for Amazon Bedrock. In July, Amazon Bedrock Knowledge Bases boosted its options for vector databases by adding support for Amazon OpenSearch Service via managed clusters. In addition to batch ingestion, Amazon Bedrock Knowledge Bases enables programmatic document ingestion for streaming data. In this scenario, the streaming data is chunked, with embeddings for each chunk generated and stored in the selected vector storage platform. Supported sources include Salesforce, S3 buckets, Confluent, and SharePoint. 

Additionally, Amazon Aurora PostgreSQL and Amazon RDS for PostgreSQL support the pgvector extension. Vector search for Amazon MemoryDB also supports storing vectors, while Amazon DocumentDB (with MongoDB compatibility) supports vector search as well. Finally, Amazon S3 Vectors for object storage has vector search capabilities too. 

AWS is positioned as a Challenger and Fast Mover in the Innovation/Feature Play quadrant of the vector databases Radar chart.

Strengths
AWS scored well on a number of decision criteria, including:

• Results optimization: In Amazon Bedrock Knowledge Bases, organizations can tailor the results of queries with filters generated from foundation models and user-specified filters based on metadata. Access control list (ACL) filtering and relevance tuning are enabled by the Kendra Retriever API. Organizations can access rerankers through Amazon Bedrock via the Amazon Bedrock API, which conveys the query, supporting documents, and pertinent configurations to the indicated model. The system also contains RetrieveAndGenerate and Retrieve operations, which use a reranker to override the default rankings determined by the system. Thus, organizations can employ rerankers directly or during the results retrieval process for the knowledge base query. Amazon Bedrock Knowledge Bases enhances reranking for dense vectors with rerankers such as Amazon Rerank 1.0 and Cohere Rerank 3.5. Amazon Kendra Intelligent Ranking relies on semantic search mechanisms to rerank the results of queries. Predominantly used for lexical searches, this feature accounts for information contained in the underlying documents as well as the business context of the search query. 

• Embedding flexibility: Amazon Bedrock is known for its large selection of LLMs, including those for producing vector embeddings, and the full gamut of these models is available to Amazon Bedrock Knowledge Bases. Users can choose from hundreds of foundation models, many of which are available via integrations to other solutions reviewed in this report. The embedding process can also be handled by customized models from external sources. Organizations can switch embedding models for their applications without needing to write any code. Bedrock supplies an evaluation tool that allows organizations to contrast and take into account facets of accuracy, performance, and cost for models. 

• Multimodality support: Several of the models available for use in Amazon Bedrock Knowledge Bases support multimodality. Most embedding models supported by Amazon Bedrock use floating-point vectors by default. However, some models require binary vectors. If you choose a binary embedding model, you must also choose a model and vector store that supports binary vectors. If you choose Amazon Neptune Analytics as a vector store, Amazon Bedrock Knowledge Bases automatically creates embeddings and graphs that link related content across your data sources. Bedrock Knowledge Bases leverages these content relationships with GraphRAG to improve the accuracy of retrieval, enabling more comprehensive, relevant, and explainable responses to end users.

Opportunities
AWS has room for improvement in a few decision criteria, including:

• Indexing: Broadening the array of built-in indexes available to users, to include options like approximations and precise matches for both dense and sparse vectors, would make this offering more competitive. 

• Complex data structure support: Increasing the level of support for multidimensional tensors found in Amazon’s vector computing platforms would help to narrow the gap between it and some of the specialist vendors in this space. 

Purchase Considerations
Users can get a 30-day free trial of Amazon Kendra that includes 750 hours of service usage. After the trial period, pricing is predicated on the number of connectors, storage units, and query units used. The vendor supplies discounted pricing for high-volume workloads. 

Use Cases
Amazon’s vector database options are suitable for implementing RAG, empowering dynamic agents with generative models for natural language question answering and summarization, and applications of agentic AI that don’t pertain to question answering, like accessing tools to perform tasks. 

Chroma: Chroma Cloud

Solution Overview
Chroma is an open source AI retrieval system, around which Chroma Cloud offers a managed service. Chroma allows organizations to store vector embeddings and their metadata with multimodal capabilities spanning image and text. Using a Rust-based execution engine, Chroma supplies vector search, full-text search, full-text indexes, and document storage.

The engine runs as a server and can be used from Jupyter notebooks or client SDKs for JavaScript/TypeScript and Python. Chroma Cloud, however, is a SOC 2 Type I-compliant serverless offering with an object storage-based persistence layer. Organizations can manage collections of embeddings with Chroma’s collection primitive. Collections are the basic storage unit and means of querying the data in Chroma. 

Chroma is positioned as a Challenger and Forward Mover alone in the Maturity/Platform Play quadrant of the vector databases Radar chart.

Strengths
Chroma scored well on a number of decision criteria, including:

• Embedding flexibility: The database supplies lightweight wrappers for a variety of embedding model providers. Organizations can administer embedding functions as part of the creation of a Chroma collection. With this option, the function is used automatically. Alternatively, users can directly call the embedding providers themselves. Examples of providers for which Chroma offers wrappers include Google Gemini, OpenAI, Cohere, Hugging Face, and Hugging Face Embedding Server. Additionally, Chroma ships with the OpenCLIP embedding function, which organizations can employ for text and image embeddings. Currently, only the Python SDK supports multimodal embeddings, and even in that scenario, it is limited to images and text. Users can choose to query in the context of one or both of these modalities when they’re implemented in Chroma collections.

• Search variety: In addition to full-text search, which is case-sensitive in Chroma, users can perform regular expression pattern matching, using both $regex and $not_regex operators. Operators for full-text search include $contains, $not_contains, $and, and $or. It’s possible to layer on additional filtering based on metadata; metadata filters necessitate using a “where filter dictionary” for the query. Other search constructs include document search.

Opportunities
Chroma has room for improvement in a few decision criteria, including:

• Generative feedback loop: Providing mechanisms to insert the results of prior searches into the engine to inform the results of future searches could further improve the overall quality of Chroma’s query results. 

• Multimodality support: Chroma provides multimodal support. The vendor can become even stronger in this area by expanding its capabilities to include forms of data beyond text and images, such as audio data. 

Chroma was classified as a Forward Mover due to some of its limitations for indexing and complex data support.

Purchase Considerations
Chroma Cloud is accessible through a usage-based pricing model, which begins with a zero charge per month and five dollars in free credits. The vendor charges $2.50 for each gibibyte (GiB) written, $0.33 for each GiB per month for storage, and $0.0075 for each tebibyte (TiB) queried, with an additional cost of $0.09 for each GiB returned.  

Use Cases
Chroma is a good choice for workloads dependent on vector search, as well as those pertaining to full-text search, metadata search, and regular expression (regex)-based search. 

Google: Vertex AI Vector Search* 

Solution Overview
Vertex AI Vector Search is part of a broad assortment of tools and capabilities found within the Vertex AI platform, a fully managed offering for developing and deploying generative AI models. The platform supports creating agents courtesy of Agent Builder, a module containing a no-code console for constructing, orchestrating, customizing, and grounding the actions of AI agents. Other notable modules include Vertex AI Studio, an environment for designing, training, and implementing machine learning models, and Vertex AI Training and Vertex AI Prediction, which focus on the namesake areas of the data science lifecycle.

Vertex AI Notebooks integrates with BigQuery and provides a variety of data science workbench options. There are over 200 foundation models available in Vertex AI, including Google’s own Gemini, Veo, Imagen, and Chirp. A model garden feature provides access to open models like Llama 3.2 and Gemma. Vector Search was designed for users to supply the information retrieval necessary for generative AI applications, including recommendations and agent-based workflows. 

Google is positioned as a Challenger and Fast Mover in the Innovation/Feature Play quadrant of the vector databases Radar chart.

Strengths
Google scored well on a number of decision criteria, including:

• Search variety: Vertex AI Vector Search supplies semantic search with dense vectors, lexical search with sparse vectors, and a hybrid option combining these two techniques. The system also supports key features like out-of-the-box suggestions for searches, generative model-based summarization, and spell correction. Other search modalities include brute force search, conversational search, and agent-based search, in which the engine powers agents that can perform information retrieval tasks. Search results are enhanced by controls such as metadata filtering and bury and boost options.

• Indexing: Vertex AI Vector Search offers robust support for approximation algorithms, including variations of ANN such as ScaNN and other nearest neighbor approaches. The indexes are applicable to either batch updates or streaming data updates in near-real time. Exact matches can be produced using the combination of indexes and a brute force algorithm, which systematically goes through all results.

Opportunities
Google has room for improvement in a few decision criteria, including:

• Multimodality support: Vertex AI would benefit by providing more models with multimodal search capabilities and expanding those multimodal capabilities to include audio data.

• Embedding flexibility: Adding support for byte embeddings, which represent data at the byte level, would make Vertex AI more capable on this criterion. 

Purchase Considerations
Pricing for Vertex AI depends on the specific model used, the volume of input tokens, and whether those tokens are processed via the batch API. Charges apply only to queries that succeed. 

Use Cases
In addition to canonical GenAI applications like RAG development, Vertex AI is a good option for facilitating different facets of data science, including training and deploying advanced ML models. It also delivers personalized recommendation systems and fortifies conversational agents with summarization and question-answering capabilities. 

IBM: DataStax and IBM Db2 Database

Solution Overview
IBM’s vector computing platforms include DataStax and IBM Db2 Database, a fully managed, cloud-based relational offering that supports transactional and analytics workloads. Db2 Database leverages the VECTOR data type so organizations can store and manage vectors within the database via SQL. The engine features a database assistant that supplies configuration tuning, query optimization recommendations, and anomaly detection.

DataStax is built upon three core capabilities:

• Astra DB, which is built atop open source Cassandra, provides knowledge graph and vector database functionality. It supports graph and relational models, integrates with Astra Streaming for streaming data capabilities, offers high availability, and scales linearly.

• DataStax also integrates with Langflow, an open source tool built in Python from the LangChain family of services, which provides a low-code interface for prototyping, creating, and implementing multi-agent and RAG applications.

• Astra DB’s functionality is available across the cloud and on-premises environments in what the vendor terms a hyper-converged database. Astra DB’s Langflow integration allows organizations to choose among various retrievers, LLMs, and embedding models, in addition to test data, all in a hosted solution without significant installation requirements.

IBM is positioned as a Leader and Fast Mover in the Maturity/Feature Play quadrant of the vector databases Radar chart.

Strengths
IBM scored well on a number of decision criteria, including:

• Search variety: Astra DB primarily facilitates lexical search with exact keyword matching through the BM25 algorithm. It uses dense vectors for semantic search and combines these two paradigms for hybrid search capabilities. Similarity search constructs include cosine, Euclidean, and dot product, which users specify at the time of creation of their vector store. IBM Db2 contains an automated performance tuning mechanism that enables the engine to learn from previous query patterns to improve query planning for future searches. The engine also supports hybrid, lexical, and semantic search. IBM Db2 integrates advanced AI-powered query optimization capabilities, allowing enterprises to tap into the potential of AI within their database operations. This feature provides automated performance tuning, enabling the system to continuously learn from historical query patterns and adapt execution plans for optimal performance.

• Results optimization: The results from Astra DB are refined via fine-tuned LLM models devised to provide relevance ranking. Specific rerankers involve those contained in NVIDIA NeMo Retriever reranking microservices, which includes a number of models for enhancing RAG pipelines. Organizations can employ these rerankers at the database layer without configuring them. Astra DB also has options for filtering search results based on metadata.

• Embedding flexibility: Users can choose from multiple options for creating vector embeddings when implementing them on DataStax. They can use models from third-party providers (like Hugging Face) or perform the embedding internally via Astra DB’s Assistants API. This feature wraps OpenAI’s Assistants API to deliver digital assistants that create embeddings using the specific models that the customer organization selects. With DataStax, it’s also possible to search content prior to generating its embeddings.

Opportunities
IBM has room for improvement in a few decision criteria, including:

• Complex data support:  IBM could strengthen its capabilities by enhancing support for multidimensional tensors and matrices (data structures commonly used to store and query high-dimensional vectors). Expanding its ability to index these formats would improve performance in advanced AI workloads. 

• Generative feedback loop: Vendors that integrate feedback mechanisms, such as by using language models or agents to identify preferred results and reinsert them into the vector store, can improve accuracy over time. IBM could strengthen its retrieval accuracy by implementing such feedback mechanisms that learn from user preferences.  

Purchase Considerations
The multifaceted nature of IBM Db2 is one of its chief distinctions, making it inherently value-additive by nature. In addition to processing both analytical and transactional jobs, it lets users store vector embeddings in the same engine that houses their relational data. Thus, developers can rely on the same set of SQL skills they always have to manipulate both structured and unstructured data, improving efficiency and cost-effectiveness. Astra DB also adds value by offering significant capabilities for streaming data, which can inform what data is made available in its vector store. 

Use Cases
IBM’s vector computing platforms underpin an array of generative AI use cases. As with many of its competitors, the more salient use cases include grounding the actions and outputs of agents and language models, respectively, on proprietary enterprise data. IBM’s platforms also support semantic search, hybrid search, and free-text search.  

LanceDB

Solution Overview
LanceDB’s open source, columnar-format vector engine stores an array of data types, including vectors and raw binary data from images, text, video, and audio, along with metadata, all organized in the Lance Format. This proprietary format is inspired by Apache Arrow and interoperable with Apache Parquet, complete with ACID versioning and on-disk vector indexes.

The open source offering, LanceDB OSS, is an embedded engine with Rust, Python, and TypeScript clients. LanceDB Enterprise provides a multimodal data lakehouse service that pairs distributed model training and feature engineering capabilities atop the vector engine. LanceDB Cloud is a SaaS solution with autoscaling, serverless ingestion facilities, and SOC 2 Type compliance. While LanceDB OSS is shipping, neither LanceDB Cloud nor LanceDB Enterprise is in general availability yet.

LanceDB is positioned as a Challenger and Fast Mover in the Innovation/Platform Play quadrant of the vector databases Radar chart.

Strengths
LanceDB scored well on a number of decision criteria, including:

• Multimodality support: Each data type (and its metadata) has a dedicated column within the Lance Format. Thus, one query can encompass tabular data, text, image, and other data types. Users can specify models to embed each column into a vector, and the system also supports inverted-index capabilities for lexical lookups (which can involve BM25 and exact results from the raw tokens). Hybrid search is facilitated by gathering lexical candidates from the inverted index before employing rerankers (dense k-Neural Networks) on embeddings. Since Lance Format maintains raw bytes, embeddings, and metadata next to each other in the same row, users can return results that include just vectors, just metadata, or both.

• Search variety: LanceDB allows users to choose from a wide range of search options, including lexical and keyword search with options for substring matching, fuzzy matching, full text search, and exact term matching.  Users can combine these search options with dense vectors, augmented by reranking neural networks, for hybrid search with improved recall. The engine facilitates similarity search and semantic search, and there are also options for brute force and range search. Users can avail themselves of Euclidean, cosine, and dot-product metrics for each query.

• Indexing: LanceDB provides a number of built-in index types for sparse and dense vectors. It employs an inverted index for the former and includes options like IVF-PQ, IVF-Flat, and IVF_PQ_HNSW for the latter. Metadata is indexed with both bitmap and scalar indexes. The system supports incremental index building and reindexing. 

Opportunities
LanceDB has room for improvement in a few decision criteria, including:

• Generative feedback loop: Although LanceDB can be employed as a feature store to support feature engineering workloads and it provides schema evolution and versioning for tables containing additional features, it doesn’t contain a generative feedback loop. Adding one would make this offering even more compelling. 

• Complex data structure support: LanceDB works well with and supports several scalar types, including bool, decimal, and string. Improving support for multidimensional tensors would add to its effectiveness in this area.  

Purchase Considerations
LanceDB’s open source offering uses the Apache 2.0 license and deploys as conventional software. The pay-as-you-go pricing model that LanceDB Cloud will use when generally available will be based on factors like indexed rows and include a free tier. LanceDB Enterprise will be available through an annual subscription predicated on storage volume and feature add-ons used. The vendor offers training packages and premium-level support for the enterprise and cloud offerings.

Use Cases
LanceDB supports data science workloads, for which it can be leveraged as an embedding warehouse and a feature store. It also supports RAG and agentic AI applications and can facilitate real-time personalization use cases and multimodal search across documents and media.

Marqo: Marqo Cloud*

Solution Overview
Marqo delivers enterprise-scale vector search engine capabilities with a holistic solution designed to optimize results. The engine is architected atop Apache 2.0-licensed open source Vespa, while Marqo Cloud, the managed service, includes numerous features to make the use of this engine an enterprise-grade experience. Besides information retrieval, the vendor specializes in providing embedding models. Marqo is also one of the few solutions in this space to allocate substantial resources to training the embedding models used to create vectors.

The platform includes sample datasets, making it easy to practice validating training data for correct formatting, spinning up and validating training jobs, and deploying embedding models through API calls. These capabilities are enhanced by Marqo’s use of generalized contrastive learning and a model evaluation phase that precedes deployment. The platform offers a Python-Marqo client and infrastructure health checks, delivers high availability, and features integration with DataDog for observability.

Marqo is positioned as a Challenger and Fast Mover in the Innovation/Platform Play quadrant of the vector databases Radar chart.

Strengths
Marqo scored well on a number of decision criteria, including:

• Embedding flexibility: Users can upload the embedding model of their choice to their own storage, including cloud object storage. However, Marqo has numerous embedding models available that organizations can access directly through the platform. The system integrates with Hugging Face, which has a vast range of open source models to select from. Accessing one of these models is as simple as inputting the corresponding URL for the desired model through Marqo’s API before the platform’s inference mechanisms create the embeddings with it. These options, coupled with Marqo’s attributes for training embedding models, increase the overall embedding flexibility of the platform. Marqo can generate vector embeddings using CPUs or GPUs, and can then store embeddings and indexes over them in the storage layer of the underlying Vespa database.

• Multimodality support: Marqo supports the multimodal paradigm that is pivotal for scalable vector database deployments in a few ways. The solution natively supports contrastive language-image pretraining (CLIP), designed to couple images and text in the same vector. Sigmoid loss for language image pretraining (SigLIP) is also available through the platform and, along with CLIP, offers indexes that support natural language search in more than 100 languages. Moreover, organizations can store and query code with audio, video, image, and text data through the system’s customized API. Fixed, multimodal data in the same vector (such as that involving images and text) can be paired with weighted queries. Marqo can index both string fields and tensor fields; the latter is part of its broader capabilities for generation and retrieval of tensor embeddings.

• Results optimization: Marqo is able to optimize the results of its queries using both reranking models and its numerous constructs for custom training of embedding models, which help tailor results to an organization’s particular use case.

Opportunities
Marqo has room for improvement in a few decision criteria, including:

• Generative feedback loop: The addition of a built-in feedback mechanism, beyond the support for custom training of embedding models, would improve and benefit Marqo’s offering. 

• Indexing: The inclusion of more vector index types for exact and approximate matches would make Marqo’s platform even stronger.   

Purchase Considerations
Marqo provides a fully managed cloud service, with pricing starting at approximately $500 a month for end-to-end vector generation, storage, and search. The vendor also offers a customer-managed enterprise class product that includes all of the features of its cloud solution. 

Use Cases
Marqo excels in personalization use cases. In this respect, it’s primed for e-commerce and recommendation applications, built on purchase, clickstream, and event data. It’s also a solid resource for procuring existing models or building custom-trained models for generative AI deployments. 

Microsoft: Azure AI Search, Azure SQL Database, Cosmos DB, SQL Server 2025

Solution Overview
Microsoft offers several products and services that organizations can access independently or as part of Microsoft Fabric, for storing and querying vectors. Azure Cosmos DB is a serverless, fully managed NoSQL vector database for generating, storing, indexing, and querying embeddings alongside their original data, if desired. Azure SQL Database is a fully managed relational database that supports RAG and agentic search. The forthcoming SQL Server 2025, a conventional on-premises OLTP system currently available in preview, will also feature a native vector store. Users can blend SQL data with vector embeddings for this form of hybrid search, and both data types can be managed with the T-SQL syntax.

Microsoft Azure AI Search is a vector database to which organizations can upload data from numerous third-party and Azure-based services. It supports VoiceRAG for natural language speech prompts and responses and multimodality, and integrates with Azure AI Search Lab, an open source experimentation and learning environment for devising and refining search paradigms using Azure AI Search and Azure OpenAI. Azure Database for PostgresSQL, which utilizes the pgvector extension, is also available from Microsoft.

Microsoft is positioned as a Challenger and Fast Mover in the Maturity/Feature Play quadrant of the vector databases Radar chart.

Strengths
Microsoft scored well on a number of decision criteria, including:

• Embedding flexibility: There are several options for vector embedding generation on Microsoft’s platforms. REST interfaces will allow SQL Server 2025 users to employ any third-party AI embedding model. Cosmos DB enables organizations to select models from Hugging Face on Azure and Azure OpenAI Embedding models. Microsoft Azure AI Search accesses dense and sparse vector embedding models from Azure OpenAI. Alternatively, Azure OpenAI Embeddings QnA is a GitHub-hosted open source web application that uses Azure Search as a vector store and facilitates document search via Azure OpenAI. More specifically, it relies on Azure OpenAI Service for embedding content, Azure OpenAI LLMs for conversational questions and answers, and Azure AI Search for information retrieval.

• Search variety: Azure AI Search provides a host of different search techniques, including an agentic search copilot and chat-based interactive queries that work with agents and humans. Organizations can also select from full-text search, vector similarity search, and hybrid search combining the two techniques. Multimodality is possible by using text and image models and joining their results in the same vector space. Organizations can also search in multiple languages and refine results with metadata filtering for numeric and text fields. Azure SQL Database integrates with Microsoft Copilot so that LLMs can analyze the database’s content and supply conversational interactions via RAG.

• Indexing: Cosmos DB includes a range of different indexes to support different use cases. It employs the DiskANN index type, which provides a cost-effective means for storing vector indexes on disk. Users can also choose between a flat index and the vendor’s QuantizedFlat Index. The latter compresses vectors to reduce the storage they require, which constitutes another cost-saving measure.

Opportunities
Microsoft has room for improvement in a few decision criteria, including:

• Complex data structure support: Microsoft’s vector database platforms are not as strong as those of some of its competitors when it comes to more sophisticated functionality for working with high-dimensionality tensors, such as those with mapped, indexed, and sparse dimensions.

• Generative feedback loop: A native construct for implementing a feedback loop, through which successful or preferred query results are input back into the database, would strengthen Microsoft’s support for an assortment of the use cases its vector computing services target. 

Purchase Considerations
Microsoft’s cloud relational service, Azure SQL Database (and soon, the on-premises/customer-managed SQL Server 2025), now includes a vector database that is robust in multiple contexts. The platform simplifies the management and use of vector embeddings by storing them alongside relational data, and enables the use of SQL for vector database use cases. Finally, support for hybrid search opportunities between conventional relational data and vector embeddings broadens the applications this stalwart database supports.

Use Cases
Microsoft’s range of services in the vector database category collectively accommodates a full gamut of vector search use cases. These include applications of similarity and semantic search, hybrid search, RAG, and agent-based architecture fortification. 

MongoDB: MongoDB Atlas Vector Search 

Solution Overview
MongoDB is an open source document-oriented database that provides vector computational capabilities through MongoDB Atlas Vector Search, which in turn derives a significant set of capabilities from Voyage AI (a technology company specializing in advanced embedding and reranking models). Built into the core MongoDB Atlas cloud database, Atlas Vector Search stores vectors alongside conventional scalar data stored in the engine.

Atlas couples MongoDB’s document model with several data services for comprehensive vector data management and information retrieval. One example is Chatbot Demo Builder, which is available as a tool in Atlas Search Playground. Chatbot Demo Builder lets organizations construct a chatbot for answering questions in natural language without writing code. MongoDB Atlas itself provides serverless horizontal scaling with geographically aware fault tolerance spanning the major cloud platforms (AWS, GCP, and Microsoft Azure). It also facilitates high availability and features enterprise-grade security.

MongoDB is positioned as a Leader and Fast Mover in the Innovation/Feature Play quadrant of the vector databases Radar chart.

Strengths
MongoDB scored well in a number of decision criteria, including:

• Embedding flexibility: Organizations can use proprietary and open source models to generate embeddings in MongoDB. Doing so requires either defining a function to facilitate the embedding process or writing scripts. Users can create embeddings from existing data in MongoDB, from newly ingested data, or from search terms. MongoDB has upwards of 10 embedding models from Voyage AI, which also supplies a handful of reranking models. There are models from OpenAI and open source models. After data is vectorized, it’s possible to convert these embeddings into MongoDB’s BSON binary format. 

• Indexing: MongoDB Atlas Vector Search indexes must be created prior to performing vector search. Organizations can devise and manage Atlas Vector Search indexes through the Atlas Administration API, the Atlas UI, a number of MongoDB drivers, and the Atlas command line interface (Atlas CLI). However, the vendor recommends converting embeddings to BSON BinData vectors—with subtypes of int1, float 32, or int8—to optimize storage in Atlas clusters. Supported indexes include HNSW and different variants of  KNN. Specific similarity functions include Euclidean, cosine, and dotProduct. Use of dotProduct requires normalization of the vector to unit at both index time and the time of querying. Organizations can index embedding fields inside documents via dot notation. The system also supports scalar and binary quantization.

• Multimodality support: MongoDB Atlas Vector Search enables users to perform hybrid searches of full-text and vector embeddings. Organizations can employ reciprocal rank fusion and pipelines set up to combine search results for this hybrid approach. Some of Voyage AI’s embedding models support multimodality use cases specific to text and image data. Examples include data pertaining to document screenshots, figures, slide decks, and photographs. Voyage’s multimodal embedding models vectorize content by interleaving data from images and text. Atlas also supports retrieval of image, text, and attendant metadata together in a single query.

Opportunities
MongoDB has room for improvement in a few decision criteria, including:

• Complex data structure support: Many of the pure play vector databases reviewed in this report can store and process a variety of tensors, matrices, and scalars for high-dimensionality data. Increasing Atlas’s capabilities to do the same would enhance the platform’s competitive strength. 

• Search variety: Including native spellcheck or search recommendations, either at result time or as users type, would strengthen MongoDB’s competitive showing in this criterion. 

Purchase Considerations
One big cost advantage of MongoDB Atlas Vector Search stems from customers not needing to synchronize operational or transactional databases with their vector engines, since MongoDB can now perform both of these workloads. Consequently, organizations have less infrastructure, fewer moving parts, and lower overhead costs. 

Use Cases
MongoDB Atlas Vector Search is a strong contender for creating intelligent chatbots, implementing RAG, and providing both semantic and hybrid search.

OpenSearch

Solution Overview
Architected atop Apache Lucene, OpenSearch is an open source search and observability platform with vector information retrieval capabilities. The suite of services enabled by OpenSearch spans security analytics, search, and observability in on-premises, hybrid, and multicloud environments. Its vector engine provides an array of search capabilities, including similarity, multimodal, semantic, neural sparse, conversational, and agentic search modalities.

Users can avail themselves of Query DSL, the system’s domain-specific query language that incorporates a JSON interface. There’s also a robust collection of dashboards for visualizing and exploring data, along with a standalone product, Data Prepper, for server-side data collection, enrichment, aggregation, and transformation. Additional mechanisms include those for building data pipelines for low-latency data ingestion and indexing. Other features include support for SQL and Piped Processing Language (PPL), a proprietary scripting language, alerting capabilities, and geospatial support for improving dashboards with maps alongside conventional numerical analytics.

OpenSearch is positioned as a Leader and Fast Mover in the Innovation/Platform Play quadrant of the vector databases Radar chart.

Strengths
OpenSearch scored well on a number of decision criteria, including:

• Search variety: The quantity of search types available through OpenSearch justifies the platform’s name. Its dense vector support (for semantic and similarity search) is readily coupled with neural sparse vector search for hybrid search use cases. Organizations can also deploy these search varieties individually. Hybrid search capabilities include metadata-based filtering, score blending for lexical  search and vectors, and out-of-the-box normalization via min-max and reciprocal ranking fusion. Neural sparse vector search incorporates sparse vectors for searching text, and multimodal embedding models provide further support for multimodal search. OpenSearch also supports what the vendor terms “agentic search,” in which users ask natural language questions for which a preconfigured agent reads, plans, and retrieves results. The platform also supports brute force search for appropriate use cases and RAG applications for conversational search.

• Embedding flexibility: The three chief methods of facilitating embeddings through OpenSearch involve connecting to third-party foundation models, uploading a model, or relying on the platform’s pretrained models. Organizations can also import their own vectors into the search engine. The system’s AI connector framework enables customers to create customized connectors with any REST-based AI service without having to write code. There are a number of sparse encoding models trained by OpenSearch, such as amazon/neural-sparse/opensearch-neural-sparse-encoding-v2-distill, and several others originally sourced from Amazon. The majority of the text embedding models are from Hugging Face, which OpenSearch uses for a couple of  rerankers. Regardless of the method, OpenSearch can generate embeddings once data has been ingested. Then at query time, the same model is implemented for vectorizing queries to facilitate the retrieval process.

• Results optimization: On OpenSearch’s platform, users can access cross-encoder models, such as those found on Amazon SageMaker, to perform reranking. The system’s rerank processor is implemented in a search pipeline and applies the scoring of the results based on the cross-encoder model’s output. Alternatively, organizations can rerank search results by prioritizing a number of different criteria, including items that match a user’s preferred provider via a metadata field that encodes the relationship. Metadata filtering allows users to refine the results of searches. 

Opportunities
OpenSearch has room for improvement in a few decision criteria, including:

• Indexing: By increasing the number of index types natively available in OpenSearch that users can knowingly select, the information retrieval platform could become more competitive than it currently is. 

• Complex data structure support: OpenSearch provides support for multivectors which, in some cases, can outperform single vectors for both dense and sparse vector applications. Increasing its capabilities for storing, processing, and computing multidimensional tensors could increase the solution’s viability for working with intricate vector data structures. 

Purchase Considerations
OpenSearch is available through an Apache 2.0 open source license, free of any fees. However, many cloud providers offer its engine as a managed service. Examples include Amazon OpenSearch Service and Oracle’s OCI Search with OpenSearch. The pricing varies among the managed service providers. 

Use Cases
OpenSearch facilitates a number of generative AI use cases, from RAG to intelligent chatbots and agent-driven architectures. It’s also widely used for security analytics and observability.

Oracle: Database 23ai and MySQL HeatWave* 

Solution Overview
Oracle’s products and services ecosystem offers a host of capabilities supporting generative AI and vector search. Oracle AI Vector Search, which is a component of Oracle Database 23ai, provides a vector store with the capacity to load documents, issue transformations, facilitate chunking and embedding, perform similarity search, and implement RAG. Significantly, the engine is also equipped with LLMs and capabilities for automating the embedding process. It allows organizations to combine vector search alongside that pertaining to graph, scalar/relational, spatial, and JSON data within the same engine.

Such versatility supports semantic search on unstructured data alongside more traditional search constructs for structured data. AI Vector Search also makes use of Oracle’s VECTOR data type, which can be used for storing and querying vectors inside Oracle Database 23ai’s tables. This allows SQL to be used for similarity search on vectors; when the underlying source data changes, the vectors also change. Oracle Cloud Infrastructure (OCI) also includes generative development (GenDev), which supplies development infrastructure for AI applications.

Additionally, Oracle MySQL HeatWave GenAI is a version of the popular transactional database with vector store capabilities. It also contains a variety of LLMs, scalable vector processing, and constructs for natural language queries with unstructured data. MySQL HeatWave supports its own VECTOR data type with the same advantages afforded by Oracle 23ai’s and provides vector search directly within its own relational store. The platform also provides object storage for data lakehouse interactivity and AutoML capabilities for automated ML model creation.

Oracle is positioned as a Challenger and Outperformer in the Innovation/Feature Play quadrant of the vector databases Radar chart.

Strengths
Oracle scored well on a number of decision criteria, including:

• Indexing: The three index types native to Oracle AI Vector Search are HNSW, which is an approximation algorithm that’s only available in-memory; Inverted File Flat (IVF) with vectors clustered into partitioned tables predicated on similarity; and a hybrid vector index. This final one pairs indexes available from search in Oracle Text (an Oracle Database feature for analyzing, indexing, and searching text) and its lexical applications with the indexes in Oracle AI Vector Search. Organizations can tailor the accuracy of information retrieval for specific applications by expressing it as a default percentage during the index creation phase. The vector creation and search process can be accelerated via optimizations found in Exadata System Software 24ai.

• Embedding flexibility: There are several options for accessing the models that embed enterprise content into vectors in Oracle Database 23ai. It’s possible to import vectors directly into the database. Organizations can also import open source models via ONNX or rely on APIs to create embeddings from third-party model providers. Users can streamline the path to generate vector embeddings in Oracle Database 23ai with OML4Py 2.0, a Python API that downloads pretrained models from Hugging Face. This feature also supplements those models with preprocessing and post-processing capabilities, and converts models to ONNX format prior to loading them into the database. Another model from Hugging Face, all-MiniLM-L12-v2, is also made available to the platform via this approach. There is also a SQL scoring function for generating embeddings; alternatively, integrations with Oracle Autonomous Database enable users to access embedding models from providers like Anthropic (Claude) and Google (Gemini). In MySQL HeatWave, the embedding is facilitated by in-database LLMs, as well as those from external resources. The vector store can automate the embedding generation process. Specific steps included in this automated process include locating desired documents in object storage, parsing and embedding them, and then inserting them into the vector store.

• Results optimization: Oracle’s ONNX framework adoption allows its embedding workflows to access any number of reranking models. The framework also facilitates model serving and model evaluation activities, both of which help optimize query results. Users can filter metadata to further tailor their results in Oracle Database 23ai.

Oracle was classified as an Outperformer partly due to its incorporation of language models utilized within its vector search solutions and to the breadth of models it makes available to users of those systems. 

Opportunities
Oracle has room for improvement in a few decision criteria, including:

• Search variety: Oracle provides several options for users to implement natural language querying of its various systems. Supplementing this with real-time recommendations for searches, as users type, would further enhance the platform’s value and versatility. 

• Generative feedback loop: Enhancing its constructs for providing a feedback mechanism for logging favorable query results and storing them back into the databases would further enhance vector query result quality on Oracle Database 23ai and MySQL HeatWave. 

Purchase Considerations
When deployed on Oracle Cloud Infrastructure, MySQL HeatWave is priced at $0.011 for capacity per hour and $0.02 for each GB stored, per month. Users can get started risk-free with Oracle Database 23ai using its free tier, which provides up to 2 CPUs for foreground processes, 12 GB of user data on disk, and 2 GB of RAM. 

Use Cases
In addition to allowing users to leverage relational data alongside their vector embeddings, Oracle’s vector engine services are designed to facilitate RAG, agent-based AI, and conversational search. 

Pinecone: Pinecone Serverless*

 Solution Overview
Pinecone Serverless is a cloud-native, fully managed service for Pinecone’s closed-source vector database. It features a multitenant retrieval layer facilitating on-demand information retrieval, vector-based clustering on object storage, and a separation of reads and writes. Pinecone Connect enables organizations to integrate Pinecone with external platforms via an authentication workflow, in which those third-party services and platforms can directly access Pinecone’s contents. There are also SDKs for Java, C#, Scala, Python, Node.js, and Go.

The solution has numerous security constructs, including RBAC, customer-managed encryption keys, and audit logs. Private Endpoints for AWS Private Link facilitates access to the retrieval engine via private clouds, avoiding routing traffic over the public internet. In terms of regulatory compliance, Pinecone is compliant with HIPAA and SOC 2.

Pinecone is positioned as a Challenger and Fast Mover in the Innovation/Platform Play quadrant of the vector databases Radar chart.

Strengths
Pinecone scored well on a number of decision criteria, including:

• Results optimization: Pinecone Inference is an API-based service through which organizations can access a variety of reranking and embedding models hosted on Pinecone’s infrastructure. Customers can also use rerankers, such as Cohere Rerank 3.5 and the vendor’s proprietary pinecone-rerank-v0 model, directly on Pinecone’s platform. Metadata-based filtering is accessible through Pinecone Assistant, an API service that also provides ad hoc, RAG-based, natural language Q&A. Organizations can refine query results with indexed metadata, enabling filter condition specification prior to issuing queries, to reduce latency and boost accuracy.

• Embedding flexibility: Pinecone provides multiple options for the embedding process. A model gallery provides access to external, third-party models for generating vector embeddings. The platform also integrates with a number of generative AI platforms, including Hugging Face Inference Endpoints, Cohere, Amazon SageMaker, OpenAI, and Amazon Bedrock, among others. There are also integrations with Elasticsearch, Databricks, and the vendor’s own Pinecone Inference service, the last of which provides both embedding and reranking models. Finally, Pinecone has its own models available through the platform for generating sparse vector embeddings (pinecone-sparse-english-v0). The solution also has a sparse vector index, which can be used for lexical and hybrid search options.

• Multimodality: Users can search through text and images on Pinecone’s platform. The database is well known for its dense vector support, which it relies on for similarity search and semantic search. Pinecone also provides indexes and models that support keyword search. Its cascading retrieval methodology allows users to combine these two approaches with rerankers to implement hybrid search options. 

Opportunities
Pinecone has room for improvement in a few decision criteria, including:

• Indexing: Pinecone supplies indexes for both dense and sparse vectors. Increasing the number of index types it has available for these workloads would strengthen the platform overall. 

• Complex data structure support: Pinecone users can use the engine to store, process, and query multivector embeddings. Boosting the vector computing platform’s capacity to work with high-dimensionality tensors and matrices would nicely augment these capabilities.

Purchase Considerations
Pinecone is available in cloud marketplaces on GCP, Microsoft Azure, and AWS. There’s a free tier for getting started quickly, and a three-week trial period (that includes $300 in credits) for the standard price plan, in which organizations pay as they go to use Pinecone Serverless, Pinecone Inference, and Pinecone Assistant. An enterprise price plan starts at $500 a month. 

Use Cases
Major use cases for Pinecone include generative Q&A with long-term memory, hybrid search, semantic search, and intelligent chatbots. 

PostgreSQL: pgvector*

Solution Overview
PostgreSQL was one of the first relational databases to add support for vector similarity search, which it did through pgvector, one of the numerous extensions in the ecosystem of this leading open source operational database. With pgvector, organizations can ingest, store, index, quantize, and search vectors. pgvector works in concert with numerous other PostgreSQL extensions to enhance its capacity to manage and, in some cases, optimize the retrieval and storage of vector embeddings.

Although connectivity to PostgreSQL includes an open source Type-4 JDBC driver, organizations can access pgvector in the Azure ecosystem through Azure Database for PostgreSQL and Azure Cosmos DB for PostgreSQL. On AWS, the extension is included in Amazon Relational Database Service for PostgreSQL, as well as Amazon Aurora PostgreSQL. GCP users can find and access it through Cloud SQL for PostgreSQL and AlloyDB for PostgreSQL.

PostgreSQL is positioned as a Challenger and Fast Mover in the Innovation/Feature Play quadrant of the vector databases Radar chart.

Strengths
PostgreSQL scored well on a number of decision criteria, including:

• Indexing: PostgreSQL has numerous indexing capabilities for its vector computational engine extension. Inverted File Flat (IVF) provides some utility for sparse vectors, facilitates nearest neighbor search, and is built in concert with parallelization constructs. Other index types include HNSW and StreamingDiskANN, both of which are useful for storing indexes on disk—a cost-saving measure. PostgreSQL expression indexes allow quantization of vectors, which reduces the time for index creation and boosts the number of vectors that can be indexed. B-tree indexes, which are well suited for sparse vectors, are also available.

• Embedding flexibility: Organizations can generate vector embeddings for use with pgvector in different ways. First, pgvector integrates with numerous third-party generative AI platforms and services for embedding content, and users can do so with their programming language of choice (with options ranging from C++ to Python). Another option is PostgreSQL’s pgai extension, which is compatible with pgvector and provides access to OpenAI embedding models. An AI Accelerator Pipelines feature, which is partly based on PostgreSQL’s AI database (aidb) extension, creates intelligent pipelines for generating vector embeddings, storing them, and indexing them. Users simply choose which of the supported models they’d like to use (including those available through an API from OpenAI), and the pipelines handle everything from generating the embeddings to storing them and providing similarity search over them. This approach also updates vector embeddings as their underlying data changes.   

Opportunities
PostgreSQL has room for improvement in a few decision criteria, including:

• Results optimization: The PostgreSQL/pgvector combination provides for metadata filtering of queries and the use of reranking models to refine search results. The addition of multiphase reranking would further enhance the platform’s results optimization story.

• Search variety: PostgreSQL is a solid choice for implementing similarity search, keyword search, and the hybrid of the two. Its search capabilities would stand out even more with the addition of search recommendations provided inline as users type. 

Purchase Considerations
“Vanilla” PostgreSQL is a fully open source product, available via The PostgreSQL License. Organizations can also access it as a managed service from numerous providers, including the major cloud hyperscalers. AWS provides Amazon Aurora PostgreSQL and Amazon Relational Database Service (RDS) for PostgreSQL. Microsoft has Azure Database for PostgreSQL and Azure Cosmos DB for PostgreSQL. GCP features Cloud SQL for PostgreSQL and AlloyDB for PostgreSQL. Other provider offerings, such as Aiven for PostgreSQL and OVHcloud Managed PostgreSQL, also offer pgvector support.

Use Cases
The pgvector extension for PostgreSQL works well for facilitating information retrieval for AI applications such as recommendation systems, RAG, and content-based filtering.

Qdrant: Qdrant Cloud, Qdrant Hybrid Cloud, Qdrant Cloud Inference

Solution Overview
Built in Rust, Qdrant is a high-performance, highly scalable open source vector database that utilizes single instruction, multiple data (SIMD) parallel processing. The vector database platform uses Qdrant’s “Gridstore” key-value engine and employs memory-mapped (MMAP) storage techniques to offload sizable vector datasets to disk. With native support for Docker, Kubernetes, and air-gapped installations, it deploys in edge settings, via Qdrant Cloud, private clusters, and hybrid cloud implementations. The configurable vector store enables users to refine parameters for indexing, hybrid retrieval strategies, memory use, quantization settings, and more. The developer-friendly engine features a built-in web UI and capabilities for scaling via gRPC and REST APIs.

Qdrant is positioned as a Leader and Outperformer in the Innovation/Platform Play quadrant of the vector databases Radar chart.

Strengths
Qdrant scored well on a number of decision criteria, including:

• Multimodality support: Qdrant facilitates multiple named vector fields, each of which can encode a different modality, including audio, image, or textual data, for individual data points. Hybrid retrieval across sparse and dense vectors, as well as their metadata, is possible within the same query pipeline. Additionally, organizations can choose to query metadata or filters without vector embeddings, vector fields without metadata, or any mix of vector fields (which could include image only) and metadata filters in the same query.  

• Embedding flexibility: The robust set of choices for implementing vector embeddings is one of Qdrant’s most prominent differentiators. Users can embed content with their model of choice or use those offered through Qdrant. Qdrant Cloud Inference, which launched earlier this year, generates embeddings directly in Qdrant Cloud with models providing images based on CLIP and text (for sparse and dense) vectors. The vendor’s open source embedding library, FastEmbed, is another option for the efficient generation of vector embeddings. It integrates with the vector database engine so organizations can create and store embeddings during the ingestion process. In addition to offering models for sparse and dense vectors, FastEmbed allows users to index multiple embeddings for documents or sections of documents.

• Search variety:  Qdrant’s lexical search capabilities include full-text search with multilingual tokenization, stemming, phrase matching, and stop words. Lexical search is combined with dense vector search for hybrid search within a single query. Results are merged with distribution-based score fusion or reciprocal rank fusion methods. The system also enables brute force search and dissimilarity search. Organizations can submit positive and negative vector examples, which increases search result relevance by expanding the context for more desirable results. Multivector retrieval is possible as well, via models like ColBERT.

Qdrant was classified as an Outperformer given its robust support for multimodality, embedding flexibility, and cost reductions.

Opportunities
Qdrant has room for improvement in a few decision criteria, including:

• Generative feedback loop: Organizations can create their own feedback loops with third-party resources and implement them in Qdrant through its API. However, Qdrant does not provide such functionality natively; doing so could make it more useful for certain applications. 

• Complex data structure support: Qdrant is able to treat numerous types of scalars, including numbers, geo fields, booleans, and more, as first-class metadata. Adding support for multidimensional matrices and tensors would nicely complete this metadata support. 

Purchase Considerations 
Qdrant’s open source solution is a free offering that deploys in self-hosted environments. Qdrant Cloud is a managed service. It has a free tier that comes with a single gigabyte cluster that scales up once users adopt a paid plan, based on storage, compute, and memory usage. Paid versions of the managed service include premium levels of customer support, high availability, and automated backups. Qdrant Cloud Inference is an add-on to Qdrant Cloud. 

Use Cases
Use cases include several forms of information retrieval relevant to contemporary AI applications, such as enterprise search, multimodal search, hybrid search, and similarity search. Qdrant also specializes in recommendation systems, RAG, and agentic AI deployments. 

Vespa.ai

Solution Overview
Vespa.ai’s AI retrieval platform includes a vector database, ML inference capabilities, and a full-feature text-search engine. The solution is based on a distributed architecture that intelligently distributes ML models, data, and queries across nodes to achieve low-latency, scalability, and fault tolerance. The platform is designed for customization and includes a robust set of SDKs and APIs, so organizations can develop applications with a host of data sources and contemporary tools. Vespa.ai provides logging and holistic metrics for performance tuning and monitoring. It’s particularly suitable for AI workloads requiring online, high-volume transactions. Users can implement its open source distribution or use its managed service offering.

Vespa.ai is positioned as a Leader and Outperformer in the Innovation/Platform Play quadrant of the vector databases Radar chart.

Strengths
Vespa.ai scored well on a number of decision criteria, including:

• Complex data structure support: Vespa.ai natively stores data as tensors. Accordingly, the engine works with tensors of any rank for sparse or dense vectors, including matrices and scalars. Respective values can be doubles, bits, floats, bytes, and bfloat16. The platform can index tensors of any rank if they have only a single dimension. Rank 1 sparse tensors can be indexed with a weighted set before being converted at ranking time to conventional tensors for tensor rank computations. Fields in Vespa.ai’s engine can be scalars and collections of scalars. They can also be indexed with B-trees and dictionaries so they’re accessible for ranking, aggregation, grouping, and sorting.

• Results optimization: Vespa.ai delivers both prefiltering and post-filtering of metadata to refine search results. Users can import reranker models via ONNX or train them using algorithms like XGBoost. Vespa.ai provides reranking in phases. Primary and secondary ranking occurs in distributed content partitions. This process can involve any desired mathematical function, including those supplied by ML models that are directly imported. There’s also a global phase function that can augment or replace Java-based programmatic reranking, applicable once partial results from content nodes are combined. This functionality supports GPU acceleration and API-based access for integration. Multivector models, including ColBERT and ColPali, are natively supported.

• Multimodality support: Vespa.ai’s multimodality approach enables organizations to use different types of fields together: binary data, raw text, and tensors with images and a textual embedding, for example. Because tensors support any number of dimensions, they can involve both mapped dimensions for sparse vectors and indexed dimensions for dense vectors for hybrid search possibilities. 

Vespa.ai was classified as an Outperformer because of its proficiency in storing, retrieving, and processing complex data structures at scale and in real time.

Opportunities
Vespa.ai has room for improvement in a few decision criteria, including:

• Generative feedback loop: Organizations can implement their own language model pipelines to create a generative feedback loop in Vespa.ai. For certain customers, having this capability built in would enhance the offering. 

• Advanced access controls: Although Vespa.ai provides certificate-based and token-based forms of authentication, it doesn’t presently support ABAC. Doing so would enhance the platform’s security story. 

Purchase Considerations
Vespa.ai’s cloud service is available through a pricing model based on units of GPU, CPU, disk, and memory.

Use Cases
Vespa.ai excels in use cases requiring low latency, high scalability, and adaptability. Examples include recommendation systems, ad-serving applications, and agent-based AI deployments. 

Weaviate: Weaviate Database

Solution Overview
Weaviate promotes itself as an AI-native, open source vector computing platform that blends facets of token (keyword), semantic, and hybrid search. Weaviate Database is the core product. Users can optionally choose Weaviate Embeddings, an in-platform embedding service, and Weaviate Agents as add-ons. The latter features a query agent, which performs query-related tasks; a transformation agent, which focuses on data preparation; and a personalization agent. Query Agent is generally available, while the other two agents are available in public preview. Weaviate Cloud Import Tool provides a mechanism for uploading data from CSV files into Weaviate collections. Some of the more recent embedding and reranking model integrations the platform has added are those for Voyage AI, NVIDIA, Cohere (including the multimodal Cohere v4 Embed), and Google Gemini text-embedding endpoints. The engine allows users to dynamically relocate shards within a cluster.

Weaviate is positioned as a Leader and Outperformer in the Innovation/Platform Play quadrant of the vector databases Radar chart.

Strengths
Weaviate scored well on a number of decision criteria, including:

• Embedding flexibility: Internally, customers can use Weaviate Embeddings, which is available through Weaviate Cloud. It produces embeddings for users’ data and queries them within their Weaviate instance. The service directly supports a modest number of embedding models, including Snowflake/snowflake-artic-embed-1-v2.0, and has integrations with numerous external sources for embedding models, including Hugging Face, AWS, Ollama, OpenAI, Jina AI, Databricks, Mistral, and Cohere. Collectively, these capabilities enable users to choose from hundreds of models spanning various domains, modalities, and even written languages. Significantly, the system supports a bring-your-own-embedding-model paradigm and can vectorize content during data ingestion. An integration with NVIDIA Inference Microservices (NIM) lets organizations work with that vendor’s GPUs to expedite the embedding process. This approach is also applicable for reranking and other generative AI tasks.

• Multimodality support: Weaviate facilitates multimodality through the use of vector embedding models purpose-built for combinations of video, text, images, audio, and metadata. Specific models utilized for multimodality in the platform include CLIP and ImageBind. These models create dense vectors that capture semantic meaning. Users can opt to return query results containing vectors, metadata from those vectors, or a combination of the two.

• Indexing: Weaviate has customized implementations of its dense vector indexes. These include HNSW, which provides approximations, and a flat index, which is applicable to vectors stored on disk and relies on brute force search techniques. The vendor also provides a hybrid of the two, and sparse vector indexes largely predicated on inverted indexing techniques. 

Weaviate was classified as an Outperformer in part because of its sophisticated functionality that incorporates AI agents for some of the core capabilities the platform offers.  

Opportunities
Weaviate has room for improvement in a few decision criteria, including:

• Search variety: Weaviate supports semantic, token, and hybrid search. It also has a natural language query agent, several types of search, and a personalization agent in public preview that can tailor the results to users’ preferences. Moving the personalization agent to general availability and improving its capabilities with constructs to provide search suggestions on the fly as users type would fortify Weaviate’s capabilities in this area. 

• Advanced access controls: Weaviate users can currently use RBAC. Supplementing this paradigm with native support for ABAC would solidly improve Weaviate’s access control prowess. 

Purchase Considerations
Weaviate Serverless Cloud pricing is based on factors including compression, vector dimensions, type of SLA, and object count. The service is priced monthly, pay-as-you-go, or on an annual contract basis. Users of Weaviate Enterprise Cloud are billed annually according to AI unit consumption, which in turn is based on compute and storage used.

Use Cases
Use cases for Weaviate include agentic RAG workflows, token-based search, agent-based applications, and semantic search.  

Zilliz: Milvus

Solution Overview
Open source Milvus is a vector database with a C++ based engine, column-oriented storage, and a stateless, cloud-native architecture. The AI retrieval platform allows users to store, manage, and query vector data with support for indexing, search, and generating embeddings at data ingestion time. The offering relies on hardware-savvy features, including GPUs, SIMD, and NVMe SSDs, and parallelized operations.

Zilliz Cloud, a Saas solution, is the vendor’s fully managed vector database platform with the same APIs as Milvus, and it uses an advanced, proprietary engine called Cardinal. A Spark-enabled vector data lake ETL and query service is available through Zilliz Cloud, which also contains a web-based UI and dashboard. The cloud service implements autoscaling and a distributed architecture with query node group-based sharding. 

Zilliz is positioned as a Leader and Fast Mover in the Innovation/Platform Play quadrant of the vector databases Radar chart.

Strengths
Zilliz scored well on a number of decision criteria, including:

• Indexing: The sheer quantity of index types natively supported by Milvus is itself significant. Whereas many competitors only offer one or two index types, Milvus offers a broad range, including DiskANN, HNSW, FLAT, SCANN, IVF, RabitQ-optimized indexes, and GPU-accelerated indexing courtesy of cuVS and Cagra. The majority of the indexes use Zilliz Cloud’s proprietary indexing engine, AutoIndex, which relies on ML to select and implement the most suitable algorithm. The engine works with sparse and dense vectors and tunes itself to the underlying data once a setting from 1 to 10 is selected for recall versus performance. SPARSE_INVERTED INDEX capabilities are provided. Approximate search modes include DAAT-MAXSCORE and DAAT_ WAND, while a TAAT_NATIVE index issues more exact matches. Additionally, each scalar and vector metadata field can be assigned its own type of index. 

• Embedding flexibility: Milvus provides comprehensive embedding capabilities through multiple integration approaches and built-in functions, supporting both external AI services and internally managed embedding models. The library also facilitates integrations with open source models, including Sentence Transformers, Instructor, SPLADE, and Model2Vec. Milvus 2.6 contains a feature for direct integration with third-party embedding models, allowing users to query raw text while the engine calls a user-defined model service to vectorize text with low latency. Automatic embedding generation is possible during ingestion via a function module, with which customers can use vector embedding model services. The returned embeddings are stored in expressly designated vector fields in Milvus collections: two-dimensional tables with variant rows and fixed columns.

• Multimodality support: Milvus supports multivector search and has a flexible schema design that includes multiple modalities such as video, audio, text, and structured metadata. Users can combine these modalities in a single search. There are options to query vectors, metadata, or both in a single query.  

Opportunities
Milvus has room for improvement in a few decision criteria, including:

• Results optimization: Milvus allows users to access several reranking models. Augmenting this with reranking in phases would enhance the vendor’s standing in this area.

• Search variety: Milvus provides an open source GUI, Attu, that lets users sift through related results to searches via graph-oriented visualizations. Complementing this capability with native support for on-the-fly recommendations, supplied as users type would improve the vendor’s capability in this category.  

Purchase Considerations
Open source Milvus is available via an Apache 2.0 license and can be deployed as self-managed software or on physical and virtual appliances. Zilliz Cloud is the SaaS solution. Zilliz Cloud BYOC lets users deploy the platform into their own clouds.

Use Cases
The broad number of applications supported by both open source Milvus and Zilliz’s offerings includes semantic search, fraud detection, agent-based AI applications, and RAG.

Vector search capabilities are likely to remain a permanent fixture within the general database landscape and for data management. The prominence of AI applications is only growing, and vector database platforms are central to responsible, sustainable deployments of AI technologies, particularly those reliant on language models. As such, IT decision-makers will encounter a need to incorporate these engines into their existing technology stacks sooner rather than later.

However, there is a mounting divide between the approaches taken by general-purpose database vendors, including those offering data warehouse or operational database platforms, and by vector database specialists. The former have broader, established constituencies and organizations that are already acclimated to and heavily invested in their product ecosystems. The appeal of these vendors and the vector database enhancements they have made to their platforms is clear.

On the other hand, specialists in this space typically provide more advantages in terms of cost reductions and advanced functionality, especially when working with tensors and more complex numeric representations of data. Many have been implemented by well-known organizations across diverse industries, providing a formidable challenge to established database platforms in which customers have made previous investments. Additionally, these specialists are expanding their offerings to include support for external object storage, custom training of embedding models, and other advanced abilities, making them that much more competitive with general-purpose vendors.

As is often the case, the correct decision for a specific organization comes from determining which of these approaches is more appropriate for their own use cases. Often, starting with proverbial “low-hanging fruit” applications of RAG or semantic search may provide the most straightforward means of testing a solution while delivering the greatest value in the shortest time. Evaluating each approach and platform prudently will typically involve implementing a proof-of-concept test with specific vendors. However, it also requires a considerable amount of solo effort on the part of organizations.

Astute buyer organizations will determine what data sources are necessary for their use cases and map that to the embedding, storage, and search capabilities offered by prospective vendors, which differ greatly. They’ll also want to identify a content “chunking” strategy and discern the way vendors handle that process. Some employ a generic strategy, while others stratify it according to the format and type of content to be embedded. 

Organizations must also dutifully research the automation capabilities offered by each vector database solution. One of the best means of assessing vendors early on in talks may be to have them vectorize the customer’s content (from its website, for example) and make that available for semantic search during a PoC. If the results are encouraging, organizations can hone in further on chunking and embedding model particulars and evaluate long-term costs and cost optimizations vendors offer in terms of storage, quantization, and compression.

To learn about related topics in this space, check out the following GigaOm Radar reports:

• GigaOm Sonar for AI Governance

• GigaOm Radar for Streaming Data Platforms

• GigaOm Radar for Data Observability

*Vendors marked with an asterisk did not participate in our research process for the Radar report, and their capsules and scoring were compiled via desk research.

For more information about our research process for Radar reports, please visit our Methodology.

Andrew Brust has held developer, CTO, analyst, research director, and market strategist positions at organizations ranging from the City of New York and Cap Gemini to GigaOm and Datameer. He has worked with small, medium, and Fortune 1000 clients in numerous industries and with software companies ranging from small ISVs to large clients like Microsoft. The understanding of technology and the way customers use it that resulted from this experience makes his market and product analyses relevant, credible, and empathetic.

Andrew has tracked the Big Data and Analytics industry since its inception, as GigaOm’s Research Director and as ZDNet’s original blogger for Big Data and Analytics. Andrew co-chairs Visual Studio Live!, one of the nation’s longest-running developer conferences, and currently covers data and analytics for The New Stack and VentureBeat. As a seasoned technical author and speaker in the database field, Andrew understands today’s market in the context of its extensive enterprise underpinnings.

Jelani Harper has worked as an information technology editorial consultant and journalist for over 10 years. During that time he has helped myriad vendors and publications in the data management space strategize, develop, compose, and place content in a variety of outlets, spanning mainstream outfits such as Forbes, Tech Crunch, and VentureBeat to trade journals like The New Stack, TDWI and KMWorld. He has produced an assortment of technical content including white papers, solutions briefs, contract proposals, marketing materials, thought leadership articles, bylines, and blogs for clients specializing in nearly every facet of data management.

As such, Jelani has focused extensively on the numerous dimensions of cognitive computing, cloud computing, analytics, data governance, data engineering, and data integration. His work has enabled him to conduct a number of substantive interviews with both established and progressive vendors in the arenas of High-Performance Computing, semantic technologies, quantum computing, cyber security, the Internet of Things, blockchain, and more. Content under his own byline has been cited or paraphrased by organizations such as SAS, TAMR, and DAMA International, as well as publications like VentureBeat, IT Business Edge, and The Data Administration Newsletter. He’s spent the last couple of years working with Blue Badge Insights.

GigaOm provides technical, operational, and business advice for IT’s strategic digital enterprise and business initiatives. Enterprise business leaders, CIOs, and technology organizations partner with GigaOm for practical, actionable, strategic, and visionary advice for modernizing and transforming their business. GigaOm’s advice empowers enterprises to successfully compete in an increasingly complicated business atmosphere that requires a solid understanding of constantly changing customer demands.

GigaOm works directly with enterprises both inside and outside of the IT organization to apply proven research and methodologies designed to avoid pitfalls and roadblocks while balancing risk and innovation. Research methodologies include but are not limited to adoption and benchmarking surveys, use cases, interviews, ROI/TCO, market landscapes, strategic trends, and technical benchmarks. Our analysts possess 20+ years of experience advising a spectrum of clients from early adopters to mainstream enterprises.

GigaOm’s perspective is that of the unbiased enterprise practitioner. Through this perspective, GigaOm connects with engaged and loyal subscribers on a deep and meaningful level.

© Knowingly, Inc. 2025 "GigaOm Radar for Vector Databases" is a trademark of Knowingly, Inc. For permission to reproduce this report, please contact sales@gigaom.com.