What Is Weaviate and Where Does It Fit in the AI Stack?

---

Weaviate is an open-source vector database, BSD-3-Clause at its core, that stores embeddings and retrieves them through native hybrid search: BM25 keyword matching fused with vector similarity in one query, a design that has helped its GitHub repository pass 16.8k stars as of September 2026. Code under the wl/ directory ships under the separate Weaviate License and needs a licence key, which is why GitHub reports the repository’s licence as NOASSERTION. Atlan’s context layer sits upstream of Weaviate the same way it sits upstream of any vector database, governing what gets embedded before Weaviate ever indexes it.

Founded in Amsterdam in 2019, Weaviate’s real distinction from most vector database vendors is a deployment philosophy: run it yourself, or let Weaviate run it for you, on the same underlying code.

• Native hybrid search, tuned by an alpha parameter that weighs keyword matching against vector similarity
• Collections, cross-references, and vectorizer modules, closer to structured data than a flat vector index
• Deployment choice: Weaviate Cloud, Docker, Kubernetes, AWS, GCP, DigitalOcean, Snowpark, or an experimental embedded mode
• A 2025-2026 agentic push, led by the Query Agent, routing natural-language questions across collections

---

What is Weaviate?

Weaviate is a company and an open-source project, not a single product SKU: an engine anyone can run for free, and a business that sells a managed version of the same code. Bob van Luijt and Etienne Dilocker incorporated the company in Amsterdam in 2019, building on an open-source project the two had already been developing since 2016. In an interview with Microsoft’s Agent Framework team, CEO van Luijt described the category plainly: “A vector database like Weaviate is stateful and directly integrates with generative machine learning models.”

Funding has scaled with that thesis: a $50M Series B announced in April 2023 led by Index Ventures, with Battery Ventures and existing investors. Gartner forecasts a 75.3% CAGR for the vector database segment of the DBMS market, and Fortune Business Insights projects it growing from $2.58 billion in 2025 to $17.91 billion by 2034.

A vector database exists because LLMs are stateless between calls: a model has no memory of what it retrieved a moment ago unless something outside it stores and re-serves it. Pinecone, Qdrant, Milvus, and Weaviate all answer that problem, but Weaviate’s answer to “who operates this” is structurally different from a managed-only competitor: a BSD-3-Clause core you can run yourself is Weaviate’s default identity, not a managed-only SKU.

---

How does Weaviate’s architecture work?

Three structural pieces separate Weaviate’s data model from a flat vector index: collections, vectorizer modules, and cross-references.

Collections and schema

Weaviate organizes data into collections (formerly “classes”), each defining properties, a configured vectorizer, and optional cross-references, graph-style links between objects in different collections, per Weaviate’s documentation. A support-ticket collection can cross-reference the customer it belongs to, and a query can traverse that link at retrieval time, closer to structured data than a flat vector index. Weaviate scopes the feature itself in its data concepts documentation: cross-references are beacon-based pointers, directional, described as “links that help you retrieve related information,” and “queries involving cross-references can be slower than queries that do not involve cross-references, especially at scale.” They are pointers, not the reasoning layer of a knowledge graph. See vector database vs. knowledge graph for agent memory for that tradeoff.

Vectorizer modules

Each collection’s vectorizer is set once, at creation, and can’t be swapped without re-indexing. It connects to an external embedding service, OpenAI, Cohere, Google, or HuggingFace among them, converting text or other modalities into the embeddings Weaviate stores, the same step every transformer-based vector database depends on. Generative modules extend this further, letting a query call an LLM directly against retrieved objects, sometimes called RAG inside the database rather than around it.

Underneath, Weaviate indexes with HNSW, the same approximate-nearest-neighbor algorithm Qdrant and Milvus use. Pinecone does not publish its index algorithm. HNSW isn’t what makes Weaviate different; what differs is what sits around the index, the schema and the modules, and how a query searches it.

---

What is hybrid search in Weaviate?

Hybrid search is the mechanism Weaviate leans on hardest. A query runs BM25 keyword search and vector similarity search in parallel, then fuses both into one ranked list, per Weaviate’s hybrid search documentation. An alpha parameter controls the blend, pure keyword (0) to pure vector (1), the same tradeoff covered generally in semantic search vs. keyword search.

Pinecone supports hybrid search natively too. What differs is the default: Weaviate fuses BM25 and vector scores in one query without being asked, using Relative Score Fusion. That is distinct from hybrid RAG, which fuses dense and sparse retrieval across a broader pipeline, not one query.

Whether what got retrieved deserves to be there is a separate question, and one this page comes back to.

---

What deployment options does Weaviate offer?

Deployment is where Weaviate’s identity diverges most from a managed-only vector database. Weaviate documents eight routes, per Weaviate’s deployment documentation:

That table is the direct structural answer to a question Pinecone’s own honest-limitations section leads with: there is no self-hosted or on-premises option. Weaviate’s answer to the same question is the opposite, a genuinely different starting point for a team deciding what it wants to own.

Weaviate Cloud’s current tiers, per Weaviate’s own pricing page, replaced older Serverless/Enterprise naming: a Free tier that’s always free up to 100,000 objects, Flex starting at a $45 monthly minimum, and Premium starting at $400 a month on a prepaid contract, with shared or dedicated deployment.

Whichever path a team picks, the decision is reversible in a way a closed system’s isn’t: the same open-source core runs underneath both, so migrating between them means changing operators, not re-architecting.

---

Where does Weaviate fit in the AI stack?

Most vector-database explainers skip half of this question: not just what Weaviate does, but where it sits relative to everything else an AI system needs.

Weaviate is a standard partner integration in both major orchestration ecosystems: langchain-weaviate wraps it as a retriever inside LangChain, the same retrieval-layer role it plays in the broader retrieval orchestration agentic AI depends on: orchestration decides what an agent does next, not what it’s allowed to retrieve.

More recently, Weaviate has pushed further up that stack. Its Query Agent reached general availability in September 2025, routing natural-language questions across collections, decomposing complex queries and reranking results before answering. A companion Transformation Agent augments existing objects with generative models. As memory needs grow past one collection, this overlaps with what agentic AI memory and multi-agent scaling already require: consistency across many concurrent readers and writers.

Routing intelligence isn’t certification, and the closing sections come back to that gap.

---

How much does Weaviate cost?

Two cost models mirror the two main deployment paths. Self-hosting is free and open-source; the cost shows up as infrastructure and engineering time, work Weaviate Cloud otherwise absorbs. Weaviate Cloud’s pricing follows the same Free, Flex, and Premium tiers covered in the deployment section above.

A headline number ages fast; Weaviate’s own pricing page is the source to check before committing budget. The real comparison is total cost of ownership over 12 to 18 months, not today’s Free-versus-Flex math. A team with existing Kubernetes expertise may find self-hosted materially cheaper; a team without it usually finds the ops time costs more than the subscription would have. Neither path answers whether the content indexed under either model was worth indexing, a context window problem that sits above pricing entirely.

---

What are Weaviate’s honest limitations?

Weaviate’s flexibility comes with real tradeoffs.

Self-hosting means owning sharding, scaling, and schema design that Weaviate Cloud otherwise runs. Weaviate scopes some features itself: its data concepts documentation warns that “queries involving cross-references can be slower than queries that do not involve cross-references, especially at scale.” Retrieval quality also depends on decisions taken before anything reaches the index, like chunking strategy. Cost at scale is a tradeoff every managed vector database shares, not unique to Weaviate.

The deeper point sits one layer up. The Query Agent routes a question to the right collection with real sophistication, and multi-tenancy has isolated tenants into dedicated shards since v1.20. Both operate on collections whose contents were decided upstream: whether a document was certified for AI use, whether it has gone stale, whether a business term means what the business thinks it means. That upstream decision is the same cause behind why AI agents fail in production and the broader RAG accuracy problems that have little to do with which vector database sits underneath.

Weaviate stores and retrieves. Deciding what belongs in a collection is a separate, upstream job.

---

How does Weaviate compare to other vector databases?

The right vector database depends on deployment control, existing infrastructure, and how much operational work a team wants to own. Weaviate’s open-source, self-hostable identity sets it apart from Pinecone’s managed-only model; Milvus, Qdrant, and ChromaDB occupy similar open-source territory, each with its own tradeoffs this page won’t relitigate. A team facing a knowledge-graph-shaped problem, not a retrieval one, should look at context graphs instead of stretching Weaviate’s cross-references to do that job.

Atlan covers the full landscape in Top vector databases for enterprise AI, an eight-way comparison; readers starting from fundamentals should begin with What is a vector database? instead. This page’s job was narrower: what Weaviate is, how its architecture and hybrid search work, and where its real limits sit, with semantic search implementation, knowledge base construction, and advanced RAG techniques covered elsewhere.

---

What a governed context layer adds upstream of Weaviate

Weaviate answers a narrow question well: given this query, what’s closest in what’s already indexed, increasingly with the Query Agent adding intelligence about which collection to ask. A different question gets answered earlier, wherever the embedding pipeline starts: was this data certified for AI use, and is a model safe to see it at all.

Stack AI chose Weaviate over Pinecone for cost and deployment flexibility; co-founder Antoni Rosinol put it in people terms: “The biggest benefit of using Weaviate isn’t just the technology – it’s the team behind it.” Morningstar built its “Mo” research assistant on Weaviate. Both are proof of retrieval quality, which is what the case studies set out to establish.

Atlan sits upstream of that gap, drawing on the same data catalog and systems of semantics discipline: certification so only owner-approved data enters the embedding pipeline, sensitivity classification blocking PII before it becomes a vector, freshness triggers tying a source update to a re-embed signal, and lineage tracing every vector back to its source table. Atlan’s MCP Server, through the Model Context Protocol, lets an agent querying Weaviate check whether what it retrieved was ever supposed to be there. A semantic layer closes the gap between “semantically similar” and “means what the business thinks it means.”

The enterprise context layer, built on Atlan’s Enterprise Data Graph, sits above deployment flexibility and hybrid search entirely, answering a question neither touches: not how fast retrieval runs, but whether it was retrieval of something trustworthy to begin with. Context engineering means building that check in from the start, not finding its absence after an agent already acted on something it shouldn’t have. Teams that treat implementing an enterprise context layer as a parallel track don’t rebuild the pipeline six months in.

---

Weaviate solves retrieval; trust is solved upstream

Weaviate’s genuine strengths are deployment flexibility, native hybrid search, and an increasingly agentic layer on top. Whether what gets retrieved deserves to be trusted is a context problem rather than a retrieval one, and it gets settled where the data enters the pipeline: certification, classification, and lineage. Solving deployment flexibility and hybrid search precision is a real achievement. It is a different one.

---

FAQs about Weaviate

1. What is Weaviate used for?

Weaviate is used to store and retrieve embeddings for AI applications, most commonly retrieval-augmented generation, AI agent memory, and semantic search. Its native hybrid search suits teams wanting both keyword precision and semantic recall from a single query, self-hosted or through Weaviate Cloud.

2. Is Weaviate open source?

Yes, with one qualification. Weaviate’s core is open-source under the BSD-3-Clause license, free to self-host on Docker or Kubernetes; code under the wl/ directory ships under the separate Weaviate License and needs a licence key, which is why GitHub reports the repository as NOASSERTION. Weaviate Cloud is the company’s managed alternative, priced across Free, Flex, and Premium tiers, and the same core powers both.

3. How does Weaviate differ from Pinecone?

Weaviate is open-source and deployment-flexible, self-hosted or managed; Pinecone is closed-source and managed-only, though Bring Your Own Cloud runs its data plane in a customer’s account. Both support hybrid search natively; Weaviate fuses BM25 and vector scores in one query by default, using Relative Score Fusion.

4. What is hybrid search in Weaviate?

Hybrid search in Weaviate runs a BM25 keyword search and a vector similarity search in parallel, then fuses the results into one ranked list. An alpha parameter, from 0 (pure keyword) to 1 (pure vector), controls the blend. Relative Score Fusion has been the default algorithm since Weaviate v1.24.

5. Can I self-host Weaviate?

Yes. Weaviate’s BSD-3-Clause core runs anywhere Docker or Kubernetes runs, on your own infrastructure or any cloud provider, at no licensing cost. Self-hosting means owning sharding, scaling, and schema design that Weaviate Cloud otherwise runs, a real operational tradeoff worth budgeting for.

6. Does Weaviate support multi-tenancy?

Yes. Since v1.20, Weaviate’s multi-tenancy assigns each tenant a dedicated shard within a collection, giving logical and physical isolation rather than a metadata-filter-based separation. That distinction matters for workloads needing hard isolation between customers or business units, not just a filtered view of shared data.

7. What is the Weaviate Query Agent?

The Query Agent is Weaviate’s own agentic feature, reaching general availability in September 2025. It routes a natural-language question across multiple collections, decomposing complex queries and reranking results automatically. It routes retrieval. Whether a collection’s contents were certified for AI use is decided upstream of it.

---

Sources

1. About Us, including the $50M Series B announcement, Weaviate
2. weaviate/weaviate repository and LICENSE, GitHub
3. Revolutionizing AI Search with Weaviate: An Interview with CEO Bob van Luijt, Microsoft Agent Framework Devblog
4. Managing Collections, Weaviate Documentation
5. Data Concepts, Weaviate Documentation
6. Hybrid Search, Weaviate Documentation
7. Deploying Weaviate, Weaviate Documentation
8. Weaviate Pricing, Weaviate
9. Query Agent, Now Generally Available, Weaviate Blog
10. Weaviate Multi-Tenancy Architecture Explained, Weaviate Blog
11. Case Study: Stack AI, Weaviate
12. Case Study: Morningstar, Weaviate
13. Forecast: Database Management Systems, Worldwide, Gartner
14. Vector Database Market Size, Share & Industry Analysis, Fortune Business Insights