Weaviate combines vector and keyword search for AI applications

Weaviate is a database company focused on retrieval for AI applications. Its platform combines semantic search with structured data and keyword matching, with managed cloud services alongside an open-source database. The useful question is whether a team needs a configurable retrieval system behind its own application. A database can make relevant objects discoverable, but the application still needs a clear definition of relevance and an experience that helps people act on the results.

01 / ProductThe database and the surrounding Weaviate services

Weaviate’s platform brings together vector search, retrieval-augmented generation and agent memory. The database is the foundation; cloud offerings include managed operation and additional AI services. This article concentrates on building retrieval over a known collection, which is a different project from adopting every service in the portfolio.

The hybrid search documentation describes combining a vector search with a BM25F keyword search and fusing their result sets. Relative weights and fusion methods can be configured. This gives developers a way to accommodate both descriptive questions and precise terms without pretending those requests have identical matching requirements.

That flexibility introduces a design responsibility. A product search that returns something semantically related may still be wrong if it ignores an exact connector type or model number. The retrieval system needs to represent those constraints explicitly. The final application should make the reason for a match understandable, particularly when the result will inform a purchase or an operational choice.

02 / AudienceSearch teams with both descriptive and precise queries

Weaviate is relevant to developers building catalog search, document discovery, recommendation features and evidence retrieval for assistants. It is particularly interesting when a single corpus has several kinds of information: free text, exact identifiers, categories and access boundaries. Those teams benefit from evaluating retrieval behavior directly rather than hiding every result behind generated prose.

A useful buyer already has a collection it can curate. If product descriptions are incomplete or categories conflict, adding semantic retrieval will not repair the underlying records. Start by checking a representative sample of source objects and defining which fields are authoritative. The work can expose a data maintenance problem before it becomes a search quality problem.

Organizations already centralizing AI and analytics in a larger data platform should compare the operational fit with Databricks. A specialist retrieval database can be attractive, but it adds another system to synchronize. The right boundary depends on where the authoritative data lives and who will maintain the search-specific representation.

03 / WorkflowA proposed search experience for an equipment catalog

Imagine a proposed search tool for a distributor of industrial equipment. Buyers ask for items by part number, by intended task and by an imprecise description of a replacement. This example is an editorial design exercise, not a tested Weaviate deployment. Begin with a narrow product family and retain the existing exact-match catalog search as a baseline.

Represent each item with a stable identifier, a clear description and the attributes that decide compatibility. Keep discontinued status and replacement relationships in authoritative fields. A visually or semantically similar component should not outrank a verified replacement merely because its description happens to resemble the customer’s phrasing.

Build a set of judged requests with the catalog specialist. Include an exact part number, a misspelled name, a task description and a request containing mutually incompatible requirements. The expected result may be a shortlist, a clarification or no suitable item. Recording those different outcomes prevents the evaluation from rewarding a system simply for always returning something.

Use the hybrid search controls to compare different balances of meaning and exact wording. In this proposed catalog, raise attention to lexical matching for identifiers while checking that ordinary descriptions still find plausible candidates. Do not choose a single weight from a few attractive examples; inspect the failures as carefully as the successful demonstrations.

Separate hard eligibility from ranking preference. Availability, permitted markets and supported equipment versions should determine which items can be shown. Ranking then orders eligible candidates. A generated explanation can help describe a result, but it should not change the compatibility record or invent a replacement relationship that the source catalog does not contain.

The multi-tenancy guide describes tenant data stored on separate shards and notes that multi-tenancy must be enabled. If distributors have private catalogs, decide that boundary during collection design. In the example application, authenticated account context should select the tenant; a user-supplied search phrase should only influence the search within that scope.

After the first pilot, stage changes deliberately. Compare the same judged requests before and after updating descriptions or retrieval settings. Include newly discontinued products and recently added replacements in the checks. A useful release record explains which failures were corrected and whether an improvement in descriptive search weakened exact-identifier behavior.

04 / PricingWeaviate Cloud pricing and capacity choices

The current cloud pricing page lists Free, Flex and Premium routes. The amounts below are displayed US-dollar starting figures checked on 16 September 2026. A starting price does not establish the cost of a specific object count, vector configuration or optional AI service.

For the equipment catalog, size the estimate around objects, stored text, vector dimensions and the required operating configuration. Then add the cost of refreshing source data and serving the application. A retrieval-only experience and an assistant that generates an explanation for every search have different economics even when they query the same collection.

Self-operated software also has a cost model: infrastructure, upgrades, recovery and engineering time. Treat that route as an operating choice rather than a zero-cost substitute for cloud. A team that already runs database services may value control; a small product team may prefer paying for managed operation to keep attention on catalog quality.

05 / DistinctionsRelevance is configurable at several levels

The indexing overview distinguishes vector indexes from inverted indexes and describes collection-level configuration. Indexing fields that will never be queried adds storage without helping a reader. For the proposed catalog, inventory the actual search and filter requirements before enabling every available representation.

Weaviate’s hybrid approach is meaningful because everyday search mixes exact and descriptive language. The same buyer may start with “compact pump for a washdown area” and then enter a specific part code. Those are two useful behaviors to support in one application, but success requires evaluating each rather than averaging them into an opaque overall score.

Cohere’s retrieval models provide a relevant comparison at the embedding and reranking layer. Changing the model and changing the database configuration are distinct interventions. In a controlled pilot, hold one constant while testing the other so that a quality difference leads to a clear implementation decision.

06 / QuestionsQuestions that determine the collection design

How will the application behave when a requested item is unavailable? A search system may find a close match, but the business must decide whether that is an acceptable substitute. Label candidate alternatives visibly and preserve the reason the original item was excluded. This makes a near match useful without presenting it as guaranteed compatibility.

Which deployment responsibilities belong to the team? Managed cloud, independent operation and enterprise configurations require different procedures for updates, backup validation and incident response. Write down the expected recovery outcome using the actual catalog size and update cadence. A deployment option matters only if the organization can operate it at the standard its users need.

How will tenant and catalog lifecycle changes be tested? Create a small second tenant with intentionally overlapping product names and verify that its private records remain unavailable to the first. Then remove a catalog entry and check search again. These cases are more informative for the proposed application than a polished demonstration on one static dataset.

07 / DecisionChoose Weaviate around a measurable search requirement

Weaviate merits a pilot when hybrid retrieval, structured objects and deployment choice fit a concrete application. The best next step is a small judged collection and an interface that exposes why each object was returned. Expand capacity and add generation after the retrieval behavior is understood, not simply because the platform offers them.