Elastic combines keyword and vector retrieval for AI applications

Elastic brings lexical search, vector retrieval and access controls into Elasticsearch, with distinct hosted, serverless and self-managed routes.

01 / ProductElastic is a retrieval platform with a broader operational portfolio

Elastic is the company behind Elasticsearch, a distributed search and analytics engine. Its wider products address observability and security, while its AI relevance comes from retrieving useful context from structured records, text and vectors. A language model can generate an answer; Elasticsearch supplies mechanisms for finding the information that answer should use.

The key distinction is that exact matching and semantic similarity can coexist. An equipment code, legal clause number or product identifier often needs a precise match. A question phrased differently from the source material may benefit from vector retrieval. Hybrid search combines those approaches into one result list rather than forcing a team to choose one for every query.

Elastic offers cloud-hosted, serverless and self-managed deployments. The pricing overview distinguishes control over cluster configuration from the more managed serverless model. These are operating choices as much as purchasing choices: they affect who manages versions, capacity and resilience, and how the application’s costs should be measured.

This blueprint examines public product and technical documentation. Sequenced has not benchmarked Elasticsearch latency, retrieval quality or operating costs. The workflow below is a proposed support-search implementation and should be tested against a representative document collection before wider rollout.

02 / AudienceThe fit is a search problem that needs both meaning and precision

A software support team is a useful example. Customers describe a symptom in ordinary language, but the correct answer may depend on an exact version number, error code and deployment type. A purely semantic match can retrieve a plausible answer for the wrong release; a purely literal match can miss a useful article that uses different vocabulary.

Elastic is relevant when developers need to control that combination, maintain filters and inspect why results were returned. It also becomes attractive when the organisation already uses Elasticsearch and can build on existing indexing and operational skills. Existing use does not remove the need to assess the new vector workload independently.

Compare Glean when the priority is a packaged enterprise search and assistant experience across workplace systems. Compare Pinecone when the central requirement is a dedicated vector retrieval service. Neither comparison should be reduced to one relevance score: the decision includes ingestion, exact matching, filters, operational ownership and the complete bill.

03 / WorkflowProposed workflow: retrieve the right support answer for the right version

Start with approved support articles and release notes. Give each document a stable identifier, publication state, product family, supported version range, language and access audience. Preserve the original source link. A retrieval result that points to a superseded draft can be harmful even when it is semantically close to the question.

Create an evaluation set from real categories of support questions, with sensitive information removed. Include precise error codes, descriptions with no known code, ambiguous product names and questions whose answer is absent. The expected output should identify acceptable source material, not merely an ideal sentence that a model might reproduce.

Elastic’s semantic_text workflow sends indexed text to a configured inference endpoint and stores the resulting embeddings. The managed Elastic Inference Service is one documented route, and other inference services can be configured. In the proposed implementation, record the chosen embedding configuration alongside the index so a later model change is treated as a retrieval change.

Keep literal fields for error codes and identifiers. Run a keyword query and a semantic query, then combine the rankings. Elastic recommends reciprocal rank fusion for its hybrid workflow. That combines positions in result lists rather than assuming a keyword score and a vector similarity score have the same meaning. Evaluate the resulting order on the support questions that matter to the team.

Apply product, version and publication filters as part of retrieval. An article explaining an obsolete setting may still be relevant to an older customer, so avoid simply deleting all old material. Store enough metadata to choose the correct revision for the question, and make missing version information visible to the user when it changes the answer.

For private support material, establish the requesting user’s access before returning context. Elastic documents field and document security, but those controls are intended for read-only privileged accounts. Use a separate ingestion identity and test the interaction of roles. A prompt telling an assistant to respect permissions is not a substitute for filtering the retrieved data.

Only after retrieval is useful should a language model summarise the selected passages. Return source links and let the application decline an answer when the available context is insufficient. Measure source correctness separately from prose quality. A fluent explanation of the wrong software version is still a failed support result.

04 / PricingServerless prices describe several independent usage dimensions

The Elasticsearch Serverless pricing page lists starting rates for ingest, search and machine-learning virtual compute units, plus retained storage and transfer. It also explains that a baseline of search resources remains provisioned while a project is idle. Serverless therefore does not mean every component becomes free when users stop searching.

The rates below are the page’s USD starting figures consulted on 16 September 2026, not a quote for a configured project. Region, search settings, data footprint, workload and support can affect the total. Embedding and language-model use introduce additional dimensions; estimate them separately from storage and search compute.

Elastic also documents a distinct Vector Database billing model, based on storage, reserved search capacity, ingested volume and an infrastructure fee. Do not apply the general Elasticsearch Serverless VCU table to that offering. Select the product first, then use its current calculator and billing documentation.

05 / DistinctionsHybrid retrieval makes exact business identifiers part of the AI design

The useful distinction is the ability to combine familiar search requirements with vectors. Support teams rarely want similarity in isolation. They want results that are relevant, current for a particular release, available to the requester and easy to trace to a maintained source. Elasticsearch gives engineers several mechanisms for representing those constraints in the retrieval layer.

This can also make investigation clearer. If the result misses an exact error code, inspect the literal field and query. If a paraphrased question retrieves unrelated articles, inspect the embedding and candidate selection. If an authorised article is absent, inspect filters and roles. Separating these causes is more productive than repeatedly rewriting the answer prompt.

Elastic’s wider portfolio can be relevant to organisations already operating its stack, but a shared vendor is not automatic evidence of a simpler architecture. Keep the support index’s permissions, retention and resource needs explicit. An observability workload and a customer-facing retrieval workload may have very different data sensitivity and latency requirements.

06 / QuestionsThe important limits are index freshness, permissions and deployment costs

Define how quickly an approved article revision must become searchable. An application can update its primary content store before the retrieval index reflects the change. Test a replacement and deletion through the full path, including any answer cache. A retrieval platform cannot make an outdated source correct merely by ranking it well.

Recheck the chosen security features against the deployment and subscription. Elastic’s documentation explains the mechanics of roles, but the buyer still needs to validate its exact edition and identity flow. Test two users with deliberately different access and inspect both result titles and returned fields; even an inaccessible document’s title can disclose information.

Finally, test idle as well as busy costs. A support index with infrequent traffic still occupies storage and may reserve search resources. Run a representative quiet period, record the configured search setting and compare actual usage with the estimate. Include reindexing after an embedding change, since the maintenance workload can be materially different from ordinary daily queries.

07 / DecisionChoose Elastic when retrieval control is part of the product

Elastic is a strong evaluation candidate when exact terms, semantic matching and access-aware filters must work together. Build a measured retrieval baseline before adding generated answers, and make the deployment choice against both the active workload and the resources retained between searches.