sequenced.ai
Articles/Search & research/Blueprint//8 min read

How turbopuffer searches AI application data from object storage

A rarely opened workspace and a busy one can share a search service with different cache behavior. Namespace design affects latency, retrieval quality and spend.

By Sequenced deskAI-assisted, source-led · how we work
Visit turbopuffer website ↗
Object storageDurable stateSuccessful writes persist in an object-store log.
NamespacesQuery scopeEach query addresses one namespace.
BM25 + vectorsHybrid retrievalText and semantic results can be combined.
Usage minimumBillingMonthly usage floors vary by plan.
turbopuffer mark
turbopufferturbopuffer.com · independent research

Represent this company? Verify your work email to access its workspace, or send the desk a factual correction.

At nine on Monday, an employee opens a project workspace nobody has visited in weeks. A colleague is already searching a busy workspace for the twentieth time that morning. The two people expect the same simple search box. Behind it, turbopuffer keeps durable search data in object storage and uses faster caches for active work. Designing around those two visits is a better test of the architecture than timing twenty identical searches in a row.

In brief
  1. 01The durable layer turbopuffer writes to object storage. Memory and local solid-state storage help serve hot queries.
  2. 02The retrieval layer A namespace can support both full-text and vector ranking, with options for combining or reranking the results.
  3. 03The economic choice The service charges for usage with a monthly minimum. The data being searched and its organisation influence that bill.

01 / ProductA search engine for data an application already owns

turbopuffer is a managed search engine for applications that need to retrieve text, embeddings and associated attributes. It describes object storage as its durable source of state. A successful write adds data to the namespace's write-ahead log, and indexing makes those changes efficient to search. Compute uses memory and local solid-state storage for caches. The architecture lets data persist even when a particular namespace receives very little traffic. turbopuffer engine overview Search architecture

The unit a developer encounters in a query is the namespace. Each query searches one namespace, and the service's documentation recommends separating documents into namespaces where that suits the access pattern. A project-management application might give each customer its own namespace. That layout makes a customer's usual search narrow and well-defined. A company-wide search that crosses several such boundaries would require additional application design. Query API and consistency

Its AI role is retrieval. An embedding can help find a passage about a concept even when the searcher does not remember its original wording. Full-text ranking can preserve the importance of a distinctive word or identifier. A product can show the resulting documents directly, or pass selected passages to a model as context. Grounding an answer still requires checking that the retrieved passage actually supports what the generator says. The vendor describes deployments at companies including Notion and Linear. Hybrid retrieval and rank fusion turbopuffer customer deployments

02 / AudienceWhich search patterns suit an object-store design

A product with many customer workspaces and highly uneven traffic is a natural place to investigate. Paying for a large corpus that spends much of its time idle is a different problem from keeping one small shared index consistently fast. The right question is how people experience the first query after a quiet period and how often they make follow-up searches while the workspace remains warm.

A code assistant has both precise and conceptual queries. A developer may ask where a named class is defined or describe an unexpected behaviour without knowing the responsible function. An internal knowledge tool has much the same split between a named form and a question about a policy. Find out whether users are disappointed by missing words, missing meaning or stale results. Each problem points to a different change in the retrieval system.

An application that must be predictably fast on the first lookup after a long break should examine the documented cold-query trade-off. A sustained benchmark against a popular namespace will understate that first-visit cost. Conversely, an architecture chosen solely for its cold performance may be unnecessarily expensive for thousands of low-traffic workspaces whose members accept a modest initial wait. Decide which experience matters by observing the product, not by assuming every namespace behaves like its busiest customer. turbopuffer engine overview

03 / WorkflowA proposed trial for searching customer project notes

Choose three workspaces the test team is authorised to use, with distinct traffic patterns and project vocabularies. Prepare searchable passages with stable document and project identifiers, dates, titles, and links back to their originals. Keep one copy of the canonical source so an update or deletion has a clear route into the search index. Choose which fields should be returned to the user and which should be used only for filtering.

Write queries with the people who know the projects. Include a ticket number, a project nickname, a recently renamed component and a description of a decision whose exact wording nobody remembers. For each query, identify a passage that would answer it and any obsolete note that would look plausible but lead the reader astray. A human can now distinguish retrieving useful evidence from merely returning something that talks about the same subject.

Run a BM25 full-text search and a vector search on the same namespace. Compare where the expected passages land in each list before combining them. The hybrid guide describes server-side reciprocal rank fusion and optional reranking as well as application-side choices. Fusion based on positions gives an application a way to combine different kinds of scores. A later reranker can reorder candidates it receives, but it cannot rescue a passage omitted from every earlier list. Hybrid retrieval and rank fusion

Test a newly corrected project note. Write the update, record acknowledgement and immediately search using the old and new terminology. The architecture guide describes strong consistency by default and an optional eventual mode with different freshness behaviour. A change to a decision record makes freshness visible to a tester in a way that an unchanging demo corpus never will. Use the ordinary access path and check the underlying document link as well as the search excerpt. Search architecture Query API and consistency

Repeat a representative group of queries after genuinely quiet intervals. Capture the first experience, subsequent latency and the quality of results as distinct observations. Have a different account attempt its legitimate search for a similarly named project and inspect what comes back. The application must resolve the correct namespace and permissions for each request. A technically accurate match in a workspace the reader cannot access is still an unacceptable result.

04 / PricingHow plans and usage interact

The pricing page lists Launch at a $16 monthly minimum, Scale at a $256 monthly minimum and Enterprise from $4,096 per month with a 35% usage premium. The company describes unused commitments as expiring at month end. Above the minimum, actual usage determines the charge. Its calculator asks for dimensions, attribute size, documents, writes, queries and namespace distribution. A price per thousand searches divorced from those assumptions will miss much of the bill. turbopuffer plans and usage pricing

PlanMonthly basisRelevant provision
Launch$16 minimum usageMulti-tenant database and community support
Scale$256 minimum usageAdditional controls including single sign-on and audit logs
EnterpriseFrom $4,096 and a 35% usage premiumPrivate networking and additional deployment choices

Plan figures in dollars, consulted 21 November 2026. turbopuffer plans and usage pricing

The underlying accounting uses logical bytes. Vector dimension and precision change the amount associated with each document. The pricing FAQ says half-precision vectors use fewer bytes for storage and queries than full precision, while writes count half precision as four bytes per dimension because maintaining the index still incurs work. A trial evaluating smaller vectors should compare retrieval quality and each usage component. A storage saving is not an equal percentage saving on every operation. turbopuffer plans and usage pricing

05 / DistinctionsHow to compare retrieval services on the same work

The Pinecone Blueprint is a companion for teams primarily evaluating managed vector retrieval. Use the same passages, embeddings and relevance judgments to compare candidates. Then repeat with the deployment's actual update patterns and traffic distribution. Otherwise a difference caused by a new chunking rule or a better embedding model can be incorrectly credited to the database underneath it.

The Typesense Blueprint offers another comparison for a team building hybrid search. A catalogue with a single shared collection and strong requirements for precise identifiers may have a different operating profile from a collaboration product with a long tail of quiet customer workspaces. Compare the cost of keeping the index available during quiet hours and the effect of the first search after those hours. Preserve the same user-visible output and filters during both trials.

06 / QuestionsWhich operating assumptions deserve another look

A popular workspace can conceal an awkward namespace boundary. Picture a person who can see several project teams but only part of each team's records. If the query service has to fan out across many namespaces and combine partial lists, the application's complexity and latency can exceed the cost of the individual requests. Prototype a real cross-project view as well as the straightforward single-project case before locking in the partition scheme.

A second concern is rebuilding the searchable representation when the source or embedding model changes. An updated model may produce vectors with different dimensions or meanings. Decide how an application could prepare the new representation, validate it against saved questions and switch readers to it without mixing incompatible vectors. Price the period in which two versions may coexist. Rolling out a better retrieval model can temporarily create more storage and writes than ordinary operation.

07 / DecisionChoose the next test by the job of the search box

01

A service with many quiet workspaces

Select several real traffic shapes and record the first search after an idle interval alongside a warm repeat. Include the least active workspace in the cost estimate.

Decide whether the first-visit wait fits what the product promises.
02

A team adding semantic retrieval

Label examples where exact words matter and where a paraphrase should win. Inspect the candidate lists before and after fusion while holding the underlying notes constant.

Keep the semantic path where it finds evidence people actually need.
03

A team pricing a migration

Measure document dimensions, writes, query distribution and the proportion of traffic that crosses namespace boundaries. Include plan minimums and the migration overlap.

Base the purchase on the service's actual shape and the controls it needs.
What should we explore next?

A business worth understanding.

Suggest your business or one you find interesting. Tell us what you want to understand about its product, positioning, design or workflows.

Suggestions are free. Selection and publication stay with the desk.

Sources
Filed under Search & researchCompany turbopufferNot affiliated with turbopufferRequest a correctionRequest a refresh by email

Continue reading

All in this category