sequenced.ai
Articles/Search & research/Blueprint//7 min read

How Typesense combines exact matches and semantic search in one catalogue

A part number and a description of a problem need different kinds of search. Typesense lets a team serve both from one collection and size the hardware behind it.

By Sequenced deskAI-assisted, source-led · how we work
Visit Typesense website ↗
Rank fusionHybrid rankingCombines keyword and vector result positions.
In memorySearch indexRaw records also have a copy on disk.
Scoped keysResult accessA search key can carry enforced filters.
Per hourCloud computeDedicated cluster time accrues while provisioned.
Typesense mark
Typesensetypesense.org · independent research

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

Someone who types a replacement filter's part number expects the right filter at the top of the page. Someone who types “dust comes through the vacuum” may need the same part without knowing its name. Typesense gives a catalogue operator two ways to retrieve the record and a way to combine their results. The practical decision is how much meaning to add before an exact lookup begins to get worse.

In brief
  1. 01The product An open-source search engine with typo tolerance, filters and vector search, available to run directly or as a managed cluster.
  2. 02The AI choice The application can supply embeddings or ask Typesense to create them from selected fields. Product identifiers still have a job in ordinary text search.
  3. 03The cost driver Cloud cluster capacity runs by the hour. The number of searches alone cannot tell a team what its bill will be.

01 / ProductA search engine can rank the same record twice

A Typesense collection holds documents such as products, articles or parts. A team decides which fields people can search, which values the interface can filter and which information it only displays. Someone browsing a parts shop might use the title to search, the appliance family to narrow the list and a photograph to confirm the result. The engine also offers facets, sorting and grouping, so the decision concerns the whole results page. Typesense product overview

An embedding represents information as a vector that can be compared with other vectors. Typesense can build one from selected document fields or accept one produced elsewhere. A query can then find records with similar meaning even when its wording differs from the title. In a hybrid query the engine runs text and semantic retrieval and combines their rank positions. The version 30.2 guide gives keyword results more weight by default and exposes an alpha parameter to change the mix. Vector and hybrid search

That distinction matters for diagnosis. If the intended part is absent from both candidate sets, changing their relative weight cannot bring it back. Start with the record, the indexed fields and any applied filter. If the record ranks well in one set and poorly in the combined list, then weighting becomes an interesting experiment. This sequence gives a small search team a way to explain a change to the person who maintains the catalogue.

02 / AudienceWho benefits from controlling the collection

A parts distributor has unusually good reasons to keep field-level control. A product specialist knows which numbers identify compatible parts, which names customers use interchangeably and which descriptions are ambiguous. An engineer can turn those distinctions into schema choices. Together they can answer two different questions about a disappointing result. Did the buyer use unfamiliar wording, or did the supplier never specify the machine the part fits?

A publisher has a different problem. Someone looking for a particular article may remember an exact headline fragment. Someone researching a theme may remember only the idea. Search can accommodate both, provided the publisher has useful text and a means to judge whether neighboring stories actually cover the subject. For a small list of clearly named items, a simple text index may already solve the job. More storage and another ranking signal need a measurable benefit.

03 / WorkflowA proposed spare-parts search trial

Pick one appliance family with enough compatible and incompatible parts to expose mistakes. Keep the original supplier identifier in a dedicated searchable field. For the embedding, try the plain-language product name, appliance type and description of the part's function. A string of shipping codes repeated across every product contributes little descriptive information. Keep compatibility as structured catalogue data so a similarity match cannot silently overrule a known exclusion.

Build a small query set from actual customer language that the business can use for evaluation. Separate exact identifiers, incomplete names and descriptions of symptoms. Have a product specialist record which items are acceptable for each query and which would be an expensive wrong answer. For an ambiguous symptom, the desired next step might be a question about the appliance model. The review sheet should allow for that outcome instead of rewarding every returned item.

Run the keyword, semantic and combined configurations against the same catalogue snapshot. Record the first few positions, where the correct result first appears, and whether an incompatible part gets above it. Follow up on individual reversals. An embedding field may need a better product description, a filter may be suppressing a valid match, or a few exact identifiers may deserve a separate lookup path. These proposed observations would be more actionable than one average across all three kinds of search.

If different trade customers have different catalogues, factor access into the trial. Typesense documents scoped search keys that can enforce filters. Issue keys for the intended set of records from a trusted application and exercise the ordinary search and facet views for two permitted customer accounts with different catalogues. Check the returned fields as well as the visible titles. A correct part number is still an incorrect result if the account has no entitlement to see it. API key scopes and filters

04 / PricingAn idle managed cluster still has a price

The managed service allocates dedicated capacity. Its billing explanation describes an hourly cluster fee, bandwidth and optional charges for backups, extra disk and prioritized support. Extra disk is calculated per node. The cluster calculator exposes hardware, region and availability choices. Save the proposed configuration with the estimate so a quotation for a small trial cannot be mistaken for the price of its production replacement.

ComponentBilling basisWhat to estimate
ClusterHourly while runningCapacity and hours provisioned, including quiet periods
BandwidthData transferredSize of records returned and traffic sent
Optional capacityBackups and extra diskBackup storage and extra disk per node

Managed cost components, consulted 19 November 2026. Typesense Cloud billing.

Memory is an important part of this arithmetic. Typesense keeps the search index in memory and a copy of the raw documents on disk. The vendor distinguishes indexed fields from display-only fields and gives separate estimates for vectors and for loading a built-in embedding model. A remote embedding service moves the model out of the cluster but still leaves the vectors in the index. Budget its provider charges separately. System requirements and sizing examples

As a proposed budgeting exercise, price the same catalogue at launch and after a second supplier is added. Include the number of unique products, vector dimensions and fields actually searched, then add the expected hours each cluster remains up. A temporary import rehearsal and a permanent replica have different cost implications even if each handles the same records. If operating the engine internally is plausible, compare the installation and ongoing operating work too. Installation options

05 / DistinctionsWhere this fits beside managed search and a vector database

The Algolia Blueprint is a useful comparison for a team choosing who will own the search experience and its commercial arrangement. Price a specific catalogue and a realistic pattern of browsing against both offers. A service measured by dedicated capacity behaves differently when actual traffic falls below forecast. Also compare the daily work of the catalogue editor who is asked to explain why a particular item moved.

The Pinecone Blueprint addresses a neighboring infrastructure decision. An application that already has a strong keyword experience may chiefly need a vector retrieval layer. A team starting from the catalogue itself may prefer a single engine where filtering, exact matching and vector results come together. An informative comparison holds the query set, compatibility rules and human judgments constant while changing which components the team operates.

06 / QuestionsWhat could invalidate the trial

A successful test on richly described products says little about a supplier feed full of abbreviated names. Choose examples from the least complete supplier and inspect what actually enters the embedding. Repeat an import with changed compatibility information and a discontinued item. The time from that change to a correct result is a business requirement in its own right, especially when a support agent uses the result to advise a caller.

Consider how people recover when they get the wrong item. A query reformulated with a precise identifier should give the user a dependable way back. Record cases where it fails and set a threshold that catalogue owners understand. The choice of model or cluster size matters less if the team cannot distinguish an irrelevant match from a correct part attached to an outdated supplier record.

07 / DecisionChoose according to the way people look things up

01

A distributor with reliable identifiers

Establish an exact-match baseline with live compatibility rules. Add a semantic path for symptom descriptions and inspect every case where a known identifier falls in rank.

Keep the identifier test in every relevance review.
02

A publisher with a thematic archive

Select stories whose subject is easy to recognize but hard to name. Ask an editor to judge the first few matches for both known-headline and exploratory queries.

Invest in embeddings where readers gain useful routes into the archive.
03

A team comparing deployment options

Measure the actual indexed fields and vector footprint. Price a managed configuration and name the people who would maintain an equivalent self-hosted setup.

Choose the arrangement that fits both the data and the operating team.
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 TypesenseNot affiliated with TypesenseRequest a correctionRequest a refresh by email

Continue reading

All in this category