A help centre can have the answer to a customer's question and still lose the search. The customer writes “money went back on my old card” while the article is called “Refunds after card replacement.” Meilisearch can use an embedding model to bring those phrases closer together. Yet the same help centre may also have an article for a particular payment error code, where the exact characters matter much more than the subject. Its hybrid search gives the publisher a choice about both.
- 01The engine A single index can return keyword and semantic matches. An embedder configuration connects the documents and incoming queries to a model.
- 02The editorial choice A document template governs what goes into the embedding. Titles, symptoms and product context can be more useful than a full page of repeated navigation.
- 03The operating choice Meilisearch Cloud prices managed instances by region and size. The open-source engine is also available for self-hosting.
01 / ProductHow a help article becomes a candidate result
In a full-text search, wording, spelling and the fields chosen for the index help determine which document appears. Semantic search gives the application a second signal. An embedding model processes selected text and produces a vector representation. Meilisearch can use it to find documents connected by meaning. The hybrid search overview describes automatic embedding and a choice of providers. Teams that already generate vectors can also bring them to the engine.
The mechanism gives an editor two controls that are easy to confuse. At indexing time the document template determines what the model gets to read. At query time the semantic ratio changes the mix of semantic and literal matching. Moving the ratio cannot restore a distinguishing product name or symptom that was excluded from the model's input. Changing a template, meanwhile, changes what an entire set of documents means to the retrieval system. Meilisearch's explanation of the two controls
Meilisearch returns matching records. A help centre can present the article titles, summaries and links directly, giving the reader the underlying instruction and its owner. If another product later writes an answer based on those records, that adds a separate decision about accuracy and attribution. A team can learn a great deal about its support corpus before introducing that extra step.
02 / AudienceThe best starting point is a library with recognizable answers
A useful first collection has articles that a support team already uses to resolve identifiable cases. An agent can say which document answers a question and why an older one is wrong. A help centre with both consumer and business editions is an interesting example. “How do I change a billing address?” looks similar in either edition, but a plausible answer for the wrong account type can send a customer looking for settings they cannot see.
It is harder to evaluate semantic retrieval on an undifferentiated archive of meeting notes. Two notes may be close in meaning while only one describes a current decision. If there is no owner, publication state or date policy in the source system, a model cannot decide who has authority. Before migrating that archive, identify what constitutes a legitimate result. A smaller approved collection may provide a more honest first deployment.
03 / WorkflowA proposed experiment for a payments help centre
Take a set of resolved questions and remove customer information. Have agents identify the current answer, any acceptable alternatives and the reason an apparently similar page would be wrong. Include exact error codes, informal descriptions, questions with two meanings and cases that have no published answer. Preserve the original wording. Cleaning every question into the language of the knowledge base would eliminate the problem the project is supposed to examine.
Build a document template around article title, affected product, symptoms and the substance of the resolution. Leave recurring footers and general company messaging out of the text sent to the model. Include a code in a searchable field when customers actually look it up. If an article covers multiple unrelated errors, consider splitting the article at the source. That editorial change gives future readers a more precise destination regardless of which search mechanism finds it.
First run the queries with literal matching only. Repeat with hybrid search and compare the actual order of the top few results. Tune the semantic ratio on one group, then try it on a second group that played no part in the tuning. The getting-started guide also documents a minimum ranking score. Try it on deliberately unanswerable questions. An empty list with an escalation route may be kinder than a convincing link to the wrong refund policy.
Access can be tested separately from relevance. Mark a set of articles for a single merchant audience and use the search rules embedded in a tenant token to restrict its records. Meilisearch describes deriving these tokens from an API key and applying document filters to queries. In the proposed trial, ask the same question as two users with different entitlements. Compare which records they are allowed to retrieve before asking whether those records are ordered well. Multi-tenant search guide
04 / PricingThe cloud estimate begins with an instance and a region
The directly consulted Cloud pricing page lists a starting price of $20 per month, with regional differences, and a 14-day trial. Its estimator uses document count, average document size and search traffic to recommend a configuration. The displayed starting amount is a guide to entry, not a quote for a library full of vector data. Record the selected region and the instance and disk assumptions when comparing managed and self-hosted operating costs.
| Route | Published basis | What to budget |
|---|---|---|
| Managed Cloud | From $20 per month, varies by region | Instance, disk, outbound bandwidth and embedding provider |
| Enterprise | Custom quote | Required support, service level and features in the agreement |
| Self-hosted open-source engine | No engine licence fee | Servers, storage, backups, upgrades and embedding costs |
Hosting options from Meilisearch pricing, consulted 19 November 2026. Dollar amounts follow the vendor’s display.
The pricing FAQ lists a 32 GiB minimum for disk and outbound bandwidth at $0.15 per GB. Include the size of the results that leave the service as well as the region where it runs. A monthly hosting estimate also leaves out the work of changing a corpus. Rewriting a title is routine editorial work. Changing which fields are sent for embedding can require fresh model processing across many records. To put a value on the distinction, estimate one initial import, routine article edits and one full reprocessing exercise. An externally billed embedding service may treat those operations differently from the search engine that stores the result.
Merchandise and support teams can also ask to control particular result positions. The search rules product page describes pinning, boosting or hiding records for matching conditions, with its own allowances and optional charges. Keep these requirements visible when discussing a plan. A small index that needs extensive changes chosen by editors may have quite different purchasing needs from a larger reference index with stable ordering.
05 / DistinctionsWhen a ranking rule is worth more than a different model
Suppose the help centre issues a time-limited announcement about a service change. For a precisely defined query, an editor may want that notice to appear before every background explanation. That is an explicit editorial decision and can be tested as such. Increasing semantic weight across all queries to promote one announcement risks shifting many unrelated results. Conversely, maintaining individual pins for every way someone can describe a refund is a poor substitute for measuring general relevance.
The Algolia Blueprint offers a neighboring comparison for teams that want to put search operations into a product manager's hands. Consider how each tool lets staff inspect an individual query and account for what a rule did. Elastic belongs on a different shortlist when search is part of a wider data and analytics estate. The practical question is which system the team can tune and operate with the expertise it actually has.
06 / QuestionsWhat happens when the best-looking match has expired
An article about an old payments interface may describe the customer's problem more clearly than its successor. Semantic proximity alone gives no reason to withhold it. Decide whether retired articles should be excluded during ingestion, filtered during retrieval or shown only to staff. Then change the state of a real sample record in the source and measure when the correct behavior reaches search. An excellent result yesterday can become a costly support detour tomorrow.
Keep a small collection of queries where the right response is to ask another question. A user may say “refund failed” with nothing to distinguish a rejected transaction from a delayed bank posting. A search interface that exposes those two paths with useful summaries can outperform a single confident result on that task. Judge the experience by whether a person reaches the right instruction, including the step of selecting the relevant product or account.
07 / DecisionPick the first deployment by the cost of a wrong result
A public help centre
Choose a product with clear article ownership and a meaningful sample of informal questions. Compare literal and hybrid results on both error codes and descriptions.
A private customer portal
Define who can see each article and test the same queries through separate tenant identities. Revisit relevance after the allowed result set is correct.
A team considering managed hosting
Run a representative import during the trial, record the proposed region and size, and estimate model processing for a full content refresh.
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.
- Meilisearch hybrid searchConsulted
- Getting started with hybrid searchConsulted
- Meilisearch Cloud pricingConsulted
- Meilisearch guide to multi-tenancyConsulted
- Comparing full-text and semantic searchConsulted
- Meilisearch search rulesConsulted

