New announcementCheck out Typesense on San Francisco buses View Photos
background

Typesense vs Meilisearch

Both products offer fast, typo-tolerant search with filters, facets, geo search, and hybrid search. Meilisearch is interesting when its memory-mapped storage lets an index exceed available RAM and that associated performance tradeoff is acceptable. Its production replication can keep reads available, but it requires Enterprise Edition or Cloud, and a failed static leader blocks writes until manual promotion. Typesense is more battle-tested in high-scale production environments, and includes automatic Raft leader election in its open-source server, along with type validation, stable JOINs and nested filters, Natural Language Search, and RAG.

For a broader view, see the full Typesense, Algolia, Elasticsearch, and Meilisearch comparison covering all four products.

Typesense vs Meilisearch at a glance

Swipe left to see Meilisearch

AreaTypesenseMeilisearch
Core search Typo-tolerant keyword search, filtering, faceting, query-time sorting, geo search, vector search, semantic search, and hybrid search Typo-tolerant keyword search, filtering, faceting, query-time sorting, geo search, vector search, semantic search, and hybrid search
License and hosting All API capabilities, including Raft high availability, are available in the open-source server. Self-host it free or run the same server in a dedicated Typesense Cloud cluster. Community Edition is MIT licensed. Enterprise Edition code is public under a source-available BUSL license, but it cannot be freely used in production. Production sharding and replication require Enterprise Edition or Cloud.
Storage and memory Keeps fields used for search, filtering, sorting, and faceting in memory; source documents and unindexed fields remain on disk Memory-mapped LMDB storage can serve indexes larger than RAM, though Meilisearch performs best when it has enough RAM to hold the entire dataset.
Schema and validation With ingestion-time type validation, Typesense supports automatic schema detection or strict schemas to keep collections clean Schema-less. Mixed field types are accepted. Meilisearch's sorting guidance warns that they can produce unexpected order.
Typo toleranceEnabled by default and tunable per query and per field Enabled by default, with index-level thresholds and exclusions. It can be disabled for selected attributes or words, but typo thresholds cannot be changed per query.
Runtime sorting Choose one or more sort fields and directions at query time on one collection. For example, price low-to-high, price high-to-low, and newest all use sort_by. Choose one or more sort fields and directions at query time on one index after configuring sortable attributes. For example, price low-to-high, price high-to-low, and newest do not need separate indexes.
Nested arrays of objects Stable correlated nested-object filters keep multiple conditions bound to the same object inside an array Precise array relationships use experimental foreign keys. This model is unavailable with remote sharding, and a foreign filter returns an error when its related-side condition matches more than 100 documents.
Relationships between collections Stable cross-collection JOINs with filtering, nested joins, field inclusion, faceting, and sorting Experimental one-way foreign keys and hydration are unavailable with remote sharding. A foreign filter returns an error when its related-side condition matches more than 100 documents.
Geo fields per record Any number of independently queryable named geo fields. For example, one record can have separate pickup, delivery, warehouse, and service-area locations. Each record supports one reserved _geo point for distance sorting, plus one _geojson point or shape. Distance sorting does not work with _geojson.
Vector, semantic, and hybrid search HNSW semantic and hybrid search with built-in models, OpenAI, Azure OpenAI, OpenAI-compatible APIs, Google Gemini or Vertex AI, and user-provided vectors Semantic and hybrid search with DiskANN and embedders for OpenAI, Hugging Face, Cohere, Mistral, Voyage, Gemini, Cloudflare, Ollama, custom REST APIs, and user-provided vectors
Natural Language Search Typesense has built-in Natural Language Search that uses an LLM to turn requests such as “the most powerful car under $50K” into structured search terms, filters, and sort parameters. There is no built-in equivalent for translating a natural-language request into structured search, filter, and sort parameters. Semantic search and conversational RAG solve different problems.
Conversational Search / RAG Production conversational search API with streaming, conversation history, and configurable LLM models RAG chat with streaming and workspaces is available, but it remains an experimental feature.
Multi-tenancy Scoped keys can lock each user to permitted organizations, roles, fields, query limits, and an expiry time without a server round trip JWT tenant tokens can restrict indexes, enforce document filters, and expire
Search analytics Available in the open-source server and Typesense Cloud, including search, click, conversion, and visit events Meilisearch Cloud has a search analytics dashboard for queries, no-result rate, latency, clicks, conversions, and geographic distribution
Merchandising and curation The open-source API and Typesense Cloud UI support includes, excludes, query replacement, filters, sort controls, and rule-based curations Search rules are experimental and pinning-only today. Three rules per Cloud project are free; additional rules are a paid add-on.
High availabilityRaft clustering in the open-source server with automatic leader election. Typesense Cloud supports 3 to 7 nodes and reroutes traffic away from failed nodes. Enterprise Edition or Cloud replication automatically retries a network search on another replica when a remote is unavailable. One static leader handles every write. There is no automatic leader election, so leader failure blocks writes until manual promotion, a known limitation for write availability.
Geo distribution Typesense Cloud's Search Delivery Network distributes a cluster across up to 7 regions from 26 choices, routes searches to the nearest node, and fails over automatically Enterprise replication supports geographically routed reads. Optimal routing is not currently available when sharding and replication are combined.
InstantSearch UI The Typesense-InstantSearch adapter lets Algolia's widgets work with Typesense for JavaScript, React, Vue, and Angular instant-meilisearch connects Meilisearch to the InstantSearch.js ecosystem
Hosted crawler The Typesense DocSearch scraper crawls and indexes documentation sites, but the crawler must be self-hosted Meilisearch Cloud includes a hosted crawler with JavaScript rendering, DocSearch mode, and schema extraction
Serverless offeringNo. Typesense runs as provisioned persistent compute. No. Standard Meilisearch Cloud runs on provisioned persistent compute.
Support hours Typesense Cloud includes 24/7/365 critical production support for HA clusters, with faster response targets on higher support tiers Email support is included in Cloud. Enterprise offers 24/7 support targets, including a one-hour target for urgent production outages.

How the storage and replication models affect production

Typesense keeps indexed search structures in memory

Typesense sizes memory around the fields you search, filter, facet, and sort. Fields that only need to be returned for display purposes can stay unindexed on disk, so the whole source document does not need to occupy the in-memory search index.

Every HA node holds a full replica. Raft automatically elects a new leader after a node failure, healthy nodes continue serving reads, and writes resume without operator promotion. That behavior and the full API surface are included in both the open-source server and Typesense Cloud.

Meilisearch can trade RAM for disk reads, with a performance cost

Meilisearch uses memory-mapped storage. It performs best when the full dataset fits in RAM, but an index can be larger than memory when the workload and storage can tolerate uncached page reads. Once data spills beyond RAM, cache misses require disk reads and performance depends more heavily on storage.

Enterprise Edition or Cloud replication can automatically serve a network search from another replica when a remote fails. Writes still pass through one static leader. If that leader fails, writes stop until an operator manually promotes another instance.

This creates a single point of failure, making Meilisearch unsuitable for critical production workloads that cannot tolerate write downtime or manual intervention.

PUBLIC PRICING EXAMPLES

How the Cloud pricing models differ

Standard single-node Meilisearch Cloud has public estimates, while its HA and Enterprise configurations require a quote from its sales team. Typesense publishes its full configuration catalog, including HA.

Swipe left to see Meilisearch

Pricing areaTypesenseMeilisearch
Self-hosting Free to self-host forever, including Raft high availability, analytics, JOINs, nested filters, and the complete search API Community Edition is free to self-host, but is feature restricted. Production sharding and replication require an Enterprise agreement.
Managed Cloud starting point For example, one Oregon node with 0.5 GB RAM, 2 vCPU, and a 1-hour burst window is $0.03 per hour, $21.60 per month, or $259.20 per year. View the full Typesense Cloud catalog Meilisearch Cloud starts at $20 per month. Its public resource estimator shows an XS example with 0.5 vCPU, 1 GB RAM, and 32 GiB of disk at about $23 per month, or $276 per year.
How cloud costs scale Dedicated clusters are billed hourly by RAM and CPU configuration, plus bandwidth, without per-search or per-record fees. The full single-node and HA catalog is public. Resize from the dashboard or automate changes through the Cloud Management API. Standard single-node usage and resource estimates are public. HA, sharding, dedicated resources, and other Enterprise capabilities require a custom sales quote.
High availability For example, the same 0.5 GB, 2 vCPU, 1-hour burst, Oregon configuration with 3-node HA is $0.12 per hour, $86.40 per month, or $1,036.80 per year. Replication and sharding require Enterprise Edition or Cloud. HA topology, dedicated resources, and related Enterprise pricing require a custom quote.

These are examples, not equivalent workload sizes. Meilisearch can trade RAM for disk reads, while Typesense sizes memory around indexed fields. Compare the smallest configuration on each service that meets the same latency, concurrency, indexing, and availability target with your data.

When Meilisearch is a better fit

  • The index is larger than the RAM budget, and the workload can accept memory-mapped reads from fast storage.
  • Schema-less ingestion matters more than enforcing consistent field types at write time.
  • The deployment fits within Community Edition's feature set, and a permissive MIT license matters more than production sharding or replication.
  • Relevance tuning calls for direct priority ordering of ranking rules and Meilisearch's bucket-sort model.
  • Semantic and hybrid search need its provider-specific embedders, custom REST support, DiskANN storage, or multiple named embedders in one index.
  • A hosted crawler should come bundled with the Cloud service.
  • A Rust codebase is a concrete preference for the engineers who will operate or contribute to the search engine.
  • The team is comfortable with capabilities that are experimental, Cloud-only, Enterprise-only, or governed by BUSL license terms.

When Typesense is a better fit

  • Your search use case requires high availability and cannot afford a single point of failure or manual operator intervention. Typesense uses Raft with automatic leader election to keep the cluster available through node failures.
  • One open-source server should provide the full API surface, including high availability, analytics, JOINs, nested filters, semantic search, RAG, and curation.
  • The data model needs flexible or strict schemas with ingestion-time type validation, plus mature nested-object filters and cross-collection JOINs.
  • Relevance and merchandising need runtime control, including per-query and per-field typo settings, query-time sorting, filters, curations, and scoped keys.
  • The search experience combines typo-tolerant keyword search, semantic and hybrid search, RAG, and Natural Language Search that builds structured search parameters from prose.
  • Records need several independent geo fields, or global users need nearest-node routing and automatic regional failover.
  • Pricing needs to be transparent before talking to sales. Typesense Cloud publishes its full single-node and HA configuration catalog, while Meilisearch gates HA and Enterprise pricing behind a sales conversation.

What teams ask before choosing Typesense or Meilisearch

The products overlap on fast application search, but the operational model and the maturity of advanced features can change the decision.

Is Meilisearch more memory efficient than Typesense?

It can be when the index is larger than RAM and the workload tolerates disk reads. Meilisearch uses memory-mapped LMDB, but once data spills beyond RAM, cache misses move work to disk and can increase latency. Typesense keeps the fields used for search, filters, facets, and sorting in memory, while unindexed source fields stay on disk. Benchmark the smallest configuration that meets the same latency and concurrency target instead of comparing nominal RAM.

Can both engines run highly available clusters?

Yes for Typesense, and with material caveats for Meilisearch. Typesense includes Raft clustering and automatic leader election in the open-source server, which avoids single points of failure for production search use cases. Meilisearch requires Enterprise Edition or Cloud for replication. Network searches can automatically fall back to another replica, but one static leader owns writes, and a leader failure blocks them until manual promotion.

Is Typesense schema-less like Meilisearch?

Not exactly, though it can feel that way. Typesense can infer field types automatically, then validates future records against them. It also supports strict schemas. Meilisearch accepts documents without a schema, including mixed types for the same field, which is convenient for ingestion but can produce surprising sort behavior.

Can both engines filter within nested arrays of objects?

Consider a single record that has an array-of-objects field called variants: [{ color: "red", size: "M" }, { color: "blue", size: "L" }]. A filter for red and L should not falsely match by taking the color from one variant and the size from another. Typesense binds both conditions to the same nested object. Meilisearch's precise path remodels variants as related records using experimental foreign keys. That model is unavailable with remote sharding, and a foreign filter returns an error when its related-side condition matches more than 100 documents.

Do both support semantic and hybrid search?

Yes. Meilisearch offers a broad embedder catalog and DiskANN. Typesense combines HNSW vector search with typo-tolerant keyword search, filters, and facets. It also adds Natural Language Search for structured query generation and a RAG API.

Can we keep an InstantSearch frontend?

Yes! Both products provide adapters for the InstantSearch ecosystem. The Typesense-InstantSearch adapter supports InstantSearch.js and the React, Vue, and Angular integrations, so much of the existing UI can stay in place.

How do we migrate from Meilisearch to Typesense?

Export from your source of truth, map Meilisearch index settings to a Typesense schema, bulk import, translate query parameters, and replay representative searches before cutting over. For large imports, process each record's response and retry when Typesense asks the client to slow down.

Does Rust versus C++ matter when choosing a search engine?

Not by itself. Meilisearch is written in Rust and Typesense in C++, but language alone is not a meaningful buyer-level differentiator. Compare architecture and behavior under the actual workload, including latency, reliability, relevance, operations, and cost. A team that plans to contribute deeply to the engine may still have a language preference.

Compare other search platforms

Read Typesense vs Algolia or Typesense vs Elasticsearch.

Let Your Workload Make the Decision

Bring your real dataset, queries, write rate, and latency target. Launch a managed cluster or run the same open-source Typesense engine yourself, then compare the smallest configuration that meets the bar.

Typesense CloudView on GitHub
background
NO PHD REQUIRED

Lightning-Fast
Open
Source
Search

Found 63 recipes out of 2,231,142 in ...ms
  1. Ham and Pineapple Pizza with Coconut Sauce
  2. Habanero and Pineapple Pizza
  3. Ham & Pineapple Pizza
  4. Ham and Pineapple Pizza
  5. Pineapple Pizza Burger
  6. Easy Pepperoni Pineapple Pizza Bread Burgers