[In Meilisearch v1.6](https://www.meilisearch.com/blog/meilisearch-1-6), we revolutionized search by combining the precision of full-text search with the intelligence of AI. This hybrid approach isn't just another feature—it's a fundamental reimagining of how search should work. Today, we'll pull back the curtain on our implementation and show why Meilisearch's unique approach to hybrid search delivers superior results.

Join us as we explore the mechanics of hybrid search, examine conventional implementation methods, and reveal how Meilisearch's innovative solution raises the bar for search relevance.

## What is hybrid search?

Search is always about finding the documents that are the most relevant to a user’s query.

Hybrid search is about combining the results of two complementary approaches to search: full-text search and semantic search.

[Full-text search](https://www.meilisearch.com/blog/how-full-text-search-engines-work) analyzes the words that appear in a query in a lexical manner, excelling at correcting typos and completing incomplete queries. It is also fast, and its results are easy to understand in retrospect.

[Semantic search](https://www.meilisearch.com/blog/semantic-search) leverages Large Language Model (LLM) technology to understand the meaning of words in the query. It excels at finding documents that contain synonyms to the user’s query and other similar concepts.

Both search approaches build a list of relevant documents according to their respective methods. Hybrid search combines these lists to produce a final list of results.

The golden question is, how to combine them?

![Diagram containing two lists of documents as shapes (squares for full-text, diamonds or triangles for semantic), and an arrow with an interrogation mark that produces a new list of documents containing both kinds of shapes](https://unable-actionable-car.media.strapiapp.com/how_to_merge_55aea77e87.png)

## The traditional way of mixing results is flawed

The traditional method of mixing the results from both lists is called [_Reciprocal Rank Fusion_ (RRF)](https://medium.com/@devalshah1619/mathematical-intuition-behind-reciprocal-rank-fusion-rrf-explained-in-2-mins-002df0cc5e2a), because it considers the **rank** of the documents in each list of results, and uses that as a basis to compute the final list of results.

In this context, the rank of the document refers to its position in the list of results, indicating how close it appears to the top.

For example, in a list of 20 documents, the result ranked first can be attributed 20 points. RRF can then attribute points to documents in each of the lists, sum the points for each document, and reorder the documents accordingly. Here’s a [detailed explanation](https://learn.microsoft.com/en-us/azure/search/hybrid-search-ranking) about an implementation of RRF on Azure AI Search.

However, this method of mixing results is flawed, because to function properly, it assumes that documents from all the lists of results are of comparable relevance: that is, all the documents in all lists are “good documents” that we want to return.

![Keyword search. Pikachu uses "imprecise query". It's super effective!](https://unable-actionable-car.media.strapiapp.com/pikachu_hybrid_search_f4f9e6c29f.jpg)

But, like Pokémons, each search method is vulnerable to different types of search queries. Covering for the weakness of the other method is the whole point of mixing them!

So let’s say you’re using a query that is fine for semantic search, but gives underwhelming results for keyword search. Using RRF, you might get some variant of the following:

![Illustrated example of the downfall of rank fusion](https://unable-actionable-car.media.strapiapp.com/RFF_merge_789ff4e216.png)

Recursive rank fusion is like wasting fine chocolate by mixing it with spoiled milk!

To address this issue, we require additional information about the documents listed in our results.

## Ranking is not enough: Scoring documents

The main idea is that you can score _how_ relevant a document is from 0 (not relevant at all) to 1 ("perfect” match), independently of its rank in the list.

Naturally, documents closer to the top of the list are more relevant, but the first document might be anywhere between 0 and 1. Then, you compare the scores in both lists and rank the documents across both lists that way.

Fortunately, Meilisearch already scores documents for full-text search (for more details about this, refer to [our previous article about scoring](https://www.meilisearch.com/blog/from-ranking-to-scoring)), and semantic search natively ranks documents depending on how similar they are to the query (although the details are complicated).

So we have a score for both full-text search and semantic search, but can we compare them?

## Making the scores comparable

When you have two scores ranging from 0.0 to 1.0, it can be tempting to compare them, but for this to work, similarly relevant documents must be scored similarly with both methods.

One difficulty that arises when performing this comparison is that the models used in semantic search tend to use only a portion of the full vector space they could use, resulting in vectors that have a similarity of at least 0.5 even for texts that appear entirely dissimilar. Furthermore, even very similar texts might be capped in similarity around 0.7, which will cause them to rank lower than good full-text results.

Fortunately, Meilisearch provides the tools to solve that particular problem. Meilisearch allows you to perform an affine transformation to the scores of the semantic search, so that the scores can be offset by the observed mean similarity between vectors, and the results that are packed around that mean similarity can be spread more in the 0.0, 1.0 space. For more details, refer to [the documentation on the](https://www.meilisearch.com/docs/reference/api/settings#distribution) `distribution` [field](https://www.meilisearch.com/docs/reference/api/settings#distribution).

For models known to Meilisearch (such as OpenAI models), a suitable default distribution is already applied to configured embedders, so users don’t have to configure it themselves!

With this operation, our scores are comparable, and depending on the query, the method that produces the most relevant documents will be favored by the hybrid search

![Failing example, running with Meilisearch, yields better results](https://unable-actionable-car.media.strapiapp.com/meili_merge_9bb48a7b7d.png)

## What else goes into making hybrid search right?

Scoring documents is a decisive relevancy advantage of Meilisearch over the traditional approach based on Reciprocal Rank Fusion, but there are a lot more aspects involved in making hybrid search right:

- We developed our bespoke vector store, [arroy](https://github.com/meilisearch/arroy), to leverage our database backend (LMDB) and meet our performance target. Arroy supports Approximate Nearest Neighbor search, multiple distances, native document filters, and binary quantization. Arroy doesn’t block a reader while writing or a writer while reading, and allows for incremental updates.
- We kept flexibility in mind with the mantra of making easy things simple, and hard things possible: users that are new to hybrid search can quickly setup an embedder with our default provider and configuration, but users can also carefully optimize their embedding process with multiple supported provider (or even the ability to provide your own with [REST embedders](https://www.meilisearch.com/docs/learn/ai_powered_search/configure_rest_embedder) and [user-provided embeddings](https://www.meilisearch.com/docs/learn/ai_powered_search/search_with_user_provided_embeddings)) and [document templates](%5B%3Chttps://www.meilisearch.com/docs/learn/ai_powered_search/document_template_best_practices%3E%5D(%3Chttps://www.meilisearch.com/docs/learn/ai_powered_search/document_template_best_practices#sample-document%3E)) to customize how documents are turned into text to embed.
- Best-in-class full-text search engine, seamlessly integrated with semantic search, with support for efficient and expressive [sorting and filtering](https://www.meilisearch.com/docs/learn/filtering_and_sorting/working_with_dates) of documents.

Search is foundational; it needs to be rock solid to reliably serve the needs of your application. Meilisearch’s hybrid search implementation scores documents to merge results while preserving their relevancy. Scores also power the [federated search](https://www.meilisearch.com/docs/learn/multi_search/multi_search_vs_federated_search#what-is-federated-search%3F), a generalized version of semantic search that unifies results from multiple lists of intermediate hits.

Our users use federated search to perform hybrid searches across multiple vector stores, frequently combining image and text search requests.

## Why Meilisearch's hybrid search stands apart

Meilisearch's innovative approach to hybrid search goes beyond simply combining full-text and semantic search results.

By implementing a sophisticated scoring system instead of traditional rank fusion, carefully calibrating score comparability, and building a robust technical foundation with arroy, we've created a hybrid search solution that consistently delivers more relevant results.

This thoughtful implementation ensures that the strengths of each search method are preserved while compensating for their weaknesses, setting a new standard for search relevance in modern applications.

Why traditional hybrid search falls short (and how we fixed it)

When we discovered that 70% of our team was already using AI tools daily—but with wildly varying results—we knew we needed a systematic approach. Not another vague "AI transformation" initiative, but a practical framework for separating AI hype from real value.

Here's the concrete system we developed to evaluate and implement AI tools across our company and what we've learned so far.

## The surprising reality we uncovered

When we began our AI journey at Meilisearch, we expected to find pockets of adoption. Instead, we discovered something more interesting: **universal experimentation but inconsistent value**.

### The adoption paradox

Our company-wide assessment revealed that every single team member had already experimented with AI tools, with an impressive 70% incorporating them into daily workflows.

![image.png](https://unable-actionable-car.media.strapiapp.com/image_40601b4a7c.png)

![image2.png](https://unable-actionable-car.media.strapiapp.com/image2_a867a1c935.png)

But the numbers told only half the story:

Despite high adoption rates, team sentiment about AI's value showed surprising variation. Many were using the tools but questioning whether they were truly beneficial or just creating a new kind of busywork.

![image3.png](https://unable-actionable-car.media.strapiapp.com/image3_67ad5244d9.png)

![image4.png](https://unable-actionable-car.media.strapiapp.com/image4_4a6d31dcca.png)

This gap between usage and perceived value became our first critical insight.

### Four critical insights that changed our approach

<Callout intent="note">

**Previously Shared Insights:**

- 🧠 [𝐃𝐨𝐞𝐬 𝐀𝐈 𝐛𝐞𝐜𝐨𝐦𝐞 𝐥𝐞𝐬𝐬 𝐯𝐚𝐥𝐮𝐚𝐛𝐥𝐞 𝐚𝐬 𝐲𝐨𝐮𝐫 𝐞𝐱𝐩𝐞𝐫𝐭𝐢𝐬𝐞 𝐠𝐫𝐨𝐰𝐬?](https://www.linkedin.com/posts/gillianmcauliffe_aitransformation-professionalgrowth-futureofwork-activity-7292477225889660928-7UXm/?utm_source=share&utm_medium=member_desktop&rcm=ACoAAAXkmQUBijf17cl9IKQZhMWLN67r9deqvAc) How your relationship with AI changes at different levels of expertise.
- 🔍 [**Unlocking Our AI Knowledge-Sharing Blueprint:**](https://www.linkedin.com/posts/gillianmcauliffe_aitransformation-futureofwork-knowledgesharing-activity-7290061146634485762-fm1D/?utm_source=share&utm_medium=member_desktop&rcm=ACoAAAXkmQUBijf17cl9IKQZhMWLN67r9deqvAc) Understanding what people want to know to get started.
- 🤝 [Preserving Human Connection in an AI-Enhanced Workplace](https://www.linkedin.com/posts/gillianmcauliffe_aitransformation-teamwork-futureofwork-activity-7287036478092632066-boTG/?utm_source=share&utm_medium=member_desktop&rcm=ACoAAAXkmQUBijf17cl9IKQZhMWLN67r9deqvAc): Will we lose what's special about our teams by bringing in AI?

</Callout>

Through in-depth interviews across all departments, we uncovered patterns that wouldn't have appeared in simple usage statistics:

1. 🌿 **Environmental impact:** Our teams have raised thoughtful concerns about the environmental impact of AI.
2. **🤔 The expertise paradox:** As team members' domain expertise increases, their perception of AI's value often follows a U-curve. Beginners find AI helpful for learning, experts leverage it as a powerful accelerator, but those in the middle sometimes find it more hindrance than help.
3. **🎯 The trust dilemma:** AI's "inconsistent brilliance" creates a fundamental trust issue—sometimes transformative, sometimes bafflingly off-target—making teams reluctant to integrate it into mission-critical processes.
4. **✨ Creative acceleration:** Despite skepticism, nearly everyone reported dramatic improvements in creative processes—from eliminating writer's block to transforming editing workflows.
5. 🔭 **Prompt sophistication gap:** Most team members were still using basic prompting techniques, unaware of how dramatically their results could improve with more advanced approaches.

This investigation revealed we weren't facing an adoption problem but a **value extraction challenge**. People were using AI tools but were not consistently getting meaningful results. We needed a framework to help teams identify where AI could actually create value—and where it was just adding complexity.

## ⚖️ Our solution: the AI value scoring matrix

Based on these insights, we developed a systematic framework to cut through the AI hype cycle. Rather than chasing the latest tools or models, we created a data-driven method to identify where AI could create genuine value for our specific context.

**The value scoring matrix:** Each potential AI use case is evaluated across multiple weighted dimensions aligned with our organizational priorities. This transforms the abstract question of "where should we use AI?" into a concrete, comparable set of scores.

**Our Six Critical Dimensions:**

| **Dimension** | **Weight** | **What We Assess** | **Why It Matters** |
|--------------------|------------|--------------------------------------------------------------------|----------------------------------------------------------------------------------|
| Feasibility | 20% | Implementation difficulty with current skills/resources | Prevents us from chasing technically impressive but impractical applications |
| Revenue Impact | 25% | Potential to increase top-line growth | Ensures AI initiatives contribute to business success |
| Data Readiness | 15% | Availability and quality of necessary data | Many AI projects fail due to data issues, not technology limitations |
| Time to Value | 15% | Speed to tangible results | Builds momentum with quick wins |
| Risk | 10% | Potential challenges or downsides | Protects against unintended consequences |
| Internal Impact | 15% | Number of people/teams benefiting | Prioritizes broad improvements over narrow optimizations |

Here's how we applied this framework to evaluate potential AI initiatives:

This systematic approach cut through subjective debates about which AI tools to adopt, replacing them with evidence-based decisions that aligned with our strategic priorities.

## Early results: beyond the AI hype cycle

Our scoring matrix has already transformed how we approach AI at Meilisearch:

### 1\. From endless options to strategic focus

Instead of drowning in the constant stream of new AI products, our teams now have a clear filter: "Does this tool address one of our high-scoring use cases?" This has dramatically reduced distractions and FOMO while increasing meaningful adoption.

### 2\. Balanced investment: "Ferraris" and "cargo ships"

We're strategically balancing our implementation:

- **"Ferrari" projects:** Quick wins that generate excitement and momentum (e.g., automating our internal documentation updates)
- **"Cargo ship" initiatives:** Substantial efforts with transformative potential (e.g., reimagining our customer support workflow with AI assistance)

### 3\. Creating a dedicated space for experimentation

One crucial lesson: AI adoption requires protected time. We've integrated dedicated "AI exploration blocks" into our work schedules and upcoming offsite, focusing specifically on our highest-scoring use cases.

### 4\. Addressing real concerns

The scoring matrix naturally incorporates ethical considerations through the risk dimension. We've operationalized this by:

- Favoring smaller, more efficient models when possible
- Prioritizing open-source options that allow us to run models locally
- Building in time for human review of AI-generated content

## The framework in action: a case study

After applying our scoring matrix to use cases identified through internal discussions, we quickly surfaced high-impact opportunities. One standout example was improving our code review process—specifically, reducing time to first review for pull requests (PRs).

This use case scored highly across multiple dimensions:

- High feasibility with existing tools
- Direct impact on development velocity
- Clear time-to-value metrics
- Broad internal impact across engineering teams

We implemented CodeRabbit, an AI-powered code review assistant that provides immediate PR feedback. While maintaining our essential human review process, the tool identifies common issues and provides automated PR descriptions. After a two-week trial period, we measured both quantitative metrics and gathered team feedback.

The results validated our scoring framework's prediction:

- Decreased average time to first review
- Improved PR documentation quality
- Positive team adoption and engagement
- Cumulative time savings across the engineering organization

## What's next: your turn

We're sharing this framework because we believe it can help other organizations cut through AI hype and focus on value creation.

The key to successful AI adoption isn't about having the fanciest tools—it's about having the clearest framework for identifying where AI truly adds value in your specific context.

🚀 How we're making AI work at Meilisearch

## What is RAG, and why should you care?

Every day, millions of people interact with ChatGPT, Claude, and other large language models, fundamentally changing how we interact with computers. The appeal is obvious: instead of learning specific interfaces, users can simply communicate in natural language. The machine adapts to the human, not the other way around.

This shift has organizations scrambling to create similar experiences with their own data. Whether it's building a support chatbot that understands product documentation, crafting an assistant that writes personalized emails, or developing an internal tool to help employees access company knowledge, they all require the same foundation: giving LLMs access to proprietary information.

Enter Retrieval Augmented Generation (RAG), a workflow that's deceptively simple in concept:

1. User asks a question
2. The system retrieves relevant documents based on that question
3. LLM generates an answer using those retrieved documents as context

When implemented correctly, RAG provides more accurate, up-to-date responses while reducing hallucinations. But here's where most engineers go wrong: they default to vector databases without considering whether that's actually the optimal approach.

## The engineer's shortcut: Vector databases

The typical RAG implementation looks something like this:

```
User Question → Convert to Embedding → Query Vector DB → Retrieve Similar Documents → Generate Answer

```

It makes intuitive sense. Vector embeddings capture semantic meaning, so similar concepts cluster together in the vector space. By converting questions and documents to the same embedding space, you can find relevant information even when the exact keywords don't match.

Engineers love this approach because it seems elegant. You don't need complex query parsing or keyword matching – just transform everything to vectors and let similarity algorithms do the work.

But this shortcut creates two critical problems that significantly degrade the quality of the results.

## Problem 1: Unrefined queries lead to irrelevant results

The first major issue occurs at the initial step: we take the raw user query, convert it directly to an embedding, and use that to retrieve documents.

Think about what happens when you use Google. You don't always search with your actual question – you distill it to key terms that are more likely to yield relevant results. If you want to know "What's the best way to organize React components in a large project?", you might search for "React component organization best practices large applications."

LLMs excel at this kind of query refinement, but in the standard vector DB approach, we skip this crucial step. The embedding captures the semantic meaning of the original question, but that doesn't always align with the optimal search query.

## Problem 2: Vector search sacrifices precision for recall

The second problem is more fundamental: vector similarity search is strong at recall but often weaker at precision compared to traditional full-text search.

Vector search excels at finding conceptually related content (recall) - documents that discuss similar topics, even if they use different terminology. However, it's less effective at pinpointing exactly what you need (precision).

Full-text search, on the other hand, prioritizes precision. When exact keyword matches exist, they're usually highly relevant. The trade-off is that full-text search might miss conceptually similar content that uses different terminology.

This explains why many RAG systems that rely exclusively on vector databases end up implementing additional re-ranking steps after initial retrieval. It's also why vector databases increasingly add full-text capabilities – they recognize the inherent limitations of pure vector search.

## A more human approach to RAG

Since we're trying to build systems that mimic human intelligence, shouldn't we structure our RAG pipelines to mirror how humans actually search for information?

When people need to find information, they:

1. Reframe their question into effective search terms
2. Use search engines optimized for retrieving relevant information
3. Scan results to identify the most useful content

A better RAG design would follow this pattern:

```
User Question → LLM Refines Into Search Query → Hybrid Search (Combining Full-text and Semantic) → Retrieve Relevant Documents → LLM Generates Answer

```

This approach addresses both problems:

1. The LLM refines the user's original question into an optimized search query
2. Hybrid search combines the precision of full-text search with the recall of semantic search

## The solution: Simpler (and better) RAG with search engines

Here's my contrarian take: instead of jumping to vector databases for RAG, start with a quality search engine that combines full-text and semantic capabilities.

Modern search engines like Meilisearch offer hybrid search functionality that delivers better overall relevance than pure vector retrieval. They're also:

- Easier to deploy and maintain
- Optimized for fast queries at scale
- Designed with relevance as the primary objective
- More intuitive for debugging search results

For most RAG applications, the workflow becomes much simpler:

1. The user submits a question
2. LLM transforms that question into an effective search query
3. The search engine retrieves relevant documents using hybrid search
4. LLM generates a response using those documents as context

This approach aligns with how humans actually search for information, leveraging the strengths of both LLMs and search technology while avoiding unnecessary complexity.

## The power of simplicity in RAG: Back to search basics

Vector databases certainly have their place in the machine learning ecosystem, but they're not always the best foundation for RAG systems. By taking a step back and considering how humans search for information, we can build more effective, simpler RAG architectures.

The next time you're designing a RAG system, consider whether a good search engine might be the better choice. Your users (and your future self maintaining the system) will thank you.

Why you shouldn't use vector databases for RAG

AI-powered search: What you need to know [2025]

The 10 best AI enterprise search tools and platforms [2025]

When [Bookshop.org](http://Bookshop.org)'s small engineering team struggled to push past a 14% search conversion rate, they faced a familiar dilemma. Despite having Elasticsearch—a powerful but complex search solution—their eight-person team was spending more time maintaining search infrastructure than improving their core platform. They needed something different: a solution that could deliver better results with less overhead.

Their challenge reflects a broader reality in modern software development. Every click, every search query, and every page load directly impact business metrics, yet the tools meant to optimize these experiences often become bottlenecks themselves. For engineering teams, especially smaller ones, the complexity of search infrastructure can become a _hidden tax_ on innovation and growth.

This is where Meilisearch comes in. Built to work seamlessly out of the box, it's the solution that lets developers focus on building rather than maintaining. Today, we're taking that mission further with [the launch of Meilisearch AI](https://www.meilisearch.com/launch-week)—our next step in making it radically easier for developers to build world-class search experiences.

In this article, we'll explore how modern search has evolved beyond simple queries, examine the technical challenges facing development teams, and introduce Meilisearch's innovative solutions—including our new AI capabilities. You'll learn how our approach to developer experience is transforming search implementation, and discover our roadmap for making advanced search features accessible to teams of all sizes.

## Redefining modern search

Modern search experiences extend far beyond the traditional search bar, encompassing two key facets: search and discovery. 

Search can take the appearance of inputs, menus, filters, or the trending UX of conversational search (i.e. chatbots). Discovery, on the other hand, plays a key role in navigating content and suggesting recommendations.

Across industries, we see search and discovery most prevalent as follows:

- **Traditional full-text search & semantic search:** standard in all industries ranging from SaaS to ecommerce, marketplaces, and content platforms
- **Conversational search:** prominent in RAG (Retrieval-Augmented Generation) and AI-enhanced search interfaces like Perplexity
- **Discovery for navigation:** common in social media and content platforms where a feed is the primary navigation
- **Recommendations:** tie in more critically with business value in ecommerce and content platforms

For early-stage startups and SMBs, implementing these search capabilities presents significant technical challenges. The complexity multiplies when adding AI integration and personalization layers. Addressing these pain points diverts the workforce from building core product value. Meilisearch is here to relieve you from this burden, allowing you to focus on what really matters. 

We want Meilisearch built-ins to handle modern use cases like [document embedding](https://www.meilisearch.com/docs/learn/ai_powered_search/document_template_best_practices). Complex setups—such as managing [heavyweight solutions like Elasticsearch](https://www.meilisearch.com/blog/meilisearch-vs-elasticsearch), [tweaking a Postgres instance](https://www.meilisearch.com/blog/postgres-full-text-search-limitations), or handling hybrid search—are additional integration burdens. This ignores the security consideration related to exposing front-end accessible APIs or the analytics necessary to collect business KPIs—things that don’t come packaged with custom-made solutions.

![bookshop_meilisearch.gif](https://unable-actionable-car.media.strapiapp.com/bookshop_meilisearch_9498082583.gif)

These DX considerations impact business metrics down the road as engineering bottlenecks lead to longer time-to-market, hindering growth for companies that want to move fast. In ecommerce, for example, search accuracy directly translates to conversion rate, as shown by how we helped [Bookshop increase search-based purchases by 43%](https://www.meilisearch.com/blog/bookshop-increases-search-based-purchases). On SaaS and content platforms, discovery enables engagement and adoption growth, both of which are key to keeping churn in check. 

## Leveraging DX to move mountains

Meilisearch is a no-configuration, plug-and-play solution that you can install and forget. At the heart of any search tool—and often the biggest challenge for developers—is ensuring relevant results. **CascadeRank** is our answer to [relevancy fine-tuning](https://www.meilisearch.com/docs/learn/relevancy/relevancy). Our algorithm uses human-readable ranking rules that enable configuration and easy customization. With Meilisearch, simplicity is key, allowing you to set up and focus on your core tasks. 

Though customization is easy, Meilisearch comes with strong defaults to handle common requirements out of the box—things like being typo-tolerant and prioritizing word proximity. Indeed, we’ve seen [Minipouce get better relevancy with Meilisearch’s default configuration than their previous Typesense setup](https://www.meilisearch.com/blog/minipouce-top-grade-customer-experience), leading to a 5x search volume increase. 

Scalability is a concern for small teams and enterprises alike. Performance is the name of the game for user-facing search—both for search latency and indexing time. As demonstrated by [Podscan.fm](http://Podscan.fm) [searching across 600GB of data](https://x.com/arvidkahl/status/1807889668748275802), search remains fast against large datasets. Our [next-generation indexer boasts 4x faster updates](https://www.meilisearch.com/blog/introducing-indexer-2024), making Meilisearch suitable at scale for write-intensive workflows, too.

[![arvid-kahl-tweet.png](https://unable-actionable-car.media.strapiapp.com/arvid_kahl_tweet_0c41da4824.png)
](https://x.com/arvidkahl/status/1807889668748275802)

When it comes to security, no compromise is possible. So Meilisearch provides fundamental security features to handle simple and complex use cases. Meilisearch facilitates [API key management](https://www.meilisearch.com/docs/learn/security/basic_security) to facilitate multi-environments. The built-in [multi-tenancy features](https://www.meilisearch.com/docs/learn/security/multitenancy_tenant_tokens) also enable more advanced security patterns, allowing developers to customize their access rules.

Using our expertise in search requirements across various customers, we built [Meilisearch Cloud](https://www.meilisearch.com/cloud). It’s the quickest way to get started. In addition to freeing teams from managing infrastructure, it integrates analytics and monitoring, making it [the preferred production platform for companies like HitPay](https://www.meilisearch.com/blog/hitpay-streamlines-search). Open-source Meilisearch can be self-hosted via Docker and a plethora of hosting platforms. 

This feature set positions Meilisearch well as a simple solution to get the best of modern search experience. This includes benefiting from an easy LLM integration to leverage AI-powered solutions.

## Future-proofing your search strategy

With native [hybrid and vector search](https://www.meilisearch.com/docs/learn/ai_powered_search/getting_started_with_ai_search) integration, Meilisearch makes it simple to implement AI-powered search capabilities. But we’re not stopping there—our 2025 roadmap is packed with features enabling businesses to uncover more value from AI while sustaining performance and scalability.

### Improved conversions with personalization

[Our research shows that personalization significantly improves conversion rates](https://www.meilisearch.com/blog/personalization-recommendations) across diverse business types. We aim to release a personalization framework enabling the collection of end-user context on each request. This comprehensive context management system allows developers to leverage user preferences, recent interactions, and behavioral patterns to further customize the relevancy of results on a per-user basis. 

### Retrieval augmented generation without overhead 

Although Meilisearch already enables [building retrieval-augmented generation (RAG)](https://www.meilisearch.com/blog/how-to-build-rag), we want to integrate built-in RAG capabilities in the engine, seamlessly combining full-text search, vector search, and LLM interaction. Then, we want to tackle inherent problems in state-of-the-art RAG applications by tackling the orchestration of filtering, ranking, and results merging in RAG scenarios. Of course, we want this to come with strong defaults, making it easy for developers to get started.

### Advanced ranking integrations

Our goal is to provide a native way for users to leverage AI reranking models to improve the relevancy of results. This provides a more flexible approach to configuring result relevancy by allowing real-time results reranking based on user context and behavior. We also want to enable organizations to integrate with their own custom ranking models to allow for further dynamic adjustment of results based on hand-picked factors.

### Reliability as you scale

We also plan to improve Meilisearch’s infrastructure to handle additional use cases better. In particular, we want to continue improving the distributed system capabilities for better scaling and high availability by building upon our sharding experiments with [federated remote search](https://www.meilisearch.com/docs/learn/multi_search/implement_sharding). Another layer of performance is integrating more sophisticated caching mechanisms to improve out-of-the-box performance. Monitoring and observability are also continuously improved based on user feedback.

## Meilisearch AI, refreshingly simple search DX

The release of Meilisearch AI marks a new age for Meilisearch users, fully integrating with the modern landscape of AI-powered applications. But it’s also a continuous fulfillment of our core value—making good search UX easy to build. We believe focusing on developer experience enables engineers to achieve faster time-to-value and maintain focus on their core business.

Meilisearch ships frequently to keep packaging built-in features, enabling fast-shipping companies to do what they do best. For more information on the Meilisearch AI release, consult our [launch week page](https://www.meilisearch.com/launch-week). 

Start with a [14-day free trial of Meilisearch Cloud](https://cloud.meilisearch.com/) or [request a demo](https://dub.sh/meili-sales) to discuss your needs with our search experts.

Building the future of search with Meilisearch AI

What are vector embeddings? A complete guide [2025]

Semantic search transforms how we find information online by focusing on understanding the meaning behind the words and phrases in user searches rather than relying solely on word-for-word matching.

Using machine learning (ML) and natural language processing (NLP), it decodes intent, context, and relationships between words, thus allowing for more conversational queries. This makes it a vital tool for businesses needing fast, accurate, and relevant results from large or complex data sets.

Embedding models are ML techniques that render words and phrases into complex numerical representations, which are then classified based on their context and relationships. However, not all models are the same. Accuracy, speed, indexing efficiency, and pricing can vary significantly between them.

Choosing the right semantic search model is crucial for delivering precise, efficient, and scalable search experiences for specific tasks.

Understanding the trade-offs—whether for multilingual support, search latency, or computational efficiency—will help you implement a high-performing semantic search system for your brand.

## What is semantic search?

Semantic search is an advanced AI-powered approach to information retrieval that focuses on NLP, ML, and knowledge representation to deduce the intent and contextual meaning of user queries.

Traditional keyword-based search engines match exact strings of words, but that’s not always enough. Semantic search goes further by incorporating **named entity recognition**, **relationships between words**, and **contextual disambiguation** to deliver more relevant results.

Semantic search engines can understand synonyms, paraphrased queries, and even infer implicit meanings exclusively because they use deep learning models such as transformers.

Given its ability to enhance the search criteria, [AI-powered search](https://www.meilisearch.com/blog/ai-powered-search) is crucial for applications requiring faster, broader, and even less structured [data retrieval](https://www.meilisearch.com/blog/how-to-cache-semantic-search).

## Why is semantic search important?
Semantic search is important because it improves search capability and relevance, thus helping users find information faster and with fewer irrelevant results. It also reduces uncertainty by interpreting natural language queries more clearly.

Like similarity search, it allows users to phrase their questions conversationally and intuitively rather than using specific—sometimes even long-tail—keywords. Think of it like this: if you can ask a question from your neighbor on a bus, then you can ask it with a quick semantic search.

Additionally, businesses benefit from semantic search by **enhancing customer experiences**, **improving product discovery**, and **optimizing knowledge management systems**.

For example, e-commerce stores like Bookshop.org can [boost search-based sales by up to 43%](https://www.meilisearch.com/blog/bookshop-increases-search-based-purchases) by implementing semantic search tools like Meilisearch.

A core component of semantic search is embedding models through which similarity computations are made possible.

## What are embedding models?

Embedding models are machine learning models that transform words, phrases, or documents into dense numerical representations called embeddings.

These vector representations encode word relationships, which allow search engines to compare semantic meaning and context rather than rely solely on exact similarity between words.

For instance, embeddings generated for words like “**phone**,” “**mobile**,” and “**cell**” will be more closely aligned in vector space than, say, “**cell**” and “**bell**,” even though the latter two are closer in letter-by-letter similarity.

This allows a search engine to retrieve relevant results even if the exact keyword is absent, underpinning various NLP and LLM (large language model) applications.

### What’s the difference between semantic and search embeddings?

Semantic and search embeddings serve distinct purposes in NLP applications, particularly in semantic search.

**Semantic embeddings** capture the meaning and relationships between words, which is helpful for tasks like document classification, recommendation systems, language translation, and sentiment analysis. They do this by reflecting the words’ conceptual similarities, even if they are not used in the same context.

On the other hand, **search embeddings** are specifically optimized for retrieval tasks, thus ensuring queries and indexed documents align effectively in vector space to maximize search relevance.

Unlike generic semantic embeddings, search embeddings often incorporate domain- and background-specific optimizations to fine-tune the effectiveness of the specific retrieval activity.

For example, a semantic embedding model might learn that "mac" is related to pasta, the makeup company, and the Apple laptop computer. However, a _search embedding model_ trained on computers and hardware would prioritize the third meaning when processing search queries in that domain.

Embedding models serve as the backbone of semantic search by enabling similarity-based retrieval mechanisms that align with user intent.

## What is the role of embedding models in semantic search?

Embedding models power semantic search by converting text into a structured format that can be efficiently indexed and retrieved.

Instead of relying on exact keyword matches, these models compare **query** embeddings with **indexed** document embeddings. This enables nearest-neighbor searches and significantly improves retrieval accuracy over traditional algorithms that depend on keyword matching.

These models often use transformer-based architectures such as BERT (Bidirectional Encoder Representations from Transformers), GPT (Generative Pre-trained Transformer), and their derivative neural networks to capture context-aware representations.

That’s how they can handle nuanced queries at scale. For instance, even though “cat” and “bat” sound (and look) similar, only one of them may be used in context as a conventional house pet.

Several fine-tuned and pre-trained models continue to emerge as industry standards for semantic search. Each improves precision, relevance, efficiency, and/or scalability in its own way.

## What are the most commonly used models for semantic search?

Different embedding models vary in terms of vector dimensionality, context length, and certain other performance characteristics.

In the context of LLMs, dimensionality points to the number of components in a vector, with each corresponding to an attribute of the encoded variable. Additionally, context length relates to the amount of text—usually measured in tokens—a model can “remember” and refer to at any given time.

To assess these variations, a series of benchmark tests were conducted using [Meilisearch](https://www.meilisearch.com), evaluating each model's effectiveness in real-world search scenarios that you will encounter.

These tests measured factors such as **retrieval accuracy**, **indexing speed**, and **query latency**, which is how they evaluated each model’s performance under specific search conditions.

| **Model/Service** | **Dimensions** | **Context Length** |
| --- | --- | --- |
| [Cohere embed-english-v3.0](https://docs.cohere.com/docs/embeddings) | 1024 | 512 |
| [Cohere embed-english-light-v3.0](https://docs.cohere.com/docs/embeddings) | 384 | 512 |
| [Cohere embed-multilingual-v3.0](https://docs.cohere.com/docs/embeddings) | 1024 | 512 |
| [Cohere embed-multilingual-light-v3.0](https://docs.cohere.com/docs/embeddings) | 384 | 512 |
| [OpenAI text-embedding-3-small](https://platform.openai.com/docs/guides/embeddings) | 1536 | 8192 |
| [OpenAI text-embedding-3-large](https://platform.openai.com/docs/guides/embeddings) | 3072 | 8192 |
| [Mistral](https://docs.mistral.ai/capabilities/embeddings/) | 1024 | 8192 |
| [VoyageAI voyage-2](https://www.voyageai.com/) | 1024 | 4000 |
| [VoyageAI voyage-large-2](https://www.voyageai.com/) | 1536 | 16000 |
| [VoyageAI voyage-multilingual-2](https://www.voyageai.com/) | 1024 | 32000 |
| [Jina Colbert v2](https://jina.ai/news/jina-colbert-v2-multilingual-late-interaction-retriever-for-embedding-and-reranking/) | 128, 96, or 64 | 8192 |
| [OSS all-MiniLM-L6-v2](https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2) | 384 | 512 |
| [OSS bge-small-en-v1.5](https://huggingface.co/BAAI/bge-small-en-v1.5) | 1024 | 512 |
| [OSS bge-large-en-v1.5](https://huggingface.co/BAAI/bge-large-en-v1.5) | 1536 | 512 |

Your choice of a model for semantic search will depend on a variety of factors, including, but not limited to, its **accuracy**, **computational efficiency**, and **cost**.

## What factors should be considered when choosing the best model for semantic search?

### 1. Results relevancy
Relevance is the cornerstone of effective semantic search, especially in achieving optimal user experience. The right model should balance precision, recall, and speed to ensure users receive highly relevant results without excessive noise.

This balance becomes especially important when comparing hybrid approaches like vector and full-text search. When selecting an embedding model, think about the following:

- multilingual support;
- handling multi-modal data;
- domain-specific performance.

In this context, bigger might not always mean better. While larger models often provide better accuracy, smaller models can offer competitive results at lower computational costs.

Moreover, effectively structuring data, such as with optimized document templates in Meilisearch, can improve search quality.

### 2. Search performance
Search latency is a critical factor in user experience. Search-as-you-type has become the standard for customer-facing applications as fast, responsive search results improve user engagement and retention. 

Local embedding models are ideal to achieve lightning-fast performance, as they eliminate the need for round trips to external services and reduce latency. If you must rely on remote models, hosting your search service close to the embedding service can minimize delays and improve user experience.

The table below showcases latency benchmarks for various local embedding models and embedding APIs. All requests originate from a Meilisearch instance hosted in the AWS London data center.

| **Model/Service** | **Latency** |
| --- | --- | 
| [Cloudflare bge-small-en-v1.5](https://huggingface.co/BAAI/bge-small-en-v1.5) | ±800ms |
| [Cloudflare bge-large-en-v1.5](https://huggingface.co/BAAI/bge-large-en-v1.5) | ±500ms |
| [Cohere embed-english-v3.0](https://docs.cohere.com/docs/embeddings) | ±170ms |
| [Cohere embed-english-light-v3.0](https://docs.cohere.com/docs/embeddings) | ±160ms |
| [Local gte-small](https://huggingface.co/models?search=gte-small) | ±20ms |
| [Local all-MiniLM-L6-v2](https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2) | ±10ms |
| [Local bge-small-en-v1.5](https://huggingface.co/BAAI/bge-small-en-v1.5) | ±20ms |
| [Local bge-large-en-v1.5](https://huggingface.co/BAAI/bge-large-en-v1.5) | ±60ms |
| [Mistral](https://docs.mistral.ai/capabilities/embeddings/) | ±200ms |
| [Jina colbert](https://huggingface.co/jinaai/jina-colbert-v2) | ±400ms |
| [OpenAI text-embedding-3-small](https://platform.openai.com/docs/guides/embeddings) | ±460ms |
| [OpenAI text-embedding-3-large](https://platform.openai.com/docs/guides/embeddings) | ±750ms |
| [VoyageAI voyage-2](https://www.voyageai.com/) | ±350ms |
| [VoyageAI voyage-large-2](https://www.voyageai.com/) | ±400ms |

Benchmarks conducted in Meilisearch highlight the stark differences in latency across various models, with local models achieving response times as low as 10ms while some cloud-based services reach 800ms.

### 3. Indexing performance

Efficient indexing is another critical factor for the scalability of your search solution. As expected, the time required to process and store embeddings varies widely among models. Notable metrics influencing processing time are API rate limits, batch processing capabilities, and model dimensions.

Local models without GPUs may experience slower indexing due to limited processing power, while third-party services vary in speed based on their infrastructure and agreements.

As mentioned earlier, minimizing data travel time between your application and the model can reduce latency and optimize indexing. Evaluating these factors ensures your chosen model and service can effectively meet your needs.

The benchmark below compares the indexing performance of a collection of 10k e-commerce documents (with [automatic embedding generation](https://www.meilisearch.com/docs/learn/experimental/vector_search?utm_campaign=vector-search&utm_source=blog&utm_medium=models-comparison#generate-auto-embeddings)).

| **Model/Service** | **Indexation time** | 
| --- | --- | 
| [Cohere embed-english-v3.0](https://docs.cohere.com/docs/embeddings) | 43s |
| [Cohere embed-english-light-v3.0](https://docs.cohere.com/docs/embeddings) | 16s |
| [OpenAI text-embedding-3-small](https://platform.openai.com/docs/guides/embeddings) | 95s |
| [OpenAI text-embedding-3-large](https://platform.openai.com/docs/guides/embeddings) | 151s |
| [Cloudflare bge-small-en-v1.5](https://huggingface.co/BAAI/bge-small-en-v1.5) | 152s |
| [Cloudflare bge-large-en-v1.5](https://huggingface.co/BAAI/bge-large-en-v1.5) | 159s |
| [Jina Colbert V2](https://huggingface.co/jinaai/jina-colbert-v2) | 375s |
| [VoyageAI voyage-large-2](https://www.voyageai.com/) | 409s |
| [Mistral](https://docs.mistral.ai/capabilities/embeddings/) | 409s |
| [Local all-MiniLM-L6-v2](https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2) | 880s |
| [Local bge-small-en-v1.5](https://huggingface.co/BAAI/bge-small-en-v1.5)| 3379s |
| [Local bge-large-en-v1.5](https://huggingface.co/BAAI/bge-large-en-v1.5)| 9132s |

Meilisearch benchmarks indicate that indexing times range from under a minute for optimized cloud-based solutions to several hours for certain local models without GPU acceleration.

These should be important points to ponder when you weigh the frequency and volume of data updates in your applications. This is because they directly impact how quickly your system can process frequent or large data updates, which is crucial for maintaining the performance and responsiveness of your search solution.

### 4. Pricing
Embedding model costs vary depending on the provider and usage model. While local models are free to run, they require computational resources, potentially necessitating GPU investment.

On the other hand, cloud-based services charge per million tokens (or thousand neurons for Cloudflare), with costs ranging from $0.02 to $0.18 per million tokens.

| **Provider** | **Price** |
| --- | --- | 
| Cohere | $0.10 per million tokens for Embed 3 |
| OpenAI | $0.02 per million tokens for text-embedding-3-small |
| | $0.10 per million tokens for text-embedding-ada-002 |
| | $0.13 per million tokens for text-embedding-3-large | 
| Cloudflare | $0.011 per 1,000 Neurons |
| Jina | $0.18 per million tokens |
| Mistral | $0.10 per million tokens |
| VoyageAI |$0.02 per million tokens for voyage-3-lite |
| | $0.06 per million tokens for voyage-3|
| | $0.12 per million tokens for voyage-multimodal-3 |
| | $0.18 per million tokens for voyage-code-3| 
| | $0.18 per million tokens for voyage-3-large |
| Local model | Free |

As such, analyze cost-effectiveness against search demand and performance requirements. It's often a good idea to start with a well-known model that's easy to set up and has robust community support. You can migrate to a cloud provider like AWS for improved performance when necessary. 

Alternatively, you can choose an open-source model to self-host, giving you more flexibility. Just be aware that optimizing local models for high volume may require scaling your infrastructure.

### 5. Other optimization techniques

To maximize search performance, [hybrid search approaches](https://www.meilisearch.com/blog/full-text-search-vs-vector-search) combining full-text and vector search can yield the best results. To refine semantic search performance, consider the following optimizations:

- Experiment with model presets, as some models allow tuning for query vs. document embedding, which can improve relevancy.
- Assess specialized models, especially those employing retrieval-augmented generation, as domain-specific models may offer superior results for particular use cases.
- Explore models that provide reranking functions that further enhance search precision.
- Test higher-tier accounts as premium tiers may provide faster processing and reduced rate limiting.
- Optimize data transfer with quantization options to reduce API response sizes and improve efficiency.

Carefully evaluating these factors will help you select the best semantic search model to meet your needs.

## The bottom line
Now that we’ve taken this journey together, let’s round up what we’ve discovered.

| **Model/Service** | **Dimensions** | **Context Length** | **Latency** | **Indexation time** | **Pricing (per million tokens)** |
| --- | --- | --- | --- | --- | --- | 
| [Cohere embed-english-v3.0](https://docs.cohere.com/docs/embeddings) | 1024 | 512 | ±170ms | 43s | $0.10|
|[Cohere embed-english-light-v3.0](https://docs.cohere.com/docs/embeddings)|384|512|±160ms|16s|$0.10|
| [OpenAI text-embedding-3-small](https://platform.openai.com/docs/guides/embeddings) |1536| 8192| ±460ms| 95s|$0.02|
|[OpenAI text-embedding-3-large](https://platform.openai.com/docs/guides/embeddings)| 3072|8192|±750ms|151s|$0.13|
|[Mistral](https://docs.mistral.ai/capabilities/embeddings/)|1024|8192|±200ms|409s|$0.10|
|[VoyageAI voyage-2](https://www.voyageai.com/)|1024|4000|±350ms|330s|$0.10|
| [VoyageAI voyage-large-2](https://www.voyageai.com/)|1536|16000|±400ms|409s|$0.12|
|[Jina Colbert v2](https://jina.ai/news/jina-colbert-v2-multilingual-late-interaction-retriever-for-embedding-and-reranking/)|128, 96, or 64|8192|±400ms|375s|$0.18|
|[OSS all-MiniLM-L6-v2](https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2)|384|512|±10ms|880s|Free|
|[OSS bge-small-en-v1.5](https://huggingface.co/BAAI/bge-small-en-v1.5)|1024|512|±20ms|3379s|Free|
|[OSS bge-large-en-v1.5](https://huggingface.co/BAAI/bge-large-en-v1.5)|1536|512|±60ms|9132s|Free|

Choosing the best model for semantic search depends on what (and how) you aim to achieve with your specific use case, budget, and performance requirements. For most cases, cloud-based solutions such as those provided by Cohere or OpenAI may be optimal.

As the needs of your organization grow, it might be worth the cost and effort to upgrade to a local or self-hosted solution. Understanding your own needs is paramount to making a well-informed decision.

If you’re unsure which model is best for you or if you’re looking for a tailored solution, contact a search expert at Meilisearch.

## Frequently asked questions (FAQs)

### How does semantic search differ from keyword-based search?
Semantic search focuses on understanding meaning, while keyword-based search relies on exact word matches.
### What are the key features of a good semantic search model?
A good model should offer high accuracy, low latency, efficient indexing, and cost-effectiveness.
### How do transformer-based models improve semantic search?
Transformers process text in context, capturing contextual relationships between words to enhance search relevance while reducing downtime.
### How do vector databases enhance semantic search?
Vector databases work to efficiently store embeddings, enabling fast and scalable search operations.
### What are some open-source models for semantic search?
Popular open-source semantic search models include all-MiniLM-L6-v2 and Universal Sentence Encoder.




Meilisearch latest news and company updates