

By now, everyone knows that hybrid search has improved [RAG](https://zilliz.com/learn/Retrieval-Augmented-Generation) (Retrieval-Augmented Generation) search quality. While [dense embedding](https://zilliz.com/learn/sparse-and-dense-embeddings) search has shown impressive capabilities in capturing deep semantic relationships between queries and documents, it still has notable limitations. These include a lack of explainability and suboptimal performance with long-tail queries and rare terms.

Many RAG applications struggle because pre-trained models often lack domain-specific knowledge. In some scenarios, simple BM25 keyword matching outperforms these sophisticated models. This is where hybrid search bridges the gap, combining the semantic understanding of dense vector retrieval with the precision of keyword matching.


## Why Hybrid Search is Complex in Production

While frameworks like [LangChain](https://zilliz.com/learn/LangChain) or [LlamaIndex](https://zilliz.com/learn/getting-started-with-llamaindex) make building a proof-of-concept hybrid retriever easy, scaling to production with massive datasets is challenging. Traditional architectures require separate vector databases and search engines, leading to several key challenges:

- High infrastructure maintenance costs and operational complexity

- Data redundancy across multiple systems

- Difficult data consistency management

- Complex security and access control across systems

The market needs a unified solution that supports lexical and semantic search while reducing system complexity and cost.

![](https://assets.zilliz.com/elasticsearch_vs_milvus_5be6e2b69e.png)


## The Pain Points of Elasticsearch

Elasticsearch has been one of the past decade's most influential open-source search projects. Built on Apache Lucene, it gained popularity through its high performance, scalability, and distributed architecture. While it added vector ANN search in version 8.0, production deployments face several critical challenges:

**High Update and Indexing Costs:** Elasticsearch's architecture doesn't fully decouple write operations, index building, and querying. This leads to significant CPU and I/O overhead during write operations, especially in bulk updates. The resource contention between indexing and querying impacts performance, creating a major bottleneck for high-frequency update scenarios.

**Poor Real-time Performance:** As a "near real-time" search engine, Elasticsearch introduces noticeable latency in data visibility. This latency becomes particularly problematic for AI applications, such as Agent systems, where high-frequency interactions and dynamic decision-making require immediate data access.

**Difficult Shard Management:** While Elasticsearch uses sharding for distributed architecture, shard management poses significant challenges. The lack of dynamic sharding support creates a dilemma: too many shards in small datasets lead to poor performance, while too few shards in large datasets limit scalability and cause uneven data distribution.

**Non-Cloud-Native Architecture:** Developed before cloud-native architectures became prevalent, Elasticsearch's design tightly couples storage and compute, limiting its integration with modern infrastructure like public clouds and Kubernetes. Resource scaling requires simultaneous increases in both storage and compute, reducing flexibility. In multi-replica scenarios, each shard must build its index independently, increasing computational costs and reducing resource efficiency.

**Poor Vector Search Performance:** Though Elasticsearch 8.0 introduced vector ANN search, its performance significantly lags behind that of dedicated vector engines like Milvus. Based on the Lucene kernel, its index structure proves inefficient for high-dimensional data, struggling with large-scale vector search requirements. Performance becomes particularly unstable in complex scenarios involving scalar filtering and multi-tenancy, making it challenging to support high-load or diverse business needs.

**Excessive Resource Consumption:** Elasticsearch places extreme demands on memory and CPU, especially when processing large-scale data. Its JVM dependency requires frequent heap size adjustments and garbage collection tuning, severely impacting memory efficiency. Vector search operations require intensive SIMD-optimized computations, for which the JVM environment is far from ideal.

These fundamental limitations become increasingly problematic as organizations scale their AI infrastructure, making Elasticsearch particularly challenging for modern AI applications requiring high performance and reliability.


## Introducing Sparse-BM25: Reimagining Lexical Search

[Milvus 2.5](https://milvus.io/blog/introduce-milvus-2-5-full-text-search-powerful-metadata-filtering-and-more.md) introduces native lexical search support through Sparse-BM25, building upon the hybrid search capabilities introduced in version 2.4. This innovative approach includes the following key components:

- Advanced tokenization and preprocessing via Tantivy

- Distributed vocabulary and term frequency management

- Sparse vector generation using corpus TF and query TF-IDF

- Inverted index support with WAND algorithm (Block-Max WAND and graph index support in development)

Compared to Elasticsearch, Milvus offers significant advantages in algorithm flexibility. Its vector distance-based similarity computation enables more sophisticated matching, including implementing TW-BERT (Term Weighting BERT) based on "End-to-End Query Term Weighting" research. This approach has demonstrated superior performance in both in-domain and out-domain testing.

Another crucial advantage is cost efficiency. By leveraging both inverted index and dense embedding compression, Milvus achieves a fivefold performance improvement with less than 1% recall degradation. Through tail-term pruning and vector quantization, memory usage has been reduced by over 50%.

Long query optimization stands out as a particular strength. Where traditional WAND algorithms struggle with longer queries, Milvus excels by combining sparse embeddings with graph indices, delivering a tenfold performance improvement in high-dimensional sparse vector search scenarios.

![](https://assets.zilliz.com/document_in_and_out_b84771bec4.png)


## Milvus: The Ultimate Vector Database for RAG

Milvus is the premier choice for RAG applications through its comprehensive feature set. Key advantages include:

- Rich metadata support with dynamic schema capabilities and powerful filtering options

- Enterprise-grade multi-tenancy with flexible isolation through collections, partitions, and partition keys

- Industry-first disk vector index support with multi-tier storage from memory to S3

- Cloud-native scalability supporting seamless scaling from 10M to 1B+ vectors

- Comprehensive search capabilities, including grouping, range, and hybrid search

- Deep ecosystem integration with LangChain, LlamaIndex, Dify, and other AI tools

The system's diverse search capabilities encompass grouping, range, and hybrid search methodologies. Deep integration with tools like LangChain, LlamaIndex, and Dify, as well as support for numerous AI products, places Milvus at the center of the modern AI infrastructure ecosystem.


## Looking Forward

As AI transitions from POC to production, Milvus continues to evolve. We focus on making vector search more accessible and cost-effective while enhancing search quality. Whether you're a startup or an enterprise, Milvus reduces the technical barriers to AI application development.

This commitment to accessibility and innovation has led us to another major step forward. While our open-source solution continues to serve as the foundation for thousands of applications worldwide, we recognize that many organizations need a fully managed solution that eliminates operational overhead.


## Zilliz Cloud: The Managed Solution

We've built [Zilliz Cloud](https://zilliz.com/cloud), a fully managed vector database service based on Milvus, over the past three years. Through a cloud-native reimplementation of the Milvus protocol, it offers enhanced usability, cost efficiency, and security.

Drawing from our experience maintaining the world's largest vector search clusters and supporting thousands of AI application developers, Zilliz Cloud significantly reduces operational overhead and costs compared to self-hosted solutions.

Ready to experience the future of vector search? Start your free trial today with up to $200 in credits, no credit card required.


Elasticsearch is Dead, Long Live Lexical Search


In this article, we will show you how to quickly get up and running with the new full-text search feature and combine it with the conventional semantic search based on vector embeddings.

## Requirement

First, ensure you have installed Milvus 2.5:


```
pip install -U pymilvus[model]
```

and have a running instance of Milvus Standalone (e.g. on your local machine) using the [installation instructions in the Milvus docs](https://milvus.io/docs/prerequisite-docker.md).

## Building the Data Schema and Search Indices

We import the required classes and functions:

```
from pymilvus import MilvusClient, DataType, Function, FunctionType, model
```

You may have noticed two new entries for Milvus 2.5, `Function` and `FunctionType`, which we will explain shortly.

Next we open the database with Milvus Standalone, that is, locally, and create the data schema. The schema comprises an integer primary key, a text string, a dense vector of dimension 384, and a sparse vector (of unlimited dimensionality).
Note that Milvus Lite does not currently support full-text search, only Milvus Standalone and Milvus Distributed.

```
client = MilvusClient(uri="http://localhost:19530")

schema = client.create_schema()

schema.add_field(field_name="id", datatype=DataType.INT64, is_primary=True, auto_id=True)
schema.add_field(field_name="text", datatype=DataType.VARCHAR, max_length=1000, enable_analyzer=True)
schema.add_field(field_name="dense", datatype=DataType.FLOAT_VECTOR, dim=768),
schema.add_field(field_name="sparse", datatype=DataType.SPARSE_FLOAT_VECTOR)
```

```
{'auto_id': False, 'description': '', 'fields': [{'name': 'id', 'description': '', 'type': <DataType.INT64: 5>, 'is_primary': True, 'auto_id': True}, {'name': 'text', 'description': '', 'type': <DataType.VARCHAR: 21>, 'params': {'max_length': 1000, 'enable_analyzer': True}}, {'name': 'dense', 'description': '', 'type': <DataType.FLOAT_VECTOR: 101>, 'params': {'dim': 768}}, {'name': 'sparse', 'description': '', 'type': <DataType.SPARSE_FLOAT_VECTOR: 104>}], 'enable_dynamic_field': False}
```

You may have noticed the `enable_analyzer=True` parameter. This tells Milvus 2.5 to enable the lexical parser on this field and build a list of tokens and token frequencies, which are required for full-text search. The `sparse` field will hold a vector representation of the documentation as a bag-of-words produced from the parsing `text`.

But how do we connect the `text` and `sparse` fields, and tell Milvus how `sparse` should be calculated from `text`? This is where we need to invoke the `Function` object and add it to the schema:


```
bm25_function = Function(
    name="text_bm25_emb", # Function name
    input_field_names=["text"], # Name of the VARCHAR field containing raw text data
    output_field_names=["sparse"], # Name of the SPARSE_FLOAT_VECTOR field reserved to store generated embeddings
    function_type=FunctionType.BM25,
)

schema.add_function(bm25_function)
```


```
{'auto_id': False, 'description': '', 'fields': [{'name': 'id', 'description': '', 'type': <DataType.INT64: 5>, 'is_primary': True, 'auto_id': True}, {'name': 'text', 'description': '', 'type': <DataType.VARCHAR: 21>, 'params': {'max_length': 1000, 'enable_analyzer': True}}, {'name': 'dense', 'description': '', 'type': <DataType.FLOAT_VECTOR: 101>, 'params': {'dim': 768}}, {'name': 'sparse', 'description': '', 'type': <DataType.SPARSE_FLOAT_VECTOR: 104>, 'is_function_output': True}], 'enable_dynamic_field': False, 'functions': [{'name': 'text_bm25_emb', 'description': '', 'type': <FunctionType.BM25: 1>, 'input_field_names': ['text'], 'output_field_names': ['sparse'], 'params': {}}]}
```

The abstraction of the `Function` object is more general than that of applying full-text search. In the future, it may be used for other cases where one field needs to be a function of another field. In our case, we specify that `sparse` is a function of `text` via the function `FunctionType.BM25`. `BM25` refers to a common metric in information retrieval used for calculating a query's similarity to a document (relative to a collection of documents).

We use the default embedding model in Milvus, which is [paraphrase-albert-small-v2](https://huggingface.co/GPTCache/paraphrase-albert-small-v2):


```
embedding_fn = model.DefaultEmbeddingFunction()
```

The next step is to add our search indices. We have one for the dense vector and a separate one for the sparse vector. The index type is `SPARSE_INVERTED_INDEX` with `BM25` since full-text search requires a different search method than those for standard dense vectors.


```
index_params = client.prepare_index_params()

index_params.add_index(
    field_name="dense",
    index_type="AUTOINDEX", 
    metric_type="COSINE"
)

index_params.add_index(
    field_name="sparse",
    index_type="SPARSE_INVERTED_INDEX", 
    metric_type="BM25"
)
```

Finally, we create our collection:

```
client.drop_collection('demo')
client.list_collections()
```

```
[]
```

```
client.create_collection(
    collection_name='demo', 
    schema=schema, 
    index_params=index_params
)

client.list_collections()
```


```
['demo']
```
And with that, we have an empty database set up to accept text documents and perform semantic and full-text searches!

## Inserting Data and Performing Full-Text Search

Inserting data is no different than previous versions of Milvus:

```
docs = [
    'information retrieval is a field of study.',
    'information retrieval focuses on finding relevant information in large datasets.',
    'data mining and information retrieval overlap in research.'
]

embeddings = embedding_fn(docs)

client.insert('demo', [
    {'text': doc, 'dense': vec} for doc, vec in zip(docs, embeddings)
])
```

```
{'insert_count': 3, 'ids': [454387371651630485, 454387371651630486, 454387371651630487], 'cost': 0}
```

Let's first illustrate a full-text search before we move on to hybrid search:

```
search_params = {
    'params': {'drop_ratio_search': 0.2},
}

results = client.search(
    collection_name='demo', 
    data=['whats the focus of information retrieval?'],
    output_fields=['text'],
    anns_field='sparse',
    limit=3,
    search_params=search_params
)
```

The search parameter `drop_ratio_search` refers to the proportion of lower-scoring documents to drop during the search algorithm.

Let's view the results:

```
for hit in results[0]:
    print(hit)
```

```
{'id': 454387371651630485, 'distance': 1.3352930545806885, 'entity': {'text': 'information retrieval is a field of study.'}}
{'id': 454387371651630486, 'distance': 0.29726022481918335, 'entity': {'text': 'information retrieval focuses on finding relevant information in large datasets.'}}
{'id': 454387371651630487, 'distance': 0.2715056240558624, 'entity': {'text': 'data mining and information retrieval overlap in research.'}}
```

## Performing Hybrid Semantic and Full-Text Search

Let's now combine what we've learned to perform a hybrid search that combines separate semantic and full-text searches with a reranker:

```
from pymilvus import AnnSearchRequest, RRFRanker
query = 'whats the focus of information retrieval?'
query_dense_vector = embedding_fn([query])

search_param_1 = {
    "data": query_dense_vector,
    "anns_field": "dense",
    "param": {
        "metric_type": "COSINE",
    },
    "limit": 3
}
request_1 = AnnSearchRequest(**search_param_1)

search_param_2 = {
    "data": [query],
    "anns_field": "sparse",
    "param": {
        "metric_type": "BM25",
        "params": {"drop_ratio_build": 0.0}
    },
    "limit": 3
}
request_2 = AnnSearchRequest(**search_param_2)

reqs = [request_1, request_2]
```

```
ranker = RRFRanker()

res = client.hybrid_search(
    collection_name="demo",
    output_fields=['text'],
    reqs=reqs,
    ranker=ranker,
    limit=3
)
for hit in res[0]:
    print(hit)
```

```
{'id': 454387371651630485, 'distance': 0.032786883413791656, 'entity': {'text': 'information retrieval is a field of study.'}}
{'id': 454387371651630486, 'distance': 0.032258063554763794, 'entity': {'text': 'information retrieval focuses on finding relevant information in large datasets.'}}
{'id': 454387371651630487, 'distance': 0.0317460335791111, 'entity': {'text': 'data mining and information retrieval overlap in research.'}}
```

As you may have noticed, this is no different than a hybrid search with two separate semantic fields (available since Milvus 2.4). The results are identical to full-text search in this simple example, but for larger databases and keyword specific searches hybrid search typically has higher recall.

## Summary

You're now equipped with all the knowledge needed to perform full-text and hybrid semantic/full-text search with Milvus 2.5. See the following articles for more discussion on how full-text search works and why it is complementary to semantic search:

- [Introducing Milvus 2.5: Full-Text Search, More Powerful Metadata Filtering, and Usability Improvements!](https://milvus.io/blog/introduce-milvus-2-5-full-text-search-powerful-metadata-filtering-and-more.md)
- [Semantic Search v.s. Full-Text Search: Which Do I Choose in Milvus 2.5?](https://milvus.io/blog/semantic-search-vs-full-text-search-which-one-should-i-choose-with-milvus-2-5.md)


Getting Started with Hybrid Semantic / Full-Text Search with Milvus 2.5



## Overview

We are thrilled to present the latest version of Milvus, 2.5, which introduces a powerful new capability: [full-text search](https://milvus.io/docs/full-text-search.md#Full-Text-Search), also known as lexical or keyword search. If you are new to search, full-text search allows you to find documents by searching for specific words or phrases within them, similar to how you search in Google. This complements our existing semantic search capabilities, which understand the meaning behind your search rather than just matching exact words.

We use the industry-standard BM25 metric for document similarity, and our implemention is based on sparse vectors, allowing for more efficient storage and retrieval. For those unfamiliar with the term, sparse vectors are a way to represent text where most values are zero, making them very efficient to store and process—imagine a huge spreadsheet where only a few cells contain numbers, and the rest are empty. This approach fits well into Milvus's product philosophy where the vector is the core search entity.

An additional noteworthy aspect of our implementation is the capability to insert and query text _directly_ rather than having users first manually convert text into sparse vectors. This takes Milvus one step closer towards fully processing unstructured data.

But this is just the beginning. With the release of 2.5, we updated the [Milvus product roadmap](https://milvus.io/docs/roadmap.md). In future product iterations of Milvus, our focus will be on evolving Milvus's capabilities in four key directions:

- Streamlined unstructured data processing;
- Better search quality and efficiency;
- Easier data management;
- Lowering costs through algorithmic and design advances

Our aim is to build data infrastructure that can both efficiently store and effectively retrieve information in the AI era.


## Full-text Search via Sparse-BM25

Although semantic search typically has better contexual awareness and intent understanding, when a user needs to search for specific proper nouns, serial numbers, or a completely-matching phrase, full-text retrieval with keyword matching often produces more accurate results.

To illustrate this with an example:

- Semantic search excels when you ask: "Find documents about renewable energy solutions"
- Full-text search is better when you need: "Find documents mentioning _Tesla Model 3 2024_"

In our previous version (Milvus 2.4), users had to pre-process their text using a separate tool (the BM25EmbeddingFunction module to the PyMilvus) on their own machines before they could search it This approach had several limitations: it couldn't handle growing datasets well, required extra setup steps, and made the whole process more complicated than necessary. For the technically minded, the key limitations were that it could only work on a single machine; the vocabulary and other corpus statistics used for BM25 scoring couldn't be updated as the corpus changed; and converting text to vectors on the client side is less intuitive working withtext directly.

Milvus 2.5 simplifies everything. Now you can work with your text directly:

- Store your original text documents as they are
- Search using natural language queries
- Get results back in readable form

Behind the scenes, Milvus handles all the complex vector conversions automatically making it easier to work with text data. This is what we call our "Doc in, Doc out" approach—you work with readable text, and we handle the rest.


### Techical Implementation

For those interested in the technical details, Milvus 2.5 adds the full-text search capability through its built-in Sparse-BM25 implementation, including:

- **A Tokenizer built on tantivy**: Milvus now integrates with the thriving tantivy ecosystem
- **Capability to ingest and retrieve raw documents**: Support for direct ingestion and query of text data
- **BM25 relevance scoring**: Internalize BM25 scoring, implemented based on sparse vector

We chose to work with the well-developed tantivy ecosystem and build the Milvus text tokenizer on tantivy. In the future, Milvus will support more tokenizers and expose the tokenization process to help users better understand the retrieval quality. We will also explore deep learning-based tokenizers and stemmer strategies to further optimize the performance of full-text search. Below is sample code for using and configuring the tokenizer:

```Python
# Tokenizer configuration
schema.add_field(
    field_name='text',
    datatype=DataType.VARCHAR,
    max_length=65535,
    enable_analyzer=True, # Enable tokenizer on this column
    analyzer_params={"type": "english"},  # Configure tokenizer parameters, here we choose the english template, fine-grained configuration is also supported
    enable_match=True, # Build an inverted index for Text_Match
)
```

After configuring the tokenizer in the collection schema, users can register the text to bm25 function via add_function method. This will run internally in the Milvus server. All subsequent data flows such as additions, deletions, modifications, and queries can be completed by operating on the raw text string, as opposed to the vector representation. See below code example for how to ingest text and conduct full-text search with the new API:

```Python
# Define the mapping relationship between raw text data and vectors on the schema
bm25_function = Function(
    name="text_bm25_emb",
    input_field_names=["text"], # Input text field
    output_field_names=["sparse"], # Internal mapping sparse vector field
    function_type=FunctionType.BM25, # Model for processing mapping relationship
)

schema.add_function(bm25_function)
...
# Support for raw text in/out
MilvusClient.insert('demo', [
    {'text': 'Artificial intelligence was founded as an academic discipline in 1956.'},
    {'text': 'Alan Turing was the first person to conduct substantial research in AI.'},
    {'text': 'Born in Maida Vale, London, Turing was raised in southern England.'},
])

MilvusClient.search(
    collection_name='demo',
    data=['Who started AI research?'],
    anns_field='sparse',
    limit=3
)
```

We have adopted an implementation of BM25 relevance scoring that represents queries and documents as sparse vectors, called **Sparse-BM25**. This unlocks many optimizations based on sparse vector, such as:

Milvus achieves hybrid search capabilities through its cutting-edge **Sparse-BM25 implementation**, which integrates full-text search into the vector database architecture. By representing term frequencies as sparse vectors instead of traditional inverted indexes, Sparse-BM25 enables advanced optimizations, such as **graph indexing**, **product quantization (PQ)**, and **scalar quantization (SQ)**. These optimizations minimize memory usage and accelerate search performance. Similar to inverted index approach, Milvus supports taking raw text as input and generating sparse vectors internally. This make it able to work with any tokenizer and grasp any word shown in the dynamicly changing corpus.

Additionally, heuristic-based pruning discards low-value sparse vectors, further enhancing efficiency without compromising accuracy. Unlike previous approach using sparse vector, it can adapt to a growing corpus, not the accuracy of BM25 scoring.

1. Building graph indexes on the sparse vector, which performs better than inverted index on queries with long text as inverted index needs more steps to finish matching the tokens in the query;
2. Leveraging approximation techniques to speed up search with only minor impact on retrieval quality, such as vector quantization and heuristic based pruning;
3. Unifying the interface and data model for performing semantic search and full-text search, thus enhancing the user experience.

```Python
# Creating an index on the sparse column
index_params.add_index(
    field_name="sparse",
    index_type="AUTOINDEX",  # Default WAND index
    metric_type="BM25" # Configure relevance scoring through metric_type
)

# Configurable parameters at search time to speed up search
search_params = {
    'params': {'drop_ratio_search': 0.6}, # WAND search parameter configuration can speed up search
}
```

In summary, Milvus 2.5 has expanded its search capability beyond semantic search by introducing full-text retrieval, making it easier for users to build high quality AI applications. These are just initial steps in the space of Sparse-BM25 search and we anticipate that there will be further optimization measures to try in the future.


## Text Matching Search Filters

A second text search feature released with Milvus 2.5 is **Text Match**, which allows the user to filter the search to entries containing a specific text string. This feature is also built on the basis of tokenization and is activated with `enable_match=True`.

It is worth noting that with Text Match, the processing of the query text is based on the logic of OR after tokenization. For example, in the example below, the result will return all documents (using the 'text' field) that contain either 'vector' or 'database'.

```Python
filter = "TEXT_MATCH(text, 'vector database')"
```

If your scenario requires matching both 'vector' and 'database', then you need to write two separate Text Matches and overlay them with AND to achieve your goal.

```Python
filter = "TEXT_MATCH(text, 'vector') and TEXT_MATCH(text, 'database')"
```


## Significant Enhancement in Scalar Filtering Performance

Our emphasis on scalar filtering performance originates from our discovery that the combination of vector retrieval and metadata filtering can greatly improve query performance and accuracy in various scenarios. These scenarios range from image search applications such as corner case identification in autonomous driving to complex RAG scenarios in enterprise knowledge bases. Thus, it is highly suitable for enterprise users to implement in large-scale data application scenarios.

In practice, many factors like how much data you're filtering, how your data is organized, and how you're searching can affect performance. To address this, Milvus 2.5 introduces three new types of indexes—BitMap Index, Array Inverted Index, and the Inverted Index after tokenizing the Varchar text field. These new indexes can significantly improve performance in real-world use cases.

Specifically:

1. **BitMap Index** can be used to accelerate tag filtering (common operators include in, array_contains, etc.), and is suitable for scenarios with fewer field category data (data cardinality). The principle is to determine whether a row of data has a certain value on a column, with 1 for yes and 0 for no, and then maintain a BitMap list. The following chart shows the performance test comparison we conducted based on a customer's business scenario. In this scenario, the data volume is 500 million, the data category is 20, different values have different distribution proportions (1%, 5%, 10%, 50%), and the performance under different filtering amounts also varies. With 50% filtering, we can achieve a 6.8-fold performance gain through BitMap Index. It's worth noting that as cardinality increases, compared to BitMap Index, Inverted Index will show more balanced performance. 

![](https://assets.zilliz.com/QPS_comparison_f3f580d697.png)


2. **Text Match** is based on the Inverted Index after the text field is tokenized. Its performance far exceeds the Wildcard Match (i.e., like + %) function we provided in 2.4. According to our internal test results, the advantages of Text Match are very clear, especially in concurrent query scenarios, where it can achieve up to a 400-fold QPS increase. 

![](https://assets.zilliz.com/data_size_and_concurrency_e19dc44c59.png)

In terms of JSON data processing, we plan to introduce in subsequent versions of 2.5.x the building of inverted indices for user-specified keys and default location information recording for all keys to speed up parsing. We expect both of these areas to significantly enhance the query performance of JSON and Dynamic Field. We plan to showcase more information in future release notes and technical blogs, so stay tuned!


## New Management Interface

Managing a database shouldn't require a computer science degree, but we know database administrators need powerful tools. That's why we've introduced the **Cluster Management WebUI**, a new web-based interface accessible at your cluster's address on port 9091/webui. This observability tool provides:

- Real-time monitoring dashboards showing cluster-wide metrics
- Detailed memory and performance analytics per node
- Segment information and slow query tracking
- System health indicators and node status
- Easy-to-use troubleshooting tools for complex system issues

While this interface is still in beta, we're actively developing it based on user feedback from database administrators. Future updates will include AI-assisted diagnostics, more interactive management features, and enhanced cluster observability capabilities.


## Documentation and Developer Experience

We've completely revamped our **documentation** and **SDK/API** experience to make Milvus more accessible while maintaining depth for experienced users. The improvements include:

- A restructured documentation system with clearer progression from basic to advanced concepts
- Interactive tutorials and real-world examples that showcase practical implementations
- Comprehensive API references with practical code samples
- A more user-friendly SDK design that simplifies common operations
- Illustrated guides that make complex concepts easier to understand
- An AI-powered documentation assistant (ASK AI) for quick answers

The updated SDK/API focuses on improving developer experience through more intuitive interfaces and better integration with the documentation. We believe you'll notice these improvements when working with the 2.5.x series.

However, we know documentation and SDK development is an ongoing process. We'll continue optimizing both the content structure and SDK design based on community feedback. Join our Discord channel to share your suggestions and help us improve further.


## **Summary**

Milvus 2.5 contains 13 new features and several system-level optimizations, contributed not just by Zilliz but the open-source community. We have only touched on a few of them in this post and encourage you to visit our [release note](https://milvus.io/docs/release_notes.md) and [official documents](https://milvus.io/docs) for more information!


Milvus Metadata Management (1) How to View Metadata

How to View Metadata

SQLite

MySQL

相关博客

On This Page

Try Managed Milvus for Free

Keep Reading

Elasticsearch is Dead, Long Live Lexical Search

Getting Started with Hybrid Semantic / Full-Text Search with Milvus 2.5

Introducing Milvus 2.5: Full-Text Search, More Powerful Metadata Filtering, and Usability Improvements!

Semantic Search vs. Full-Text Search: Which Do I Choose in Milvus 2.5?