
![](https://assets.zilliz.com/2_72e444c8dc.JPG)

We are excited to introduce [Milvus Lite](https://milvus.io/docs/milvus_lite.md), a lightweight vector database that runs locally within your Python application. Based on the popular open-source [Milvus](https://milvus.io/intro) vector database, Milvus Lite reuses the core components for vector indexing and query parsing while removing elements designed for high scalability in distributed systems. This design makes a compact and efficient solution ideal for environments with limited computing resources, such as laptops, Jupyter Notebooks, and mobile or edge devices.

Milvus Lite integrates with various AI development stacks like LangChain and LlamaIndex, enabling its use as a vector store in Retrieval Augmented Generation (RAG) pipelines without the need for server setup. Simply run `pip install pymilvus` (version 2.4.3 or above) to incorporate it into your AI application as a Python library.

Milvus Lite shares the Milvus API, ensuring that your client-side code works for both small-scale local deployments and Milvus servers deployed on Docker or Kubernetes with billions of vectors.

<iframe style="border-radius:12px" src="https://open.spotify.com/embed/episode/5bMcZgPgPVxSuoi1M2vn1p?utm_source=generator" width="100%" height="152" frameBorder="0" allowfullscreen="" allow="autoplay; clipboard-write; encrypted-media; fullscreen; picture-in-picture" loading="lazy"></iframe>

## Why We Built Milvus Lite

Many AI applications require vector similarity search for unstructured data, including text, images, voices, and videos, for applications such as chatbots and shopping assistants. Vector databases are crafted for storing and searching vector embeddings and are a crucial part of the AI development stack, particularly for generative AI use cases like [Retrieval Augmented Generation (RAG)](https://zilliz.com/learn/Retrieval-Augmented-Generation).

Despite the availability of numerous vector search solutions, an easy-to-start option that also works for large-scale production deployments was missing. As the creators of Milvus, we designed Milvus Lite to help AI developers build applications faster while ensuring a consistent experience across various deployment options, including Milvus on Kubernetes, Docker, and managed cloud services.

Milvus Lite is a crucial addition to our suite of offerings within the Milvus ecosystem. It provides developers with a versatile tool that supports every stage of their development journey. From prototyping to production environments and from edge computing to large-scale deployments, Milvus is now the only vector database that covers use cases of any size and all stages of development.

## How Milvus Lite Works 

Milvus Lite supports all the basic operations available in Milvus, such as creating collections and inserting, searching, and deleting vectors. It will soon support advanced features like hybrid search. Milvus Lite loads data into memory for efficient searches and persists it as an SQLite file.

Milvus Lite is included in the [Python SDK of Milvus](https://github.com/milvus-io/pymilvus) and can be deployed with a simple `pip install pymilvus`. The following code snippet demonstrates how to set up a vector database with Milvus Lite by specifying a local file name and then creating a new collection. For those familiar with the Milvus API, the only difference is that the `uri` refers to a local file name instead of a network endpoint, e.g., `"milvus_demo.db"` instead of `"http://localhost:19530"` for a Milvus server. Everything else remains the same. Milvus Lite also supports storing raw text and other labels as metadata, using a dynamic or explicitly defined schema, as shown below.

```
from pymilvus import MilvusClient

client = MilvusClient("milvus_demo.db")
# This collection can take input with mandatory fields named "id", "vector" and
# any other fields as "dynamic schema". You can also define the schema explicitly.
client.create_collection(
    collection_name="demo_collection",
    dimension=384  # Dimension for vectors.
)
```
For scalability, an AI application developed with Milvus Lite can easily transition to using Milvus deployed on Docker or Kubernetes by simply specifying the `uri` with the server endpoint.


## Integration with AI Development Stack

In addition to introducing Milvus Lite to make vector search easy to start with, Milvus also integrates with many frameworks and providers of the AI development stack, including [LangChain](https://python.langchain.com/v0.2/docs/integrations/vectorstores/milvus/), [LlamaIndex](https://docs.llamaindex.ai/en/stable/examples/vector_stores/MilvusIndexDemo/), [Haystack](https://haystack.deepset.ai/integrations/milvus-document-store), [Voyage AI](https://blog.voyageai.com/2024/05/30/semantic-search-with-milvus-lite-and-voyage-ai/), [Ragas](https://milvus.io/docs/integrate_with_ragas.md), [Jina AI](https://jina.ai/news/implementing-a-chat-history-rag-with-jina-ai-and-milvus-lite/), [DSPy](https://dspy-docs.vercel.app/docs/deep-dive/retrieval_models_clients/MilvusRM), [BentoML](https://www.bentoml.com/blog/building-a-rag-app-with-bentocloud-and-milvus-lite), [WhyHow](https://chiajy.medium.com/70873c7576f1), [Relari AI](https://blog.relari.ai/case-study-using-synthetic-data-to-benchmark-rag-systems-be324904ace1), [Airbyte](https://docs.airbyte.com/integrations/destinations/milvus), [HuggingFace](https://milvus.io/docs/integrate_with_hugging-face.md) and [MemGPT](https://memgpt.readme.io/docs/storage#milvus). Thanks to their extensive tooling and services, these integrations simplify the development of AI applications with vector search capability.

And this is just the beginning—many more exciting integrations are coming soon! Stay tuned! 


## More Resources and Examples

Explore [Milvus quickstart documentation](https://milvus.io/docs/quickstart.md) for detailed guides and code examples on using Milvus Lite to build AI applications like Retrieval-Augmented Generation ([RAG](https://github.com/milvus-io/bootcamp/blob/master/bootcamp/tutorials/quickstart/build_RAG_with_milvus.ipynb)) and [image search](https://github.com/milvus-io/bootcamp/blob/master/bootcamp/tutorials/quickstart/image_search_with_milvus.ipynb).

Milvus Lite is an open-source project, and we welcome your contributions. Check out our [Contributing Guide](https://github.com/milvus-io/milvus-lite/blob/main/CONTRIBUTING.md) to get started. You can also report bugs or request features by filing an issue on the [Milvus Lite GitHub](https://github.com/milvus-io/milvus-lite) repository.


Introducing Milvus Lite: Start Building a GenAI Application in Seconds


![](https://assets.zilliz.com/What_is_new_in_Milvus_2_3_4_1847b0fa8a.png)


We are excited to unveil the latest release of Milvus 2.3.4. This update introduces a suite of features and enhancements meticulously crafted to optimize performance, boost efficiency, and deliver a seamless user experience. In this blog post, we'll delve into the highlights of Milvus 2.3.4. 

## Access logs for improved monitoring

Milvus now supports access logs, offering invaluable insights into interactions with external interfaces. These logs record method names, user requests, response times, error codes, and other interaction information, empowering developers and system administrators to conduct performance analysis, security auditing, and efficient troubleshooting. 

**_Note:_** _Currently, access logs only support gRPC interactions. However, our commitment to improvement continues, and future versions will extend this capability to include RESTful request logs._ 

For more detailed information, refer to [Configure Access Logs](https://milvus.io/docs/configure_access_logs.md).

## Parquet file imports for enhanced data processing efficiency

Milvus 2.3.4 now supports importing Parquet files, a widely embraced columnar storage format designed to enhance the efficiency of storing and processing large-scale datasets. This addition gives users increased flexibility and efficiency in their data processing endeavors. By eliminating the need for laborious data format conversions, users managing substantial datasets in the Parquet format will experience a streamlined data import process, significantly reducing the time from initial data preparation to subsequent vector retrieval.

Furthermore, our data format conversion tool, BulkWriter, has now embraced Parquet as its default output data format, ensuring a more intuitive experience for developers. 

## Binlog index on growing segments for faster searches

Milvus now leverages a binlog index on growing segments, resulting in up to tenfold faster searches in growing segments. This enhancement significantly boosts search efficiency and supports advanced indices like IVF or Fast Scan, improving the overall user experience.

## Support for up to 10,000 collections/partitions

Like tables and partitions in relational databases, collections and partitions are the core units for storing and managing vector data in Milvus. Responding to users' evolving needs for nuanced data organization, Milvus 2.3.4 now supports up to 10,000 collections/partitions in a cluster, a significant jump from the previous limit of 4,096. This enhancement benefits diverse use cases, such as knowledge base management and multi-tenant environments. The expanded support for collections/partitions stems from refinements to the time tick mechanism, goroutine management, and memory usage.

 **_Note:_** _The recommended limit for the number of collections/partitions is 10,000, as exceeding this limit may impact failure recovery and resource usage._

## Other enhancements 

In addition to the features above, Milvus 2.3.4 includes various improvements and bug fixes. These include reduced memory usage during data retrieval and variable-length data handling, refined error messaging, accelerated loading speed, and improved query shard balance. These collective enhancements contribute to a smoother and more efficient overall user experience.

For a comprehensive overview of all the changes introduced in Milvus 2.3.4, refer to our [Release Notes](https://milvus.io/docs/release_notes.md#v234).

## Stay connected!

If you have questions or feedback about Milvus, join our [Discord channel](https://discord.com/invite/8uyFbECzPX) to engage with our engineers and the community directly or join our [Milvus Community Lunch and Learn](https://discord.com/invite/RjNbk8RR4f) Every Tuesday from 12-12:30 PM PST. You’re also welcome to follow us on [Twitter](https://twitter.com/milvusio) or [LinkedIn](https://www.linkedin.com/company/the-milvus-project) for the latest news and updates about Milvus.


Milvus 2.3.4: Faster Searches, Expanded Data Support, Improved Monitoring, and More

![](https://assets.zilliz.com/What_s_New_in_Milvus_2_3_2_and_2_3_3_d3d0db03c3.png)

In the ever-evolving landscape of vector search technologies, Milvus remains at the forefront, pushing boundaries and setting new standards. Today, we are thrilled to announce the release of Milvus 2.3.2 and 2.3.3! These updates bring many exciting features, optimizations, and improvements, enhancing system performance, flexibility, and overall user experience.

## Support for Array data types - making search results more accurate and relevant

Adding the Array data type support is a pivotal enhancement for Milvus, particularly in query filtering scenarios like intersection and union. This addition ensures that search results are not only more accurate but also more relevant. In practical terms, for instance, within the e-commerce sector, product tags stored as string Arrays allow consumers to perform advanced searches, filtering out irrelevant results.

Dive into our comprehensive [documentation](https://milvus.io/docs/array_data_type.md) for an in-depth guide on leveraging Array types in Milvus.

## Support for complex delete expressions - improving your data management

In previous versions, Milvus supported primary key deletion expressions, providing a stable and streamlined architecture. With Milvus 2.3.2 or 2.3.3, users can employ complex delete expressions, facilitating sophisticated data management tasks such as rolling cleanup of old data or GDPR compliance-driven data deletion based on user IDs.

Note: Ensure you’ve loaded collections before utilizing complex expressions. Additionally, it's important to note that the deletion process does not guarantee atomicity.

## TiKV integration - scalable metadata storage with stability

Previously relying on Etcd for metadata storage, Milvus faced limited capacity and scalability challenges in metadata storage. To address these problems, Milvus added TiKV, an open-source key-value store, as one more option for metadata storage. TiKV offers enhanced scalability, stability, and efficiency, making it an ideal solution for Milvus's evolving requirements. Starting from Milvus 2.3.2, users can seamlessly transition to TiKV for their metadata storage by modifying the configuration.

## Support for FP16 vector type - embracing machine learning efficiency

Milvus 2.3.2 and later versions now support the FP16 vector type at the interface level. FP16, or 16-bit floating point, is a data format widely used in deep learning and machine learning, providing efficient representation and calculation of numerical values. While full support for FP16 is underway, various indexes in the indexing layer require converting FP16 to FP32 during construction.

We will fully support FP16, BF16, and int8 data types in later versions of Milvus. Stay tuned.

## Significant improvement in the rolling upgrade experience - seamless transition for users

Rolling upgrade is a critical feature for distributed systems, enabling system upgrades without disrupting business services or experiencing downtimes. In the latest Milvus releases, we’ve strengthened Milvus’s rolling upgrade feature, ensuring a more streamlined and efficient transition for users’ upgrading  from version 2.2.15 to 2.3.3 and all later versions. The community has also invested in extensive testing and optimizations, reducing query impact during the upgrade to less than 5 minutes, providing users with a hassle-free experience.

## Performance optimization

In addition to introducing new features, we’ve significantly optimized the performance of Milvus in the latest two releases.

-   Minimized data copy operations for optimized data loading
    
-   Simplified large-capacity inserts using batch varchar reading
    
-   Removed unnecessary offset checks during data padding to improve recall phase performance.
    
-   Addressed high CPU consumption issues in scenarios with substantial data insertions
    

These optimizations collectively contribute to a faster and more efficient Milvus experience. Check out our monitoring dashboard for a quick glance at how Milvus improved its performance.

  

## Incompatible changes

-   Permanently deleted TimeTravel-related code.
    
-   Deprecated support for MySQL as the metadata store.
    

Refer to the [Milvus release notes](https://milvus.io/docs/release_notes.md) for more detailed information about all the new features and enhancements.

## Conclusion

With the latest Milvus 2.3.2 and 2.3.3 releases, we're committed to providing a robust, feature-rich, high-performance database solution. Explore these new features, take advantage of the optimizations, and join us on this exciting journey as we evolve Milvus to meet the demands of modern data management. Download the latest version now and experience the future of data storage with Milvus!

## Let’s keep in touch!

If you have questions or feedback about Milvus, join our [Discord channel](https://discord.com/invite/8uyFbECzPX) to engage with our engineers and the community directly or join our [Milvus Community Lunch and Learn](https://discord.com/invite/RjNbk8RR4f) Every Tuesday from 12-12:30 PM PST. You’re also welcome to follow us on [Twitter](https://twitter.com/milvusio) or [LinkedIn](https://www.linkedin.com/company/the-milvus-project) for the latest news and updates about Milvus.

Revealing Milvus 2.3.2 & 2.3.3: Support for Array Data Types, Complex Delete, TiKV Integration, and More



![](https://assets.zilliz.com/Ingesting_Chaos_20231017_110103_54fe2009cb.png)


Data is being generated faster than ever before in every form imaginable. This data is the gasoline that will power a new wave of artificial intelligence applications, but these productivity enhancement engines need help ingesting this fuel. The wide range of scenarios and edge cases surrounding unstructured data makes it challenging to use in production AI systems.

For starters, there are a vast number of data sources. These export data in various file formats, each with its eccentricities. For example, how you process a PDF varies greatly depending on where it comes from. Ingesting a PDF for a securities litigation case will likely focus on textual data. In contrast, a system design specification for a rocket engineer will be full of diagrams that require visual processing. The lack of a defined schema in unstructured data further adds complexity. Even when the challenge of processing the data is overcome, the issue of ingesting it at scale remains. Files can vary significantly in size, which changes how they are processed. You can quickly process a 1MB upload on an API over HTTP, but reading in dozens of GBs from a single file requires streaming and a dedicated worker. 

Overcoming these traditional data engineering challenges is table stakes for connecting raw data to [LLMs](https://zilliz.com/glossary/large-language-models-(llms)) via [vector databases](https://zilliz.com/learn/what-is-vector-database) like [Milvus](https://github.com/milvus-io/milvus). However, emerging use cases such as performing semantic similarity searches with the help of a vector database require new processing steps like chunking the source data, orchestrating metadata for hybrid searches, picking the suitable vector embedding model, and tuning the search parameters to determine what data to feed to the LLM. These workflows are so new that there are no established best practices for developers to follow. Instead, they must experiment to find the correct configuration and use case for their data. To speed up this process, using a vector embedding pipeline to handle the data ingestion into the vector database is invaluable. 

A vector embedding pipeline like [VectorFlow](https://github.com/dgarnitz/vectorflow) will connect your raw data to your vector database, including chunking, metadata orchestration, embedding, and upload. VectorFlow enables engineering teams to focus on the core application logic, experimenting with the different retrieval parameters generated from the embedding model, the chunking strategy, the metadata fields, and aspects of the search to see what performs best. 

In our work helping engineering teams move their [retrieval augmented generation (RAG)](https://zilliz.com/use-cases/llm-retrieval-augmented-generation) systems from prototype to production, we have observed the following approach to be successful in testing the different parameters of a RAG search pipeline:

1. Use a small set of the data you are familiar with for speed of iteration, like a few PDFs where they have relevant chunks for the search queries.
2. Make a standard set of questions and answers about that subset of the data. For example, after reading the PDFs, write a list of questions and have your team agree on the answers.
3. Create an automated evaluation system that scores how the retrieval does on each question. One way to do this is to take the answer from the RAG system and run it back through the LLM with a prompt that asks if this RAG result answers the question given the correct answer. This should be a “yes” or “no” answer. For example, if you have 25 questions on your documents, and the system gets 20 correct, you can use this to benchmark against other approaches.  
4. Ensure you use a different LLM for the evaluation than you used to encode the vector embeddings stored in the database. The evaluation LLM is typically a decoder-type of a model like GPT-4. One thing to remember is the cost of these evaluations when run repeatedly. Open-source models like Llama2 70B or the Deci AI LLM 6B, which can run on a single, smaller GPU, have roughly the same performance at a fraction of the cost.
5. Run each test multiple times and average the score to smooth out the stochasticity of the LLM.

Holding every option constant except one, you can quickly determine which parameters work best for your use case. A vector embedding pipeline like VectorFlow makes this especially easy on the ingestion side, where you can quickly try out different chunking strategies, chunk lengths, chunk overlaps, and open-source embedding models to see what leads to the best results. This is especially useful when your dataset has various file types and data sources that require custom logic. 

Once the team knows what works for its use case, the vector embedding pipeline enables them to quickly move to production without having to redesign the system to consider things like reliability and monitoring. With technologies like VectorFlow and [Milvus](https://zilliz.com/what-is-milvus), which are open-source and platform-agnostic, the team can efficiently test across different environments while complying with privacy and security requirements. 





A Small Step for Milvus - 10,000 Stars

Special event

On This Page

Try Managed Milvus for Free

Keep Reading

Introducing Milvus Lite: Start Building a GenAI Application in Seconds

Milvus 2.3.4: Faster Searches, Expanded Data Support, Improved Monitoring, and More

Revealing Milvus 2.3.2 & 2.3.3: Support for Array Data Types, Complex Delete, TiKV Integration, and More

Ingesting Chaos: The MLOps Behind Handling Unstructured Data Reliably at Scale for RAG