Tutorials on Building Deep Research with Open source, Multimodal Semantic Search, and More!
In this issue:?
?? Build Deep Research with Open Source
OpenAI ’s Deep Research just dropped, promising AI-powered synthesis of complex topics. But what if you could build your own version running locally and fully open-source?
In this tutorial, we explore a DIY research agent that can:
?? Reason & Plan: Break down complex questions into subtopics
?? Search Wikipedia: Retrieve relevant info using Milvus for vector storage
?? Synthesize Reports: Use DeepSeek AI R1 + LangChain for structured summaries
?? Example: Asking “How has The Simpsons changed over time?”—our agent refines the query, retrieves insights from Wikipedia, and compiles a structured research report.
?? Why does this matter? Open-source AI gives flexibility & control for academia, content creation, or next-gen assistants. Future iterations could integrate real-time web search, reflection, and multi-step reasoning.
Building on top of this idea, we are open-sourcing an implementation of self-reflection and search that works with any LLM and data sources! It can search both public web and your private knowledge-base on a vector database like Milvus / Zilliz Cloud.?
Run an agent like Deep Research in your local laptop: GitHub - zilliztech/deep-searcher: Deep Research Alternative to Reasoning About Private Data
The tool works with:
Embedding models: OpenAI & Voyage AI (part of MongoDB)
LLM Services: OpenAI, DeepSeek, SiliconFlow Together AI
PDF Parsing: unstructured.io
?? Multimodal Semantic Search with Images and Text
Humans interpret the world through multiple senses. Why shouldn’t AI do the same? To truly match human understanding, AI must process text, images, and context together.
In this tutorial, we explore multimodal semantic search, showing how AI can connect words and visuals to improve search accuracy. We’ll build a retrieve-and-rerank search app that goes beyond keywords, using:
? Milvus for vector storage
? Visualized BGE for text-image embeddings
? Phi-3 Vision for reranking results
?? Example: Searching for a leopard print phone case using both text (“a phone case with”) and an image of a leopard.?
?? What’s next? Multimodal AI is unlocking new possibilities—from e-commerce to scientific research. And with open-source tools, anyone can build and experiment.
?? Check out the video walkthrough to see it in action!
Get Started: Multimodal Semantic Search with Images and Text?
?? Why DeepSeek V3 is Taking the AI World by Storm
Big news in AI! DeepSeek V3 is taking the AI world by storm, delivering GPT-4-level performance at a fraction of the cost. Here’s why everyone’s talking about it:
?? Smarter, Faster, More Efficient DeepSeek V3 introduces Multi-Head Latent Attention (MLA) to speed up processing and reduce memory use. Faster responses, lower compute costs. ??
?? Massive Power, Minimal Waste Thanks to Mixture of Experts (MoE), the model only activates the parameters it actually needs. Think of it as AI with a built-in efficiency mode! ?
领英推荐
?? Predicting the Future (Well, Almost) ?? DeepSeek V3 doesn’t just predict one word at a time, it uses Multi-Token Prediction (MTP) to generate multiple tokens in parallel. That means smoother, more natural responses.
?? It’s Cheaper While GPT-4 cost an estimated $100M+ to train, DeepSeek V3 pulled it off for just $5.6M.
?? Fully Open-Source & Ready to Use Unlike closed models, DeepSeek V3 is MIT-licensed, meaning developers can tinker, test, and build freely.?
Milvus X DeepSeek: Build RAG with Milvus and DeepSeek?
???Unstructured Data Podcast
We have a podcast! Listen to our first episode all about the AI Agent Revolution with Zilliz Developer Advocate Stephen Batifol and host Stefan Webb . They dive deep into the evolving world of AI agents—what they are, how they work, and their growing impact on technology and society.
??? Upcoming Events
Feb 20: Deploying a Multimodal RAG System Using Open Source Milvus, LlamaIndex, and vLLM (virtual)?
Discover how to build multimodal RAG systems using open-source tools! Learn to process images, audio, and text together using Milvus, LlamaIndex , and vLLM.
??Join our live demo showcasing:
Feb 26: San Francisco Unstructured Data Meetup (in-person)
Join us at the AWS GenAI Loft in San Francisco for our first Bay Area Unstructured Data Meetup of 2025! We look forward to exciting talks about the latest AI innovations and more.?
??? Arnab Sinha & Jean Malha will speak about Bedrock’s Latest feature: Bedrock Data Automation (Preview) to simplify processing unstructured data
?? Stefan Webb will be speaking on Combining Lexical and Semantic Search with Milvus 2.5.?
?? Anushrut Gupta from Hasura will speak about pushing AI's accuracy on unstructured data to 100% with PromptQL
March 11: Product Demo: Discover the Power of Zilliz Cloud (virtual)?
Join our monthly demo for a technical overview of Zilliz Cloud, a highly scalable and performant vector database service for AI applications
This webinar is an excellent opportunity for developers to learn about Zilliz Cloud's capabilities and how it can support their AI projects. Register now to join our community and stay up-to-date with the latest vector database technology.
?? Stay Connected
?? Join a community of developers on Discord?
?? Check out our latest articles on the Zilliz blog
?? Subscribe to this Milvus LinkedIn Newsletter for weekly updates