If you're building with agents and graphs, we'd love to hear from you, so read the post, share around, and build more with @kuzudb!. 4/4.

For most teams, context engineering and managing context is usually one of the hardest parts of building an agent. As more and more people build agents, we think that graphs, and graph databases like @kuzudb, are a natural fit, and for short and long term memory, and more. 3/4.

These days, "context engineering" has become a term of increasing importance. Put simply, it's the practice of packing just the right context for models during the various stages of operation in an agentic workflow. 2/4.

Read this new blog post from our CPO, @ArdanArac to learn "why knowledge graphs are critical to agent context".🚀. @kuzudb is excited to follow along the latest developments with agents and actively develop features to support this growing ecosystem!. 1/4.

The entire @kuzudb engineering team has worked very hard on putting these great research ideas into practice, so your graph and vector queries can run in one system, faster and more efficiently. 🚀. See you in London! 🇬🇧.

Over the coming weeks, we will be publishing detailed technical blogs about the algorithm and discuss other implementation details. In the meantime, we encourage you to read our paper to learn more.

What's remarkable is that it went from academic research into a usable solution for industry in relatively short time -- the implementation is already in @kuzudb's core. You can use it with your graph query workloads right away by installing the latest release of Kuzu!.

📣 @kuzudb is thrilled to announce our research work on "NaviX": a new on-disk HNSW vector index has been accepted to @VLDBconf 🇬🇧! We'll be presenting a novel, efficient algorithm for vector search with arbitrary filtering on top of a graph-based HNSW index.

By grounding the workshop in practical examples and real tools, we hope this hands-on session helps you design & build agentic systems that are robust, interpretable, and adaptable to complex enterprise environments!🚀.

Join our AI engineer @tech_optimist and many other builders at the @_odsc Agentic AI summit (online) in July! We'll be covering how to build intelligent agent workflows powered by Graph RAG and production-ready knowledge graphs. Registration link here:.

Here's wishing everyone a happy Canada Day 🇨🇦 and a great year building awesome things with @kuzudb!. Kùzu Inc. was incorporated in early 2023, in Waterloo, Ontario, Canada. From small beginnings, today, we have thousands of users, from all over the globe. Onward & upward 🚀.

Thank you for following our journey and helping us reach 1k followers here on X! The next several months are incredibly exciting for @kuzudb, and there will be loads more updates and great new features. Please spread the word, and we appreciate your continued usage of Kuzu!.

If you're abuzz with excitement about doing Graph RAG with @DSPyOSS, our amazing community member @arundsharma has got you covered! Check out the GitHub repo in the linked post, and give it a try!👇🏽.

This post shows the beginnings of an agentic Graph RAG workflow that's more robust to a broader range of user questions. No agent frameworks were used - only BAML prompts under the hood. @boundaryML's BAML make prompt engineering a breeze! 👇🏽. 9/9.

With the router agent (that picks the right vector search tool vis the BAML prompt), the results are significantly improved from the baseline!. 8/9

How do we know we improved the results from vanilla Graph RAG (i.e., a single pass at Text2Cypher)? We evaluate our router on a suite of 10 queries that ask a variety of questions to the @kuzudb database! Regardless of the LLM used, most tests fail to produce a response. 7/9

Retry logic is added using a while loop in case the wrong tool was picked. If the system still cannot return a valid response from the Cypher query after 2 tries, it gives up and the workflow ends. Along the way, the #BAML prompts are engineered to provide helpful responses. 6/9

How does routing work? There's a `PickTool` prompt defined in BAML that chooses one of the vector indices to query for similar terms, should an initial pass at Text2Cypher fail to return a response. All the queries are run on the @kuzudb database. 5/9

The vector search "tool" is a Python function under the hood that queries the vector index in @kuzudb. The function is shown below. This is exposed via a REST API to the frontend UI. 4/9

We first create multiple vector indices in @kuzudb, on the `Condition` and `Symptom` node tables, and expose these functions as tools that can be called by an LLM. 3/9