The Cloud Shuttle team were chuffed to be part of Neo4j’s GraphSummit World Tour 2024. This year, the Australian leg of the tour saw close to 150 attendees gather at Beta Events in Sydney on Thursday, 9 May for a jam-packed day full of keynotes, sessions, workshops and panels around how Aussie and Kiwi industry leaders are applying graph technologies and AI models to solve real-world business problems, from optimising supply chains to accelerating the discovery of life-saving drugs. 

Keynote

Peter Phillip, General Manager A/NZ of Neo4j, opened the summit with a fantastic keynote, illustrating how graph technology is far more integrated in our daily lives than we realise. Recounting his commute into the city, he shared how so much of what he encountered on his journey – from the rerouting of train schedules to the fraud detection systems used by credit card companies for ticket purchases – relies on graph databases to perform tasks that traditional databases struggle with.

The keynote set a fitting tone for the rest of the summit, which was all about the integral and innovative real-world use cases of graph databases being implemented by some of the region's most well-known organisations today. 

The Cloud Shuttle team have put together a recap of some of our key takeaways from the day.

Commonwealth Bank of Australia introduces GraphIT, their network digital twin

One of the standout presentations of the summit was by Stephanie Ranft (Data Scientist) and Jon Whitear (Principal Engineer) from the Commonwealth Bank of Australia. 

They unveiled GraphIT, the digital twin of their network infrastructure and workloads. The team created it to accelerate and simplify the exploration of the network events and configuration. Using GraphIT’s graph database for Retrieval-Augmented Generation (RAG) in conjunction with an LLM allows for interactive communication to solve real-time network security and architecture problems. 

The tool is the result of the collection of near-time telemetry data and ServiceNow CMDB, and allows the exploration of physical and virtual network information. 

It was cool to see how the project team has conducted the experimentation in an iterative way, starting from simple prompt engineering with zero-shot prompt, with several iterations along the way to its current state of using RAG to add real-time and domain knowledge to the application. 

Project history

• v0: Initiated with FewShots prompts varying in style and Cypher queries.
• v1: Integrated llama2 with Langchain, introduced tools and agents, and added LLM memory capabilities.
• v2: Focused on optimising query examples, transitioned from CamelCase to snake_case, and ceased using Langchain in favor of a custom framework.
• v3: Implemented dynamic prompting and semantic layers.

Looking forward, the team plans to create a separate database for examples, log user interactions, and refine the balance between context and LLM optimisation. They have chosen to maintain the foundational model without fine-tuning to stay aligned with core functionality.

Generative AI reset by Nicolas Hohn

Next up, we had Nicolas Hohn, Distinguished Data Scientist at McKinsey & Company in Melbourne and the Chief Data Scientist for QuantumBlack Australia, who gave a talk on how we need a reset when it comes to GenAI.

The initial enthusiasm and flurry of activity in this space is giving way to second thoughts and recalibrations as companies come to realise that capturing the enormous potential of GenAI’s value is harder than expected. Launching pilots is (relatively) easy; getting pilots to scale and create meaningful value is hard because they require a broad set of changes to the way work actually gets done.

Because a large swathe of roles in every sector will be impacted in some way, he emphasised the importance of keeping humans in the loop. This means involving subject matter experts and data scientists to optimise outcomes and manage change. He estimates that for every $1 invested in technology, another $4-5 needs to be invested in change management. (Even the most clever bot will be useless if no one uses it!) His talk was a crucial reminder of the challenges and steps needed to manage and make the most of disruptive innovations.  

SparkNZ: Creating a connected data-driven operating model

After morning tea, Anshuman Banerjee, SparkNZ's Tribe and Chapter Area Lead, shared their journey of how they used GenAI and graph databases in two key company initiatives: 1) a graph database to automate and improve their response to RFPs (Request for Proposals); and 2) a company digital twin to enable deep graph analysis and cross-domain simulation. 

Automated pre-filled RFP system

As a telco, SparkNZ answers hundreds of RFPs, many of these often ending up as redundant work with wasted effort due to the large amount of information that needs to be collected. The company decided to build a virtual assistant capable of assisting a Sales engineer in completing a new RFP based on the knowledge of previous answers. This means less ‘rebuilding of the wheel’ and freeing up the sales engineers to focus on where they can add more value.

Initially, SparkNZ used vector search, having developed vector embeddings of their historical documents. While it was able to provide answers when there was a large amount of available information, it came up short when information was sparse among several anchored sections of documents. To overcome this, they created an ontology of their documentation by creating relationships between document sections. SparkNZ’s sales engineers can now do fast search with an initial vector search that returns a section node, which can then be used to retrieve related sections to enhance the LLM context and provide more complete and accurate results. 

Company digital twin

Like Commonwealth Bank, SparkNZ also showcased their company digital twin at the summit. Thanks to the implementation of a graph database across several domains at SparkNZ, the company can lead with deep graph analysis and perform cross-domain simulations, such as simulating customer responses to price changes by using customer, sales and marketing data. Their decision engine is aptly named BRAIN, and it was really interesting to see the different applications of digital twins in Commonwealth Bank (networking) and SparkNZ’s (sales and marketing) settings, showing the great breadth and promise of graph database implementations.

Neo4j product vision

One of the benefits of going to a GraphSummit is getting some early insights into Neo4j’s roadmap. Emil Pastor,  Solution Architect Manager A/NZ at Neo4j, painted a product vision, which largely focused on:

• Security (Customer Managed Key for encryption, SecGovCompliance features)
• Scalability (Cloud first, autonomous Cluster, sharding)
• Data (CDC, Timetravel, GQL is now an ISO-standard!)

As we see graph database technology increasingly adopted, we can see how Neo4j is beefing up security features, scalability, and increased compliance capabilities to meet the growing data security and operational efficiency needs of Aussie and Kiwi businesses.

Building Generative AI on Google Cloud

Robert Sibo, the Head of AI/ML Customer Engineering ANZ at Google Cloud, discussed the rapid evolution of GenAI alongside ongoing advancements in traditional ML. During his presentation, Robert introduced Google's graph and GenAI capabilities, noting Google's reliance on Neo4j for graph technology since it does not offer its own graph database solutions to customers.

He also detailed Google's development of features that align with those offered by Microsoft and AWS in the GenAI ecosystem. These include state-of-the-art models, Responsible AI initiatives with built-in guardrails, and integrations with open-source platforms and Huggingface. He also touched on VertexAI, Google’s service for building AI agents and models, illustrating Google's approach to making AI more accessible and manageable for developers and businesses.

Q&A panel

Just before lunch, Peter Philip moderated a panel with the summit speakers Anushman Banerjee (SparkNZ), Robert Sibo (Google Cloud), Jon Whitear (Commonwealth Bank), Stephanie Ranft (Commonwealth Bank) and Nicolas Hohn (McKinsey & Company, QuantumBlack Australia). 

The discussion opened with how to efficiently transition from SQL to GraphQL. Panellists advised feeding the SQL schema directly into GraphQL or starting anew from an Entity-Relationship Diagram (ERD), focusing on relationships rather than relational structures. When asked about developing graph-based applications, the consensus was to avoid overcomplicating the process: start with the core need, implement, and then iterate based on feedback.

The panel also shared key tips on how to avoid pitfalls when implementing GenAI projects. It's important to involve people in the process and start with easy, "low-hanging fruit" projects to demonstrate value early on. The rollout should involve collecting feedback, performing acceptance tests, and then refining and expanding the implementation based on these insights. 

Workshop highlights

After the lunch break, the summit broke into two parallel workshops for the rest of the afternoon:

1) Architecting innovative graph applications; and

2) Enabling GenAI breakthroughs with knowledge graphs.

The first was designed for beginners and walked participants through the step-by-step development of a graph solution using a real-life dataset. Because the Cloud Shuttle team has had some experience with building graph applications (read more about chatbot prototype here!), we decided to focus on the second workshop.

Enabling GenAI Breakthroughs with Knowledge Graphs

The second workshop, "Enabling GenAI Breakthroughs with Knowledge Graphs," was all about practical techniques to enhance the accuracy, transparency, and explainability of GenAI systems. This session was where participants could get hands-on experience in integrating relationships and Large Language Models (LLMs) for domain-specific capabilities. In this workshop, we had a use case of tailoring marketing content to customer interests and purchase histories. 

During the workshop, we went through the following process:

1. We created text embeddings from product descriptions, leveraging Neo4j’s vector index for efficient vector search and semantic similarity matching.
2. We then integrated the index with LangChain, a popular framework for developing applications with large language models to simplify the process of vector search and retrieval from Neo4j.
3. We applied graph patterns in Cypher (Neo4j's query language) to enhance semantic search by incorporating detailed customer behaviour and purchase history.
4. We stored the node embeddings within the nodes themselves. The embeddings could later be retrieved and used to cluster nodes and identify the most similar ones using algorithms like KNN. 
5. We then created graph embeddings using customers and products. The embeddings created matched customers and the products they frequently bought to create customer-specific recommendations. 

The workshop was not only great fun, but provided our team with a practical, enriched understanding of Neo4j's capabilities and gave our team even more hands-on experience on how knowledge graphs can be applied to solve complex challenges in various domains.

Learn about how Cloud Shuttle built a RAG-LLM chatbot for the DataEngBytes conference in this blog post and share your feedback on what features you'd like to see in our chatbot!

Expo hall

As you may know, the hallway track is among the team's highlight of any summit or conference. Thanks to our partnership with Neo4j, Cloud Shuttle was able to proudly showcase our graph database prowess at the GraphSummit Sydney expo. Our booth featured a demo of the DataEngBytes chatbot built on Neo4j, Langchain, and Amazon Bedrock.

The chatbot is specifically designed to assist attendees and potential participants by providing essential information like event dates, times, venues, and details about past company participants. The response was overwhelmingly positive, with lots of great feedback from attendees to help us iterate on the next phase of the project.

Conclusion

All in all, it was an incredible event of learning, hands-on experience, networking and relationship-building. If you get the chance to attend a future GraphSummit, we couldn't recommend it more - it's the perfect opportunity to get your graph on!

If you're looking to harness the power of graph solutions to support your business, Cloud Shuttle has the expertise to help. Reach out to us for a no-obligation, complimentary 30-minute discovery call to explore how we can help your business to leverage GenAI to innovate!