TigerGraph has pitched a combination of dis-aggregated infrastructure and composite queries as a way for enterprises to get a grip on soaring costs as enterprises try to use Gen AI for data heavy applications.
The firm’s Savanna cloud offering will make it easier to deliver composite queries which cut the need for LLM calls, it claims. At the same time, Savanna allows users to scale compute, memory and storage separately, allowing them to deploy systems faster, and configure them more efficiently.
The launch comes as enterprises focus on the soaring cost of using AI systems and their actual return. Research by IBM last year claimed that the average cost of computing is expected to climb by 89% between 2023 and 2025, with gen AI being the key factor.
See also: This free tool uses eBPF to identify, slash cloud networking costs
Graph databases are very different from traditional databases. Rather than rows and columns, the focus is on “nodes” and edges. This makes them particularly suited for uncovering relationships and connections in a pile of data.
Enterprises use graph-based system to power recommendation systems or fraud detection, while investigative journalists have used graph-based systems to uncover relationships and track asset movements when investigating large amounts of leaked data.
However, TigerGraph CEO Rajeev Shrivastava told The Stack that the technology was increasingly useful for “supercharging” AI systems, by making LLM queries more efficient and increasing their explainability.
“That's where graph databases are extremely powerful, and that's where a lot of our customers use us.”
He cited the example of asking the question, “what are the features of phone X and the price of its compatible accessories?” That call would mean multiple queries, each of which would mean at least three to five LLM calls, with results having to be composed, validated, and checked for accuracy.
Using a graph database to create a composite query, “You'll create a relationship between phone X, the features and the accessories. And we will make a single API call to the LLM.”
That will reduce the cost overall, he said, as well as reducing the chance of hallucinations. “It’s about how can you do RAG better,” he said.
Shrivastava pitched further cost savings as the firm has overhauled its parallel processing architecture to allow compute and storage to be decoupled. This in turn means they can be scaled separately, depending on the sort of applications a customer is running.
Real time transactions, such as fraud checks, might be more compute intensive. “You’re not going to wait two minutes for a result to come back.
Together with other elasticity features, including scheduled expansion/shrink and auto stop/resume, the firm reckons this can deliver cost savings of 25%.
The service is initially available on AWS, with GCP and Azure in the pipeline. Savanna is offered as a fully managed service, or as a bring your own cloud service.
However, setting up graph database technology, never mind hitching it up to an LLM, is not exactly trivial, and TigerGraph has launched nine further solution kits, including for Transaction Fraud, Application Fraud, Mule Account Detection, and Supply Chain Management.
“We have worked with number of large banks and institutions on transaction fraud. There's no reason when we go to a new customer, they should be thinking about ‘How do I do it?’” He claimed the kits could slash deployment time from “from months to days or right away.”
Shrivastava said the firm had also invested in additional data sources, with connectors for Snowflake Spark, Delta Lake, Iceberg and Postgres, as well as AWS Azure and GCP’ object stores.
“These large customers have data sources in multiple places, and each of those investments are important for them.” So, he said, it just wasn’t on for a vendor to come in and say, “Now you got to go rip and replace or no, I can't connect to data x and data y”.
