How do you turn siloed operational data into trustworthy context for interactive agents? That's one of the big questions we set out to answer with Bilt. Their live context layer is now in production, behind the agents and microservices that power their commerce network. Learn how an up-to-the-second context graph keeps RAG and vector search fresh, and lets their engineers ship in days what used to take months, on smaller and cheaper models. Read the full post: https://lnkd.in/gemj5YPE
How do you turn siloed operational data into trustworthy context for interactive agents? What would you encode in the data model itself to get more out of every token? These are some of the big questions we set out to answer as we partnered with Bilt to create a live context layer for the agents and microservices that make up their commerce network. It’s running in production and unlocking new experiences for millions of members along with new ways for their merchant partners to engage customers in real time. The architecture: - keeps RAG and vector search continuously fresh, incrementally updating only the documents affected by each human or agent action - maintains an up-to-the-second context graph that long-running agents can reason and act over - improves token efficiency at scale, letting them optimize their inference economics by blending in more cost-effective models Bilt’s engineers now ship faster: projects that once took months ship in days. Each new context building block expands their agents’ worldview and makes the next use case easier to build, creating a flywheel for new capability development. Here’s a deep dive into how they did it: https://lnkd.in/gemj5YPE