Beam Summit’s 2026 program lists a June 22 session by Samir Sengupta on using Apache Beam with large language models and retrieval-augmented generation for real-time inference on high-velocity data streams.
Beam Summit has listed a June 22, 2026 session on embedding large language models into Apache Beam for live inference on fast-moving data streams.
According to the official Beam Summit program, the session is titled “Real-Time AI Pipelines at Scale: Embedding LLMs into Apache Beam for Live Inference.” The Beam Summit session page names Samir Sengupta as the speaker and says the talk will focus on bringing large language models and retrieval-augmented generation systems directly into Apache Beam for real-time inference.
Beam Summit describes the session as relevant to production AI systems that need low-latency responses while handling high-velocity data. The official program says the talk will cover Hugging Face and vLLM inference, retrieval-augmented generation, vector databases including Pinecone and FAISS, batching, GPU optimization, AWS Bedrock, SageMaker, and Kubernetes orchestration.
The Beam Summit listing indicates that the session is aimed at teams working with live data streams and generative AI models. In that setting, the technical challenge is not just sending a request to a model. Teams also need to manage response time, throughput, cost, and reliability while data continues to arrive.
The official session description names Hugging Face and vLLM as inference technologies, suggesting that the talk will address serving options for open or self-hosted models. Beam Summit also lists AWS Bedrock and SageMaker, indicating that managed cloud AI services will be part of the discussion as well.
Beam Summit’s materials also mention retrieval-augmented generation with Pinecone and FAISS. That points to systems in which model outputs are supported by retrieved context from vector databases. This approach is commonly used when organizations want model responses to draw on enterprise, product, or domain-specific information rather than relying only on a model’s existing parameters.
The official program specifically lists batching and GPU optimization among the session topics. Those are practical concerns for organizations running model inference under variable workloads. Batching can improve hardware use, while GPU tuning can affect both latency and infrastructure cost.
Kubernetes orchestration is also included in the Beam Summit program description. In production environments, orchestration is often used to manage deployment, scaling, and reliability for services that run continuously. The Beam Summit pages do not provide implementation details, but the topic list indicates that infrastructure management will be part of the session’s scope.
Apache Beam is used to define data processing jobs that can run across multiple execution engines. Beam Summit’s session listing positions Beam as a framework where LLM inference and retrieval-augmented generation can be integrated into streaming data applications rather than treated only as separate downstream services.
The Beam Summit 2026 sessions index also includes the same session title, names Samir Sengupta as the presenter, and summarizes the Apache Beam live LLM inference topic. Based on the official Beam Summit materials, attendees can expect a technical discussion of how live data processing, model inference, retrieval systems, and cloud infrastructure can fit together in Beam-based applications.
The sources do not state that a new product, Apache Beam feature, or open-source release will be announced. They describe a scheduled conference session focused on design and operational practices for real-time LLM inference in Apache Beam.
Beam Summit has listed a June 22, 2026 session on embedding large language models into Apache Beam for live inference on fast moving data streams.
Beam Summit describes the session as relevant to production AI systems that need low latency responses while handling high velocity data.
What the session is expected to cover The Beam Summit listing indicates that the session is aimed at teams working with live data streams and generative AI models.
Continue reading
