If someone is tired of “hello world” Python scripts that crumble the moment they hit production, this presentation will show how modern Retrieval-Augmented Generation (RAG) is being built using Golang, Onyx, and PostgreSQL.
This is a look under the hood at how a multi-agent Retrieval-Augmented Generation (RAG) system is actually built to handle hundreds of gigabytes of data without breaking a sweat.
In this session, Shahin Salehi from SEB dives into a pragmatic, robust architecture for multi-agent systems that moves beyond the hype and into production-ready code.
If one ever moved a Jupyter notebook into production only to watch it crumble under memory leaks or messy document structures, this video is the reality check for the solution.
Why Your Current RAG Might Be Failing
Standard RAG systems often rely on “one-size-fits-all” embeddings. When one deals with complex business data or giant PDFs, a simple semantic search usually returns more noise than signal. We dive into why ColBERT is a game-changer for retrieval accuracy. By encoding every word in a query, it finds the specific nuances that standard models miss.
The Tech Stack for Grown-Ups
The move is away from the “all-Python” ecosystem to show a more robust, production-ready architecture:
- Golang for Stability: Why a data scientist chose Go for the tool backends to ensure memory safety and high-concurrency performance.
- The Postgress Advantage: From specialized GPU databases to using pgvector and HNSW indexing to run complex SQL logic combined with vector search on massive datasets.
- The Inference Module: A look at running Onyx Runtime on CUDA, allowing to run state-of-the-art embedding and re-ranking models on own hardware.
Solving the “Real-Time” Problem
How do you serve 100 users at once when each agent call might take several seconds? The video explains a broker-based architecture using Server-Sent Events (SSE). This separates the query request from the response, creating that “typing” AI experience users love while keeping the backend separate and resilient.
Ingesting the Mess: Tables, Layouts, and RabbitMQ
Data ingestion is the unglamorous part of AI that makes or breaks the system. The video contains more about:
- Docling: A powerful new tool from IBM Research that handles table layout analysis better than almost anything on the market.
- RabbitMQ Pipelines: Building an offline ingestion journey where files are chunked and embedded asynchronously, ensuring that if a container crashes, the data doesn’t disappear.
Whether an AI-driven person is an architect looking for a more stable stack or it is about a developer tired of fighting with document parsers, there are code examples and architectural patterns here that they can try out at home tonight.
You can get the architectural breakdown and code examples you can run on your own server today.
