On Monday, 18 May, I hosted a workshop at the ch-open Open Source AI Workshops in Room C. The premise was a question: can a team with no template build a working RAG system in six hours, armed only with an open-source coding agent and an API specification? We set out to find out.
Four teams each took ownership of one core component of a retrieval-augmented generation system: an LLM inference server, a vector database, a document processor, and a chat frontend. Every team built its component from scratch, without finished frameworks, using only base libraries and opencode-ai as their copilot. At the end, we wired everything together live into a single working system.
The whole stack stayed open. We worked throughout with open technologies: opencode-ai as the coding agent, open-weights models such as DeepSeek or GLM-4, and a final system that ran on a local open-source LLM. No proprietary APIs, no hidden magic, just the moving parts of a RAG system laid bare.
The day was split into four sessions:
The workshop was for anyone who wants to understand how AI systems work under the hood and try AI-assisted programming with open-source tools. You did not need to be an experienced developer; the willingness to experiment mattered more. The only prerequisites were a laptop with a browser and a GitHub account, since everything else ran in a preconfigured GitHub Codespace.
Participants walked away with hands-on experience of vibe coding, a working understanding of the core components of a RAG system, and their own code to keep building on.
The experiment asked whether a team with no template could build a working RAG system in six hours using only an open-source coding agent and an API specification. Four teams each owned one core component and integrated their work live into a single running system at the end.
Four teams each built one core component of the RAG system from scratch: an LLM inference server, a vector database, a document processor, and a chat frontend. They used only base libraries and the opencode-ai coding agent as their copilot, with no finished frameworks.
The workshop used opencode-ai as the coding agent and open-weights models such as DeepSeek or GLM-4. The final integrated system ran on a local open-source LLM, so the entire stack stayed open and self-hosted.
It was aimed at anyone who wants to understand how AI systems work under the hood and try AI-assisted programming with open-source tools. No prior experience as a developer was required, only the willingness to experiment. Participants needed a laptop with a browser and a GitHub account; everything else ran in a preconfigured GitHub Codespace.
.