Members-Only
Recent Talks & Demos are for members only
You must be an AI Tinkerers active member to view these talks and demos.
How great context selection leveled up Bismuth
Explore Bismuth, a terminal coding agent. This talk details its advanced context selection, file and symbol discovery, and graph RAG system for tackling large open-source projects like Django.
Bismuth is a coding agent that works right in your terminal. It can build projects from scratch, or work in well established projects to help boost productivity. Bismuth is able to automatically select the right context from projects (even some of the largest open source projects like Django), discover and run build commands on its own and complete development tasks without human interaction.
The demo will be a live running of our project and talk through what is happening step by step as things run live, no slides!
We plan to demo the file selection step, the code symbol selection step, how both of those steps seed our graph rag system for larger projects in our pipeline and how this all needs to come together to select the right file context for an agent to be able to implement a user’s change request for a large open source project like Django so people can get a feel for the steps needed to search and build context for coding agents tackling large code projects.
The audience will see visually our agent working on a system, side by side with the live backend logs showing all of these steps executing, including files selected, symbols selected, results from our rag system and then the execution and updates on the open source project with the ultimate result being a patch that applies the actual changes we’re going to ask it to make for the demo.
Then we’ll switch to our project and dive in to the specific implementation code in our IDE for each step to understand what design considerations we had to make to make that all happen. This includes considerations like, what model are we using to vectorize the code, what libraries are we using the build our graph rag system since it is all homegrown, were there any optimization issues we ran into (spoilers there were) and explaining wow are we doing all of our code analysis and how much of it is custom knowledge versus how far can you get off the shelf and what tools we can suggest their like TreeSitter.
This demo will include how we handle Bismuth’s context selection which includes a bunch of custom code parsing, file selection, AST breakdown and code symbol selection. This is a key piece of coding agents which all begin with selecting the right context based on the users request.
Python framework for AI agents: graph execution, LLM inference, caching, state.
Compose Email
Loading recent emails...