Skip to content

Braid Architecture.md

Latest commit

 

History

History
52 lines (42 loc) · 2.68 KB

Braid Architecture.md

File metadata and controls

52 lines (42 loc) · 2.68 KB
Raw

BRaid AI eDitor

The BRaid AI eDitor will be a an agentic command line editor similar in spirit to Aider and Claude Code but taking an agentic approach, and focusing on local LLM's that are run on the user's GPU, while maximizing its speed and context use.

Agentic Components

There will be a series of agents which work in concert to answer the user's query.

  • Orchestrator -- watches agents for crashes, failures, repetition, restarts, etc.
  • Architect (large LLM) -- develops the initial plan and tracks its progress using a TODO list
    • add TODO
    • complete TODO
    • re-evaluate plan with new information
  • Editor (medium-small LLM) -- actually writes code
    • may have multiple of these depending on the complexity of the task
  • Evaluator (medium-small LLM) -- evaluates whether the code written satisfies the tests
    • applies diffs
    • lints results
  • ContextManager
    • manages context chunks
    • content based addressing (deduplication)
    • save/restore (e.g. queries by Architect, Editor, Evaluator will often re-use the same chunks of code)
  • DocumentationFinder -- finds documentation using RAG or web searching
    • includes ChatBTC RAG
    • docstring/inline documentation extraction
  • RepositoryTracker -- keeps track of code including filesystem layout, git integration, and AST/GraphRAG
    • "What does this change affect?"
    • "Which file needs to be edited?"
    • "Who calls this function?"
    • "What's in scope to this function?"
    • "What's the function signature of this thing I need to call and what does it do?"

Simplifications

  • Use a standardized format for feeding code to avoid variance in the output. The code should be stripped of comments (they can be in a separate documentation context chunk) and formatted with a linter.
  • Use AST parsing to apply diffs, so as to be insensitive to "diff format".
  • Context Chunks, not prompts: select the context chunks you need and add instructions.

Evaluation

  • We need a way to find small models that work well
    • Start with simple tests (change line 3, change variable x to y etc)
    • Refine sampling parameters to remove repetition
    • Find where they fail
  • Evaluate standard coding benchmarks (exercism)
  • Find large-codebase evals
  • Find documentation/hallucination benchmarks

Questions

  • Using SGLang/vLLM, can KV cache chunks be moved between models? e.g. if we're running Qwen3:8b for orchestrating and Qwen3:0.6b for applying diffs, can they share context chunks?
    • I think so because I think speculative decoding uses it