Hanan Mikael Ouazan

💻Building a self-corrective coding assistant from scratch with LangGraph Code generation / analysis is one of the most important LLM applications. @itamar_mar and team recently released AlphaCodium, which showed that code generation can be improved by using a flow paradigm to build answers iteratively through testing and reflection on test failures. We recently launched LangGraph to support flow engineering, giving the user the ability to represent flows as a graph. We show that LangGraph can be used to implement code generation with the same types of cycles and decisions points as the AlphaCodium work. To test this, we built a coding assistant for LangChain Expression Language (LCEL). We stuff the LCEL docs into the context window of GPT-4-128k and include checks for imports and execution w/ feedback in LangGraph. On a 20 question eval set, we see ~47% improvement in code execution success rate w/ LangGraph vs w/o. We provide this as a notebook that can be easily adapted for any codebase or documentation of interest. Video: youtu.be/MvNdgmM7uyc Code: github.com/langchain-ai/lang… Blog: blog.langchain.dev/code-exec… Reference to flow engineering w/ AlphaCodium: xitter.gq/karpathy/status/…