chintn/h7brid-ai

Hybrid Repo AI Agent is an experimental Python project that tries to make AI coding assistants work better across an entire codebase, not just one file at a time. The author built it after noticing that local AI models slow down and become unreliable once asked to reason about a whole repository, mainly because they cannot easily figure out which files matter or how those files connect to each other. To address this, the project adds several layers between a user's prompt and the AI model. It indexes the repository and builds embeddings so it can search for relevant code semantically, then expands that search using a graph of how files relate, followed by a reranking step meant to prioritize the best matches. A context builder assembles this information before sending it to either a local model, run through Ollama, or a cloud model, accessed through OpenRouter, depending on how complex the task looks. Simple requests are meant to use the faster local model, while complex ones needing deeper reasoning go to the cloud. The system also includes an editing pipeline that can generate patches across multiple files, a diff viewer to show proposed changes, and a memory component meant to give the agent some lasting understanding of the repository over time. A debug tracing pipeline is included for following what the system is doing internally. Setup requires Python 3.10 or later, cloning the repo, creating a virtual environment, installing dependencies from a requirements file, installing Ollama and pulling a local model such as qwen2.5-coder, and adding an OpenRouter API key to a .env file. On first run, the tool scans the repository, builds embeddings, and creates its indexes, which can take a while depending on repository size. The README describes this plainly as an experimental, actively developed prototype, not a finished product. The author lists planned future work including adaptive retrieval, AST-aware patching, a planner-agent system, and a proper vector database, which suggests the current version is closer to a research sketch than a stable, ready-to-use tool.