Local-First, Actionable Triangulation and Automated Remediation: Mitigating Cloud Inference Costs via Multi-Modal Vector-Graph Retrieval

The adoption of Large Language Models (LLMs) in software engineering introduces privacy considerations and computational costs, particularly when relying on cloud-based application programming interfaces (APIs). Furthermore, traditional Retrieval-Augmented Generation (RAG) systems often function passively, summarizing information rather than executing code modifications. This paper presents a local-first, zero-telemetry system designed for actionable information retrieval, offline diagnosis, and automated remediation within distributed microservice architectures. By decoupling state representation from the LLM's memory window, the architecture ingests engineering data including version control repositories, CI/CD logs, container telemetry, and issue tracking tickets into a localized hybrid vector and graph database. A multi-agent orchestrator triangulates cross-service anomalies and translates retrieved findings into programmatic code changes. The system submits these cross-repository pull requests autonomously, reducing the overhead of cloud models and mitigating intellectual property exposure. To facilitate further research, the implementation is open-sourced at: https://github.com/dasarathirout/dasarathi-aistudio.