Software development revolves around two core activities: continuously seeking information and producing code. While coding has strong tool support, information seeking across distributed knowledge sources remains less systematically supported. Prior work shows that a major source of information seeking effort in software development arises from frequent switching between heterogeneous knowledge sources. Retrieval-Augmented Generation (RAG) offers a promising approach by grounding generative models in organization-specific knowledge, enabling more context-sensitive information access. However, many RAG implementations rely on cloud-hosted Large Language Models (LLMs), raising concerns in privacy-sensitive contexts where proprietary data cannot be externally processed. Moreover, such systems often assume infrastructure, computational resources, and specialized expertise, conflicting with resource-constrained organizational contexts shaped by limited budgets, as is typically the case in small and medium-sized enterprises (SMEs). Local Small Language Models (SLMs) mitigate these constraints but are more prone to hallucinations. Following a Design Science Research approach, this study contributes a nascent design theory for cross-source information seeking under these constraints. In particular, we propose a local-first, modular design that focuses on retrieval quality. The prototype demonstrates how AI-assisted information access can be realized while retaining control over model execution. An evaluation conducted in an SME indicates organizational feasibility and improved access to information.