研究室AI基盤の構築（第4回：研究室ナレッジについて考えていたら自作してしまった件（後半））The fourth installment of a lab AI infrastructure series details the…

The fourth installment of a research lab AI infrastructure series turns to a practical problem that many teams face: how to make accumulated project knowledge accessible to AI assistants in a controlled, self-hosted way. Building on earlier entries, this post documents the second half of the author's work on project-level knowledge management, focusing on the implementation details and day-to-day operation of a custom MCP server the author ended up writing themselves.

The series has followed a clear progression. The first installment introduced a student-facing AI hub designed to be easy to administer within a lab environment. The second covered embedding "good AI usage practices" directly into custom models, effectively encoding recommended workflows so that students are nudged toward productive habits. The third laid out the concept and indexing approach for managing knowledge on a per-project basis. This fourth entry closes that thread by moving from concept to a working system, showing how research labs can enable AI-assisted knowledge sharing across individual student projects.

At the center of the design is the Model Context Protocol, or MCP, which the tags and title identify as the core technology. MCP is an open standard for connecting AI assistants to external data sources and tools through a consistent interface. Rather than hard-coding integrations into each application, an MCP server exposes resources and actions that any MCP-compatible client can call. In this case, the author appears to have implemented their own MCP server so that the lab's knowledge base can be queried directly by AI tools, keeping the retrieval logic and data firmly under local control. That self-hosted stance is consistent with the article's tags, which include self-hosted alongside knowledge-management, rag, and mcp-server">mcp-server.

The presence of the rag tag signals that the system is likely built around retrieval-augmented generation. RAG is a widely used pattern in which a model's responses are grounded in documents fetched from an external store at query time, rather than relying solely on what the model learned during training. The typical pipeline involves splitting documents into chunks, converting them into vector embeddings, and storing them in an index so that the most relevant passages can be retrieved and supplied to the model as context. The third installment's emphasis on indexing fits this pattern, and this fourth post's focus on implementation suggests it describes how that index is wired into the MCP server and surfaced to AI clients during actual use.

Organizing knowledge by project is a meaningful design choice for a lab setting. Research groups often accumulate scattered material across experiments, code repositories, meeting notes, and papers, and a project-scoped structure keeps retrieval relevant to the task at hand while reducing noise from unrelated work. It also maps naturally onto how students collaborate, since each project can maintain its own body of documents that the AI can draw on. The article's stated goal of enabling knowledge sharing across student projects implies that this structure is intended to help newer members benefit from work done by others without manually digging through archives.

For readers weighing similar systems, it is worth noting the broader ecosystem around these components. MCP was introduced by Anthropic and has since seen growing adoption, with server implementations available for services ranging from file systems and databases to issue trackers and search tools. On the retrieval side, self-hosted options such as vector databases and embedding models make it feasible to keep data on local infrastructure, which is often a priority in academic environments where research data may be sensitive or unpublished. The choice to build rather than adopt an off-the-shelf tool, as the playful title suggests, is common when existing solutions do not quite fit an organization's specific workflow or governance requirements.

Taken together with the earlier posts, this installment rounds out a fairly complete picture of a lab-scale AI stack: an accessible hub for students, custom models that embed good practice, and a retrieval layer that grounds AI responses in the lab's own project knowledge. Readers should treat the specifics as one team's implementation rather than a universal blueprint, but the described approach offers a concrete, self-hosted reference for research groups considering how to combine MCP and RAG for internal knowledge management.