RAG MCP for Cline

[wodecki]

Pre-submission Checklist

• I have checked that this question would not be more appropriate as an issue in a specific repository
• I have searched existing discussions and documentation for answers

Question Category

• Protocol Specification
• SDK Usage
• Server Implementation
• General Implementation
• Documentation
• Other

Your Question

I am using Cline (https://github.com/cline/cline) with Claude 3.5 Sonnet to support my projects realized with fast-changing libraries like LangGraph.
Unfortunately, Claude is not trained on its current API/docs. I would like to configure Cline in a way that when asked to create/correct my codebase, it would consult actual docs.
Do you have an idea how to achieve this?
I was thinking about a custom MDP server with a RAG system based on these docs. I can see two options:

• offline: I periodically download and index the library, and MDP has it "at hand"
• online: it directly connects to online docs.

For the first case, a very simple reference RAG can be found e.g. here:
https://github.com/langchain-ai/intro-to-langsmith/tree/main/notebooks/module_0

I would appreciate Your help both in the high-level approach, and implementation.

[Mehdi-Bl]

When no using websearch to get the docs on the fly.
With RAG you will end up facing same issue, aside of doing all the indexing. They key will be ranking and fetching the best result. Google/Bing do that already for you. Or you can past the URL ask the MCP to fetch it.
Some time simple solutions work file.

Otherwise if you want to build a RAG, what is the blocker? MCP is only a wrapper over the tools you have. Check the SDK And examples:

https://github.com/modelcontextprotocol/python-sdk

https://github.com/modelcontextprotocol/python-sdk/tree/main/examples/servers/simple-tool

I thinkn the examples are already simple.

[PederHP]

I think it's an important to distinguish between (classic) RAG and tool-based (sometimes called agentic) RAG. The former doesn't use tools, but is based on the application post-processing user messages and injecting context (Augmenting them as the per the A) with data retrieved (as per the R) from somewhere (could be a vector indexed db or some other source).

This happens at the application-level, not the tool level. MCP sort of support this via Resource Templates, but it's very awkward. I have an open proposal PR to add a capability specifically for this kind of application layer injection.

Tool RAG is a very different kind of beast, as it relies on the model deciding when to retrieve additional context and is usually not based on embeddings and similarity search. It also adds latency and higher token expenditure compared to classic RAG. It is still the superior choice for some use cases. I believe some of the agentic SE applications use agentic search / RAG as a core feature.

But classic RAG doesn't really have a good fit in the current MCP capability landscape. You can retrofit it pretty easily by adding your own methods to the server - but that doesn't provide hosts like Cline with any way to use them.

[Mehdi-Bl]

Why using Resources, this is static. Could be used for DOCS or similar what they do with projects.
But if you want dynamic injection best reading files and pointing to files to read.
Classic RAG once filled face similar issues with ranking and extracting relevant informations. While if you are pointing what is needed and let explore more works fine.

This is mostly tools and I think the team adding PROMPTS/RESOURCES made MCP confusing a bit bloated.
Tools are the core and the most important part. extra like prompts are useful and should eventually split aside.