(page_rag)=

## PETSc RAG MCP server

<http://mcp.petsc-ai.org:8082/mcp> or  <https://mcp.petsc-ai.org:8445/mcp>

Provides a MCP agent that provides RAG context for a prompt related to PETSc

### Purpose

The purpose of this MCP agent is to enable the implementation of "higher-level" PETSc MCP agents that use high-quality PETSc appropriate
context without needing to implement the RAG infrastructure themselves.

### Usage

The MCP agent may be tested from the command line using

```bash
python petsc_rag_mcp_client.py  [--stdio] [--url PETSC_MCP_SERVERS_URL] --get_documentation_rag_prompt request
```

or

```bash
python petsc_rag_mcp_client.py  [--stdio] [--url PETSC_MCP_SERVERS_URL] --get_developer_rag_prompt request
```

The client Python file also contains a `PetscRAGMCPClient()` class that connects to the server and has a (mirror)
method for each tool provided by the server.
This class can be used when writing new MCP agents that need to utilize RAG.

To start the MCP server locally

```bash
PETSC_DIR=xxx PETSC_ARCH=yyy python NVIDIA_API_KEY=zzz  petsc_rag_mcp_server.py [--stdio]
```

The RAG vector databases are created with

```bash
PETSC_DIR=xxx PETSC_ARCH=yyy python NVIDIA_API_KEY=zzz petsc_rag_mcp_generate_data.py
```


### Limitations

This currently uses the `nvidia/nv-embed-v1` embedding by default because that is the best one I could get working in a portable, open source environment.
It would be better to default to a better embedding (the cost is not the issue). This currently uses `nv-rerank-qa-mistral-4b:1` for reranking, perhaps
better rerankers are available.