mcpField
N
nemo-retriever-mcp

nemo-retriever-mcp

github.com/nvidiaDocuments3.8k67 installsUpdated today
Repository

About nemo-retriever-mcp

nemo-retriever-mcp is a skill for using NeMo Retriever MCP tools to search documents and, when needed, ingest documents available to the server. It guides evidence-based retrieval and citation handling.

Key features

  1. 01Search indexed documents
  2. 02Ingest server-readable documents when needed
  3. 03Retrieve evidence from new documents
  4. 04Preserve document identifiers and page numbers

Use cases

  1. 01Find evidence in documents already searchable
  2. 02Add documents available on the Retriever server
  3. 03Prepare cited answers with document and page references

MCP support

It uses the NeMo Retriever MCP server's tools under `retriever`, including `query` and `ingest_documents`. The server exposes tools rather than MCP resources.

Install

Copy the command and run it in a terminal.

$npx skills add https://github.com/nvidia/skills --skill nemo-retriever-mcp

After installing, type/nemo-retriever-mcpin Claude Code to confirm it loaded.

Bundled files

References and scripts installed along with SKILL.md.

BENCHMARK.md
skill-card.md
skill.oms.sig
evals/
evals.json

Details

From the project README

NeMo Retriever MCP

Use the NeMo Retriever MCP tools exposed under retriever instead of the retriever CLI.

The server exposes MCP tools, not MCP resources. Do not call list_mcp_resources to discover documents or decide whether Retriever is available; an empty MCP resource list is expected.

Workflow

  • Use query to search documents already available to Retriever.
  • Never ingest documents that are already searchable, including documents in a prebuilt index.
  • Use ingest_documents only when required documents are not yet searchable and server-readable paths are explicitly available. Do not assume that paths visible to the agent are also readable by the Retriever server.
  • After ingestion completes, use query to retrieve evidence from the added documents.
  • Start with top_k=5 and format="hits". Request more only when the first successful response does not contain enough evidence to answer the task.
  • Read evidence hits from the tool response and ground your answer in them.
  • Preserve available document identifiers such as document_id, source_id, or pdf_basename, along with page numbers, when the task requests citations or structured output.
  • If query fails, inspect the error and retry with corrected parameters or a shorter equivalent query. If it still fails, report the error before using another retrieval method.

Comments

Markdown supported