ResDAC Knowledge Base (resdac-knowledge-base aka rkb)

A Python library and command-line toolkit for crawling, archiving, parsing, and searching public CMS (Centers for Medicare & Medicaid Services) and ResDAC (Research Data Assistance Center) documentation.

It enables researchers, data engineers, and AI agents to query dataset metadata, variable definitions, and document structures offline with strict citation-backed provenance.

Key Features

• Offline-First Preservation: Rebuild the entire knowledge base from pre-packaged raw snapshots containing HTML, PDF, and spreadsheet documentation.
• Pipeline Processing: Pipeline commands to crawl/inventory, archive, parse, extract variables, and perform quality assurance audits.
• Lexical Search: Query the compiled SQLite FTS5 index for variables, datasets, and document chunks.
• Model Context Protocol (MCP) Support: Run an MCP server to connect the knowledge base directly to AI editors and tools (e.g., Claude Desktop, Cursor).
• Provenance-First Design: Every retrieved definition or document chunk is mapped back to its source URL and local archived path.

Installation

Install using pip or uv:

pip install resdac-doc-archive

For optional semantic search support:

pip install "resdac-doc-archive[semantic]"

Quick Start

1. Build the Search Index

Before querying, compile the SQLite FTS5 search index:

cms-kb-index

2. Search variables & definitions

Query the search index using the command-line interface:

cms-kb-search --query "BENE_ID" --limit 5

3. Expose as MCP Server

Start the Model Context Protocol server to connect your AI assistant:

cms-kb-mcp

See more options and entry points/paths in the user manual.

Programmatic API Example

You can query the knowledge base and retrieve citation-packed context directly in Python:

from pathlib import Path
from cms_kb import AgentContextConfig, build_agent_context

# Configure paths to manifests and search indices
config = AgentContextConfig(
  archive_manifest_path=Path("manifests/archive_manifest.csv")
)

# Search for a variable and print definition and source citation
response = build_agent_context(config, query="DUAL_ELG", limit=1)

for hit in response.results:
  print(f"Record: {hit.title} ({hit.record_type})")
  print(f"Snippet: {hit.snippet}")
  print(f"Source URL: {hit.citation.source_url}")
  if hit.citation.variable_document:
    print(f"Local Archive Path: {hit.citation.variable_document}")

Documentation

For full details on the data model, pipeline steps, and development guidelines, refer to the repository documentation:

• User Manual
• Developer Guide
• Data Model Reference