SF Documentation Knowledge System

Collect, process, and serve Salesforce documentation for LLM agents — using Context Engineering + MCP, not RAG.

Overview

This system programmatically collects all Salesforce documentation from developer.salesforce.com (121 domains, 31,000+ pages), processes it into structured, curated knowledge files, and serves them to LLM agents via:

1. Context Engineering — Pre-compiled Markdown files with _index.md routing tables
2. MCP Server — 5 Model Context Protocol tools for on-demand knowledge access
3. Knowledge Graph — 53,000+ nodes and 450,000+ edges connecting SF concepts, namespaces, services, and cross-references

No embeddings. No vector stores. No blind chunking.

Prerequisites

• Node.js 20 or higher
• Git
• (Optional) Claude Desktop or an MCP-compatible agent

1. Installation

git clone https://github.com/Avinava/sf-documentation-knowledge.git
cd sf-documentation-knowledge
npm install
npm run build

2. Using the Pre-Compiled Knowledge Base

The repository comes pre-loaded with 31,000+ curated markdown files and a Knowledge Graph (53,000+ nodes, 450,000+ edges) covering 121 domains of Salesforce documentation.

Option A: Context Engineering (File-based)

Point your AI agent to the _index.md file in any domain folder. The index acts as a routing table telling the AI which files contain which topics:

knowledge/current/<domain-name>/_index.md

Each domain folder also has a SKILL.md in skills/<domain-name>/SKILL.md that teaches AI agents how to navigate the knowledge.

Option B: Knowledge Graph

The graph at knowledge/current/graph.json connects all documentation with semantic relationships:

Inspect it with:

npm run graph:stats

Option C: MCP Server (Model Context Protocol)

The MCP server exposes 5 tools and 3 prompt templates for AI agents:

Tools:

Prompt Templates:

Quick start (local):

npm run mcp:start

Quick start (npx — no clone needed):

npx @sfdxy/sf-documentation-knowledge

Claude Desktop — add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "sf-docs": {
      "command": "npx",
      "args": ["-y", "@sfdxy/sf-documentation-knowledge"]
    }
  }
}

Cursor / Windsurf / Other MCP clients — point the MCP config to:

npx -y @sfdxy/sf-documentation-knowledge

The server loads the full 53k-node knowledge graph into memory on startup (~3s) and serves all queries instantly.

3. Running the Data Pipeline

To update the knowledge base with the latest Salesforce releases, run the pipeline in order:

Step 1: Discover Available Deliverables

See what documentation deliverables are available from the SF index API:

npm run discover

Step 2: Collect Raw Data

Downloads raw HTML documentation from the Salesforce Atlas metadata APIs.

# Collect a specific domain
npm run collect -- --domain cli-commands

# Collect all configured (P0) domains
npm run collect

# Collect ALL deliverables from the SF index API (121 domains, ~31k pages)
npm run collect -- --discover

Step 3: Process HTML to Markdown

Cleans the raw HTML, strips noise, parses tables, formats code blocks, creates clean Markdown, and automatically redacts any Salesforce tokens or secrets.

# Process a specific domain
npm run process -- --domain cli-commands

# Process ALL collected domains (including discovered ones)
npm run process -- --discover

Step 4: Generate Knowledge Graph & Artifacts

Builds the knowledge graph (with cross-references, namespaces, service categories, and doctype clustering), generates context files for each domain, and updates the README and docs/inventory.md.

# Generate a specific domain
npm run generate -- --domain cli-commands

# Generate ALL collected domains and rebuild the full Knowledge Graph
npm run generate -- --discover

Step 5: Inspect the Graph

npm run graph:stats

Architecture & Contributions

License

Copyright © 2026. All rights reserved.

Documentation

121 domains | 33,188 knowledge files