finance-docs-mcp

AI-searchable finance & investment knowledge base. MCP server + RAG chatbot for financial literacy education, built on BM25 + metadata filtering. Korean + English bilingual.

What This Does

You write MDX documentation with structured frontmatter. This system indexes those documents and exposes them as MCP tools that Claude Code (or any MCP client) can use to search, retrieve, and reason about your docs.

MDX Documents  -->  Indexer  -->  Search Index  -->  MCP Server  -->  AI Assistant
                                  (BM25)             (tools)

Quick Start

# 1. Clone and install
git clone <repo-url> my-docs-mcp
cd my-docs-mcp
npm install

# 2. Configure
cp .env.example .env
# Edit .env: set OPENAI_API_KEY, DOCS_ROOT

# 3. Build the search index
npm run reindex

# 4. Start the MCP server
npm run start

Architecture

docs-mcp/
|
+-- content/
|   +-- docs/              # Your MDX documentation
|   |   +-- domain-a/      # Organized by domain
|   |   +-- domain-b/
|   +-- conventions/        # How-to-write-docs guides
|
+-- src/
|   +-- indexer/            # MDX parser + BM25 index builder
|   +-- retriever/          # Search engine (BM25 + metadata filters)
|   +-- tools/              # MCP tool definitions
|   +-- chat/               # RAG chatbot (OpenAI)
|
+-- analysis/
|   +-- scripts/            # Analysis pipeline scripts
|   +-- templates/          # MDX templates for generated docs
|
+-- data/
|   +-- search-index.json   # Built index (gitignored)
|
+-- dist/                   # Compiled output (gitignored)

Data Flow

  MDX files (content/docs/)
       |
       v
  [Indexer] -- parses frontmatter + content
       |
       v
  search-index.json (BM25 terms + metadata)
       |
       v
  [Retriever] -- query -> ranked results
       |
       v
  [MCP Tools] -- exposed to AI assistants
       |
       v
  [Chat] -- RAG: retriever results + OpenAI completion

MCP Tools

Tool	Description
search_docs	BM25 keyword search with optional category, domain, platform, and tag filters
get_document	Retrieve a single document by its path
list_categories	List all categories with document counts
find_related	Find documents related to a given document (uses related frontmatter)

Example Queries

search_docs("user registration flow")
search_docs("traps", { category: "traps", domain: "user-management" })
get_document("example/sample-api")
list_categories()
find_related("example/index")

Configuration

Environment Variables

Variable	Required	Default	Description
DOCS_ROOT	No	./content/docs	Path to MDX documents
OPENAI_API_KEY	For chat	--	OpenAI API key for RAG chatbot
PORT	No	3000	Server port
LOG_LEVEL	No	info	Log verbosity: debug, info, warn, error

.mcp.json (Client Configuration)

Add to your project's .mcp.json to connect Claude Code:

{
  "mcpServers": {
    "project-docs": {
      "command": "node",
      "args": ["/path/to/docs-mcp/dist/server.js"],
      "env": {
        "DOCS_ROOT": "/path/to/docs-mcp/content/docs"
      }
    }
  }
}

Writing Documentation

Documents are MDX files with YAML frontmatter. Three fields are required:

---
title: "Document Title"
description: "One-line summary for search results"
category: endpoint
---

See content/conventions/ for full documentation:

• Frontmatter Schema -- All available fields
• Category System -- When to use each category
• FE/BE Separation -- Directory organization

Analysis Pipeline

The analysis/ directory contains scripts and templates for automated documentation generation. The pipeline:

1. Extract -- Parse source code (backend routes, frontend components) to identify endpoints and pages
2. Analyze -- Generate analysis documents (flows, traps, dependencies) using structured templates
3. Generate -- Output MDX files with correct frontmatter into content/docs/
4. Index -- Rebuild the search index with npm run reindex

Templates in analysis/templates/ define the structure for each document category.

Deployment (Fly.io)

# First time
fly launch --name my-docs-mcp

# Deploy
fly deploy

# Set secrets
fly secrets set OPENAI_API_KEY=sk-...

The server runs as a long-lived process. The search index is built at startup from the bundled MDX files.

Development

# Build
npm run build

# Run in development mode (auto-rebuild)
npm run dev

# Rebuild search index
npm run reindex

# Type check
npm run typecheck

Tech Stack

• Runtime: Node.js 20+, TypeScript
• Search: BM25 (no vector DB)
• Protocol: MCP (Model Context Protocol)
• Chat: OpenAI GPT-4 (RAG only, optional)
• Deployment: Fly.io (or any Node.js host)