mcp-markdown-ragdocs
A Model Context Protocol server that provides semantic search over local Markdown documentation using hybrid retrieval.
What it is
This is an MCP server that indexes local Markdown files and exposes a query_documents tool for hybrid semantic search. The server identifies relevant document sections using semantic search, keyword matching, and graph traversal, enabling efficient discovery without loading entire documentation collections into LLM context.
Why it exists
Technical documentation, personal notes, and project wikis are typically stored as Markdown files. Searching these collections manually or with grep is inefficient. This server provides a conversational interface to query documentation using natural language while automatically keeping the index synchronized with file changes.
Existing RAG solutions require manual database setup, explicit indexing steps, and ongoing maintenance. This server eliminates that friction with automatic file watching, zero-configuration defaults, and built-in index versioning.
Features
- Hybrid search combining semantic embeddings (FAISS), keyword search (Whoosh), and graph traversal (NetworkX)
- Community-based boosting: Louvain clustering detects document communities; co-community results receive score boost
- Score-aware dynamic fusion: Adjusts vector/keyword weights based on score variance per query
- HyDE (Hypothetical Document Embeddings):
search_with_hypothesistool for vague queries - Cross-encoder re-ranking for improved precision (optional, ~50ms latency)
- Query expansion via concept vocabulary for better recall
- Git history search: Semantic search over commit history with metadata and delta context (parallel indexing for 2-4x speedup)
- Multi-project support: Manage isolated indices for multiple projects on one machine with automatic project detection
- Server-Sent Events (SSE) streaming for real-time response delivery
- CLI query command with rich formatted output
- Automatic file watching with debounced incremental indexing
- Zero-configuration operation with sensible defaults
- Index versioning with automatic rebuild on configuration changes
- Pluggable parser architecture: Markdown and plain text (.txt) support out-of-the-box
- Rich Markdown parsing: frontmatter, wikilinks, tags, transclusions
- Reciprocal Rank Fusion for multi-strategy result merging
- Recency bias for recently modified documents
- Memory Management System: Persistent AI memory bank with cross-corpus linking, calibrated scoring (sigmoid expansion for interpretable 0-1 range), exponential decay, and ghost node graph traversal
- Local-first architecture with no external dependencies
Installation
Requires Python 3.13+.
git clone https://github.com/yourusername/mcp-markdown-ragdocs.git
cd mcp-markdown-ragdocs
uv sync
Quick Start
For VS Code / MCP Clients (Recommended)
Start the stdio-based MCP server for use with VS Code or other MCP clients:
uv run mcp-markdown-ragdocs mcp
The server will:
- Scan for
*.mdand*.txtfiles in the current directory - Build vector, keyword, and graph indices
- Start file watching for automatic updates
- Expose query_documents tool via stdio transport
See MCP Integration below for VS Code configuration.
For HTTP API / Development
Start the HTTP server on default port 8000:
uv run mcp-markdown-ragdocs run
The server will:
- Index documents (same as mcp command)
- Expose HTTP API at
http://127.0.0.1:8000 - Provide REST endpoints for queries
See API Endpoints below for HTTP usage.
Basic Usage
Configuration
Create .mcp-markdown-ragdocs/config.toml in your project directory or at ~/.config/mcp-markdown-ragdocs/config.toml:
[server]
host = "127.0.0.1"
port = 8000
[indexing]
documents_path = "~/Documents/Notes" # Path to your Markdown files
index_path = ".index_data/" # Where to store indices
[parsers]
"**/*.md" = "MarkdownParser" # Markdown files
"**/*.txt" = "PlainTextParser" # Plain text files
[search]
semantic_weight = 1.0 # Weight for semantic search results
keyword_weight = 1.0 # Weight for keyword search results
recency_bias = 0.5 # Boost for recently modified documents
rrf_k_constant = 60 # Reciprocal Rank Fusion constant
min_confidence = 0.3 # Score threshold (default: 0.3)
max_chunks_per_doc = 2 # Per-document limit (default: 2)
dedup_enabled = true # Semantic deduplication (default: true)
The server searches for configuration files in this order:
.mcp-markdown-ragdocs/config.tomlin current directory.mcp-markdown-ragdocs/config.tomlin parent directories (walks up to root)~/.config/mcp-markdown-ragdocs/config.toml(global fallback)
This supports monorepo workflows where you can place a shared configuration in the repository root.
If no configuration file exists, the server uses these defaults:
- Documents path:
.(current directory) - Server:
127.0.0.1:8000 - Index storage:
.index_data/
CLI Commands
Start MCP Server (stdio)
uv run mcp-markdown-ragdocs mcp
Starts stdio-based MCP server for VS Code and compatible MCP clients. Runs persistently until stopped.
Start HTTP Server
uv run mcp-markdown-ragdocs run
Starts HTTP API server on port 8000 (default). Override with:
uv run mcp-markdown-ragdocs run --host 0.0.0.0 --port 8080
Query Documents (CLI)
Query documents directly from command line:
uv run mcp-markdown-ragdocs query "How do I configure authentication?"
With options:
# JSON output for scripting
uv run mcp-markdown-ragdocs query "authentication" --json
# Limit number of results
uv run mcp-markdown-ragdocs query "authentication" --top-n 3
# Specify project context
uv run mcp-markdown-ragdocs query "authentication" --project my-project
# Enable debug mode to see search internals
uv run mcp-markdown-ragdocs query "authentication" --debug
Debug Mode (--debug):
Displays two formatted tables showing search internals:
-
Search Strategy Results: Result counts from each search strategy
- Vector (Semantic): FAISS embedding search results
- Keyword (BM25): Whoosh keyword search results
- Graph (PageRank): NetworkX graph traversal results
- Code: Code-specific search results (if enabled)
- Tag Expansion: Query expansion results (if enabled)
-
Compression Pipeline: Filtering stages with counts and items removed
- Original (RRF Fusion): Combined results before filtering
- Confidence Filter: Low-score results removed (threshold:
min_confidence) - Content Dedup: Exact duplicate content removed
- N-gram Dedup: Near-duplicate content removed (character n-grams)
- Semantic Dedup: Semantically similar results removed (cosine similarity)
- Doc Limit: Per-document chunk limit applied (
max_chunks_per_doc)
Example output:
┌─ Search Strategy Results ─────────────┐
│ Strategy │ Count │
├───────────────────┼──────────────────┤
│ Vector (Semantic) │ 15 │
│ Keyword (BM25) │ 8 │
│ Graph (PageRank) │ 3 │
└───────────────────┴──────────────────┘
┌─ Compression Pipeline ─────────────────────┐
│ Stage │ Count │ Removed│
├────────────────────────────┼───────┼────────┤
│ Original (RRF Fusion) │ 26 │ - │
│ After Confidence Filter │ 20 │ 6 │
│ After Content Dedup │ 18 │ 2 │
│ After N-gram Dedup │ 16 │ 2 │
│ After Semantic Dedup │ 12 │ 4 │
│ After Doc Limit │ 5 │ 7 │
└────────────────────────────┴───────┴────────┘
Use debug mode to:
- Understand why certain results appear or are filtered
- Tune search configuration (weights, thresholds, dedup settings)
- Diagnose low-quality results (check if semantic or keyword search dominates)
- Identify over-aggressive deduplication (high removal in dedup stages)
- Optimize performance (balance precision vs. recall via thresholds)
Configuration Management
Check your configuration:
uv run mcp-markdown-ragdocs check-config
Force a full index rebuild:
uv run mcp-markdown-ragdocs rebuild-index
| Command | Purpose | Use When |
|---|---|---|
mcp | Stdio MCP server | Integrating with VS Code or MCP clients |
run | HTTP API server | Development, testing, or HTTP-based integrations |
query | CLI query with optional --debug flag | Scripting, quick searches, or debugging search behavior |
check-config | Validate config | Debugging configuration issues |
rebuild-index | Force full reindex (documents, git commits, vocabulary) | Config changes, corrupted indices, or force rebuild |
MCP Integration
VS Code Configuration
Configure the MCP server in VS Code user settings or workspace settings.
File: .vscode/settings.json or ~/.config/Code/User/mcp.json
{
"mcpServers": {
"markdown-docs": {
"command": "uv",
"args": [
"--directory",
"/absolute/path/to/mcp-markdown-ragdocs",
"run",
"mcp-markdown-ragdocs",
"mcp"
],
"type": "stdio"
}
}
}
With project override:
{
"mcpServers": {
"markdown-docs": {
"command": "uv",
"args": [
"--directory",
"/absolute/path/to/mcp-markdown-ragdocs",
"run",
"mcp-markdown-ragdocs",
"mcp",
"--project",
"my-project"
],
"type": "stdio"
}
}
}
Claude Desktop Configuration
File: ~/Library/Application Support/Claude/claude_desktop_config.json (macOS)
{
"mcpServers": {
"markdown-docs": {
"command": "uv",
"args": [
"--directory",
"/absolute/path/to/mcp-markdown-ragdocs",
"run",
"mcp-markdown-ragdocs",
"mcp"
]
}
}
}
Available Tools
The server exposes two MCP tools:
query_documents: Search indexed documents using hybrid search and return ranked document chunks.
search_git_history: Search git commit history using natural language queries. Returns relevant commits with metadata, message, and diff context.
search_with_hypothesis: Search using a hypothesis about expected documentation content. Embeds the hypothesis directly for semantic search (HyDE technique). Useful for vague queries where describing expected content yields better results than the query itself.
Parameters:
query(required): Natural language query or questiontop_n(optional): Maximum results to return (1-100, default: 5)uniqueness_mode(optional): Result uniqueness strategy -"off"(allow duplicates),"document"(one chunk per document), or"content"(semantic deduplication, default)min_score(optional): Minimum confidence threshold (0.0-1.0, default: 0.3)similarity_threshold(optional): Semantic deduplication threshold (0.5-1.0, default: 0.85)show_stats(optional): Show compression statistics (default: false)
Note: Compression is enabled by default (min_score=0.3, max_chunks_per_doc=2, dedup_enabled=true) to reduce token overhead by 40-60%. Results use compact format: [N] file § section (score)\ncontent
Usage Pattern:
- Call
query_documentsto identify relevant sections - Review returned chunks to locate specific files and sections
- Use file reading tools to access full document context
Example query from MCP client:
{
"query": "How do I configure authentication in the API?",
"top_n": 5,
"min_score": 0.3
}
The server returns ranked document chunks with file paths, header hierarchies, and relevance scores.
search_git_history: Search git commit history using natural language queries.
Parameters:
query(required): Natural language query describing commits to findtop_n(optional): Maximum commits to return (1-100, default: 5)min_score(optional): Minimum relevance threshold (0.0-1.0, default: 0.0)file_pattern(optional): Glob pattern to filter by changed files (e.g.,src/**/*.py)author(optional): Filter commits by author name or emailafter(optional): Unix timestamp to filter commits after this datebefore(optional): Unix timestamp to filter commits before this date
Note: Git history search indexes up to 200 lines of diff per commit. Indexing processes 60 commits/sec on average. Search latency averages 5ms for 10k commits.
Example query:
{
"query": "fix authentication bug",
"top_n": 5,
"file_pattern": "src/auth/**",
"after": 1704067200
}
The server returns ranked commits with hash, title, author, timestamp, message, files changed, and truncated diff.
Memory Management
The server supports an AI memory bank for persistent cross-session knowledge storage.
Enable in configuration:
[memory]
enabled = true
storage_strategy = "project" # "project" or "user"
score_threshold = 0.2 # Minimum score after calibration (0.0-1.0)
# Per-type exponential decay (defaults shown)
[memory.decay_journal]
decay_rate = 0.90 # 7-day half-life
floor_multiplier = 0.1
Available tools:
create_memory,read_memory,update_memory,append_memory,delete_memory: CRUD operations (system auto-generates frontmatter forcreate_memory)search_memories: Hybrid search with recency boost, tag/type filtering, and time range filtering (absolute timestamps or relative days)search_linked_memories: Find memories linking to a specific document via ghost nodesget_memory_stats: Memory bank statisticsmerge_memories: Consolidate multiple memories into one
Time range filtering examples:
// Last 7 days
{"query": "bug fixes", "relative_days": 7}
// Absolute range (Jan 1-31, 2024)
{"query": "features", "after_timestamp": 1704067200, "before_timestamp": 1706745600}
// Combined with tag filtering
{"query": "auth improvements", "relative_days": 30, "filter_tags": ["security"]}
See Memory Management for complete documentation.
API Endpoints
Health check:
curl http://127.0.0.1:8000/health
Server status (document count, queue size, failed files):
curl http://127.0.0.1:8000/status
Query endpoint (standard):
curl -X POST http://127.0.0.1:8000/query_documents \
-H "Content-Type: application/json" \
-d '{"query": "authentication configuration"}'
Query endpoint (streaming SSE):
curl -X POST http://127.0.0.1:8000/query_documents_stream \
-H "Content-Type: application/json" \
-d '{"query": "authentication configuration", "top_n": 3}' \
-N
The streaming endpoint returns Server-Sent Events:
event: search_complete
data: {"count": 3}
event: token
data: {"token": "Authentication"}
event: token
data: {"token": " is"}
event: done
data: {"results": [{"content": "...", "file_path": "auth.md", "header_path": ["Configuration"], "score": 1.0}]}
Example response (standard endpoint):
{
"results": [
{
"chunk_id": "authentication_0",
"content": "Authentication is configured in the auth section...",
"file_path": "docs/authentication.md",
"header_path": ["Configuration", "Authentication"],
"score": 0.92
},
{
"chunk_id": "security_2",
"content": "Security settings include authentication tokens...",
"file_path": "docs/security.md",
"header_path": ["Security", "API Keys"],
"score": 0.78
}
]
}
MCP Stdio Format (Compact):
For MCP clients (VS Code, Claude Desktop), results use compact format:
[1] docs/authentication.md § Configuration > Authentication (0.92)
Authentication is configured in the auth section...
[2] docs/security.md § Security > API Keys (0.78)
Security settings include authentication tokens...
Factual queries (e.g., "getUserById function", "configure auth") truncate content to 200 characters:
[1] docs/api.md § Functions > getUserById (0.92)
getUserById(id: string): User | null
Retrieves user by ID. Returns null if not found. Example:
const user = getUserById("123");
if (user) ...
Each result contains:
- `chunk_id`: Unique identifier for the document chunk
- `content`: The text from the matching document chunk
- `file_path`: Source file path relative to documents directory
- `header_path`: Document structure showing nested headers (semantic "breadcrumbs")
- `score`: Calibrated confidence score [0, 1] representing absolute match quality (>0.9 = excellent, 0.7-0.9 = good, 0.5-0.7 = moderate, <0.3 = noise)
## Configuration Details
See [docs/configuration.md](docs/configuration.md) for exhaustive configuration reference including all TOML options, defaults, and environment variable support.
## Documentation
- [Architecture](docs/architecture.md) - System design, component overview, data flow
- [Configuration](docs/configuration.md) - Complete configuration reference
- [Hybrid Search](docs/hybrid-search.md) - Search strategies and RRF fusion algorithm
- [Integration](docs/integration.md) - VS Code MCP setup and client integration
- [Memory Management](docs/memory.md) - AI memory bank, CRUD tools, ghost nodes
- [Troubleshooting](docs/troubleshooting.md) - Debug mode usage, tuning guide, common issues
- [Development](docs/development.md) - Development setup, testing, contributing
## License
MIT