MCP Local RAG

Local RAG for developers via MCP or CLI. Semantic search with keyword boost for exact technical terms — fully private, zero setup.

Features

• Semantic search with keyword boost Vector search first, then keyword matching boosts exact matches. Terms like useEffect, error codes, and class names rank higher—not just semantically guessed.

• Smart semantic chunking Chunks documents by meaning, not character count. Uses embedding similarity to find natural topic boundaries—keeping related content together and splitting where topics change.

• Quality-first result filtering Groups results by relevance gaps instead of arbitrary top-K cutoffs. Get fewer but more trustworthy chunks.

• Runs entirely locally No API keys, no cloud, no data leaving your machine. Works fully offline after the first model download.

• Zero-friction setup One npx command. No Docker, no Python, no servers to manage. Use via MCP, CLI, or both. Optional Agent Skills help AI assistants form better queries and interpret results.

Quick Start

Set BASE_DIR to the folder you want to search (or BASE_DIRS for multiple roots — see Configuration). Documents must live under one of the configured roots.

Add the MCP server to your AI coding tool:

For Cursor — Add to ~/.cursor/mcp.json:

{
  "mcpServers": {
    "local-rag": {
      "command": "npx",
      "args": ["-y", "mcp-local-rag"],
      "env": {
        "BASE_DIR": "/path/to/your/documents"
      }
    }
  }
}

For Codex — Add to ~/.codex/config.toml:

[mcp_servers.local-rag]
command = "npx"
args = ["-y", "mcp-local-rag"]

[mcp_servers.local-rag.env]
BASE_DIR = "/path/to/your/documents"

For Claude Code — Run this command:

claude mcp add local-rag --scope user --env BASE_DIR=/path/to/your/documents -- npx -y mcp-local-rag

Restart your tool, then start using it:

You: "Ingest api-spec.pdf"
Assistant: Successfully ingested api-spec.pdf (47 chunks created)

You: "What does the API documentation say about authentication?"
Assistant: Based on the documentation, authentication uses OAuth 2.0 with JWT tokens.
          The flow is described in section 3.2...

Or use directly as CLI — no MCP server needed:

npx mcp-local-rag ingest ./docs/
npx mcp-local-rag query "authentication API"

That's it. No Docker, no Python, no server setup.

Why This Exists

You want AI to search your documents—technical specs, research papers, internal docs. But most solutions send your files to external APIs.

Privacy. Your documents might contain sensitive data. This runs entirely locally.

Cost. External embedding APIs charge per use. This is free after the initial model download.

Offline. Works without internet after setup.

Code search. Pure semantic search misses exact terms like useEffect or ERR_CONNECTION_REFUSED. Keyword boost catches both meaning and exact matches.

Agent reality. In practice, many AI environments mainly use tool calling. CLI support and Agent Skills make the same workflows available even without full MCP integration.

Usage

mcp-local-rag provides two interfaces: an MCP server for AI coding tools and a CLI for direct use from the terminal.

Using with MCP

The MCP server provides 7 tools: ingest_file, ingest_data, query_documents, read_chunk_neighbors, list_files, delete_file, status.

Ingesting Documents

"Ingest the document at /Users/me/docs/api-spec.pdf"

Supports PDF, DOCX, TXT, and Markdown. The server extracts text, splits it into chunks, generates embeddings locally, and stores everything in a local vector database.

Re-ingesting the same file replaces the old version automatically.

Ingesting PDFs with figures (visual mode)

PDFs with charts, tables, or diagrams can optionally add local VLM-generated captions to the document index, giving visual content some searchable representation in the same vector + FTS pipeline. Captions are auxiliary text — not image search, not OCR, and not a faithful transcription of the figure.

Via MCP:

"Ingest /Users/me/docs/api-spec.pdf with visual: true"

Via CLI:

npx mcp-local-rag ingest ./docs/spec.pdf --visual

Each caption is emitted as its own chunk with the envelope [Visual content on page N: …], alongside the page-body chunks. It flows through the existing embedder and FTS index — no schema differences, no separate index.

Visual mode is opt-in; normal ingest does not load the VLM. Per-page VLM failures are tolerated — that page proceeds with text only.

Choosing a visual-quality profile

Visual mode offers two profiles, selected per ingest call:

The numbers above are measured on CPU during development on the project's probe PDFs; they may shift with model updates or differ on your hardware.

Via MCP — ingest_file accepts an optional visualQuality parameter (enum: 'fast' | 'quality', default 'fast'; ignored when visual is false):

"Ingest /Users/me/docs/research-paper.pdf with visual: true and visualQuality: 'quality'"

Via CLI — --visual-quality fast|quality (default fast; silently ignored when --visual is absent):

npx mcp-local-rag ingest ./docs/research-paper.pdf --visual --visual-quality quality

Profile model identifiers and quantization variants are fixed per release. Both profiles share the same CACHE_DIR (default: ./models/); the first run on each profile downloads its model.

Behavior change from v0.14.0: Captions are now emitted as dedicated chunks rather than appended to the page text before chunking. As a side effect, metadata.fileSize for visual ingests no longer includes the caption character count — it measures the post-extraction body length only. The underlying PDF is unchanged; only the reported fileSize for visual-ingested PDFs may shrink across the release boundary.

Security note: Visual captions are derived from PDF contents and may inherit attacker-controlled text. Downstream LLM consumers should treat retrieved chunks as untrusted data, not as instructions. The [Visual content on page N: …] envelope helps consumers distinguish caption text from prose.

Ingesting HTML Content

Use ingest_data to ingest HTML content retrieved by your AI assistant (via web fetch, curl, browser tools, etc.):

"Fetch https://example.com/docs and ingest the HTML"

The server extracts main content using Readability (removes navigation, ads, etc.), converts to Markdown, and indexes it. Perfect for:

• Web documentation
• HTML retrieved by the AI assistant
• Clipboard content

HTML is automatically cleaned—you get the article content, not the boilerplate.

Note: The RAG server itself doesn't fetch web content—your AI assistant retrieves it and passes the HTML to ingest_data. This keeps the server fully local while letting you index any content your assistant can access. Please respect website terms of service and copyright when ingesting external content.

Searching Documents

"What does the API documentation say about authentication?"
"Find information about rate limiting"
"Search for error handling best practices"

Search uses semantic similarity with keyword boost. This means useEffect finds documents containing that exact term, not just semantically similar React concepts.

Results include text content, source file, document title, and relevance score. The document title provides context for each chunk, helping identify which document a result belongs to. Adjust result count with limit (1-20, default 10).

Narrow a search to part of your corpus with scope — one path prefix or a list of them. Results are restricted to chunks whose file path equals a prefix or sits under it (exact-or-descendant). For example, "/docs/api" matches /docs/api and /docs/api/auth.md but not /docs/apiv2; a file prefix like "/docs/readme.md" matches just that file. Pass prefixes in the server's OS path style.

Expanding Context Around a Result

When a search result needs more surrounding context, use read_chunk_neighbors to read the chunks before and after it:

"That result about authentication looks relevant — read the surrounding chunks for the full explanation"

Pass the filePath and chunkIndex from the search result. The response includes the target chunk (marked isTarget: true) plus its neighbors, sorted by chunk index. Defaults to 2 chunks before and 2 after (adjustable up to 50 each).

Managing Files

"List all files in configured base directories and their ingested status"   # See what's indexed
"Delete old-spec.pdf from RAG"     # Remove a file
"Show RAG server status"           # Check system health

Using as CLI

All MCP tools are also available as CLI commands — no MCP server needed:

npx mcp-local-rag ingest ./docs/               # Bulk ingest files
npx mcp-local-rag query "authentication API"    # Search documents
npx mcp-local-rag query "auth" --scope /docs/api --scope /docs/guide  # Restrict to path prefixes (repeatable)
npx mcp-local-rag read-neighbors --file-path /abs/path.md --chunk-index 5  # Expand context
npx mcp-local-rag list                          # Show ingestion status
npx mcp-local-rag status                        # Database stats
npx mcp-local-rag delete ./docs/old.pdf         # Remove content
npx mcp-local-rag delete --source "https://..."  # Remove by source URL

query, read-neighbors, list, status, and delete output JSON to stdout for piping (e.g., | jq). ingest outputs progress to stderr. Global options (--db-path, --cache-dir, --model-name) go before the subcommand. Run npx mcp-local-rag --help for details.

⚠️ The CLI does not read your MCP client config (mcp.json, config.toml, etc.). Configure the CLI via flags or environment variables as shown below.

Configuration

CLI flags — global options go before the subcommand, subcommand options go after:

npx mcp-local-rag --db-path ./my-db query "auth" --base-dir ./docs

The --base-dir flag is repeatable on ingest and list; pass it once per root:

npx mcp-local-rag ingest --base-dir ./docs --base-dir ./specs ./docs/readme.md
npx mcp-local-rag list --base-dir ./docs --base-dir ./specs

The positional path to ingest must sit inside one of the configured roots. When at least one --base-dir is supplied, CLI roots replace any env-var roots (no merge).

Environment variables — set in your shell:

export DB_PATH=./my-db
export BASE_DIR=./docs
npx mcp-local-rag query "auth"

For multiple roots, use BASE_DIRS (JSON array of non-empty path strings):

export BASE_DIRS='["/Users/me/Documents/work","/Users/me/Projects/specs"]'
npx mcp-local-rag list

Sharing config between MCP and CLI — if your MCP client inherits shell environment variables, you can set them in your shell profile (e.g., ~/.zshrc) so both use the same values. Otherwise, set them explicitly in your MCP config as well.

export BASE_DIR=/path/to/your/documents
export DB_PATH=/path/to/lancedb

Configuration is resolved in this order:

1. CLI flags (highest priority)
2. Environment variables
3. Defaults

For the full list of CLI flags, environment variables, and defaults, see Configuration.

For CLI-only setups (no MCP server), install Agent Skills so your AI assistant can form better queries and interpret results consistently.

⚠️ CLI --model-name must match the MCP server's MODEL_NAME env var. Using a different embedding model against an existing database produces incompatible vectors, silently degrading search quality.

Search Tuning

Adjust these for your use case:

Code-focused tuning

For codebases and API specs, increase keyword boost so exact identifiers (useEffect, ERR_*, class names) dominate ranking:

"env": {
  "RAG_HYBRID_WEIGHT": "0.7",
  "RAG_GROUPING": "similar"
}

• 0.7 — balanced semantic + keyword
• 1.0 — aggressive; exact matches strongly rerank results

Keyword boost is applied after semantic filtering, so it improves precision without surfacing unrelated matches.

How It Works

TL;DR:

• Documents are chunked by semantic similarity, not fixed character counts
• Each chunk is embedded locally using Transformers.js
• Search uses semantic similarity with keyword boost for exact matches
• Results are filtered based on relevance gaps, not raw scores

Details

When you ingest a document, the parser extracts text based on file type (PDF via mupdf, DOCX via mammoth, text files directly).

The semantic chunker splits text into sentences, then groups them using embedding similarity. It finds natural topic boundaries where the meaning shifts—keeping related content together instead of cutting at arbitrary character limits. This produces chunks that are coherent units of meaning, typically 500-1000 characters. Markdown code blocks are kept intact—never split mid-block—preserving copy-pastable code in search results.

Each chunk goes through a Transformers.js embedding model (default: all-MiniLM-L6-v2, configurable via MODEL_NAME), converting text into vectors. Vectors are stored in LanceDB, a file-based vector database requiring no server process.

When you search:

1. Your query becomes a vector using the same model
2. Semantic (vector) search finds the most relevant chunks
3. Quality filters apply (distance threshold, grouping)
4. Keyword matches boost rankings for exact term matching

The keyword boost ensures exact terms like useEffect or error codes rank higher when they match.

Agent Skills

Agent Skills provide optimized prompts that help AI assistants use RAG tools more effectively. Install skills for better query formulation, result interpretation, and ingestion workflows:

# Claude Code (project-level)
npx mcp-local-rag skills install --claude-code

# Claude Code (user-level)
npx mcp-local-rag skills install --claude-code --global

# Codex
npx mcp-local-rag skills install --codex

Skills include:

• Query optimization: Better search query formulation
• Result interpretation: Score thresholds and filtering guidelines
• HTML ingestion: Format selection and source naming

Ensuring Skill Activation

Skills are loaded automatically in most cases—AI assistants scan skill metadata and load relevant instructions when needed. For consistent behavior:

Option 1: Explicit request (natural language) Before RAG operations, request in natural language:

• "Use the mcp-local-rag skill for this search"
• "Apply RAG best practices from skills"

Option 2: Add to agent instruction file Add to your AGENTS.md, CLAUDE.md, or other agent instruction file:

When using query_documents, ingest_file, or ingest_data tools,
apply the mcp-local-rag skill for better query formulation and result interpretation.

Configuration

Environment Variables and CLI Flags

The MCP server is configured by environment variables only — pass them through your MCP client's env block. The CLI accepts the same env vars plus equivalent flags (priority: CLI flag > env > default). CLI flags are not accepted on the bare mcp-local-rag (MCP server) launch.

Model choice tips:

• Multilingual docs → e.g., onnx-community/embeddinggemma-300m-ONNX (100+ languages)
• Scientific papers → e.g., sentence-transformers/allenai-specter (citation-aware)
• Code repositories → default often suffices; keyword boost matters more (or jinaai/jina-embeddings-v2-base-code)

⚠️ Changing MODEL_NAME changes embedding dimensions. Delete DB_PATH and re-ingest after switching models.

Document Roots (BASE_DIR and BASE_DIRS)

mcp-local-rag enforces a security boundary: only files under a configured root are accessible to ingest, list, delete, or read-neighbor operations.

Single root — use BASE_DIR:

export BASE_DIR=/Users/me/Documents/work

Multiple roots — use BASE_DIRS with a JSON array:

export BASE_DIRS='["/Users/me/Documents/work","/Users/me/Projects/specs"]'

Only JSON-array syntax is supported. Delimiter syntax such as BASE_DIRS=/a:/b is intentionally not supported (avoids ambiguity with spaces, colons, commas, and Windows paths).

Resolution order (highest precedence first):

1. CLI --base-dir <path> flags (repeatable on ingest and list)
2. BASE_DIRS environment variable
3. BASE_DIR environment variable
4. process.cwd() (current working directory)

CLI roots replace env roots — they are never merged. BASE_DIRS and BASE_DIR are never merged either: BASE_DIRS wins when both are set.

Precedence warning — when BASE_DIRS and BASE_DIR are both set (and no CLI --base-dir is supplied), BASE_DIR is ignored and a warning is surfaced. The warning is visible:

• In MCP tool responses (as an additional content block, on every tool — including status, query_documents, ingest_file, ingest_data, list_files, delete_file, read_chunk_neighbors).
• On CLI stderr.

Unset BASE_DIR (or remove BASE_DIRS) to silence the warning.

Nested-root pruning — if one configured root sits inside another after realpath resolution, the nested child is dropped to avoid duplicate scan results. A pruning warning is surfaced the same way as the precedence warning. The surviving parent root still defines the security boundary.

Invalid BASE_DIRS — when BASE_DIRS is not a valid JSON array of non-empty strings (malformed JSON, empty array, non-string elements, ...), root-dependent MCP tools return a structured error and CLI subcommands exit non-zero. There is no silent fallback to BASE_DIR or cwd. The MCP status tool remains callable so you can diagnose the config error through your MCP client.

MCP client examples — multi-root setup:

Cursor (~/.cursor/mcp.json):

{
  "mcpServers": {
    "local-rag": {
      "command": "npx",
      "args": ["-y", "mcp-local-rag"],
      "env": {
        "BASE_DIRS": "[\"/Users/me/Documents/work\",\"/Users/me/Projects/specs\"]"
      }
    }
  }
}

Codex (~/.codex/config.toml):

[mcp_servers.local-rag]
command = "npx"
args = ["-y", "mcp-local-rag"]

[mcp_servers.local-rag.env]
BASE_DIRS = "[\"/Users/me/Documents/work\",\"/Users/me/Projects/specs\"]"

Claude Code:

claude mcp add local-rag --scope user \
  --env BASE_DIRS='["/Users/me/Documents/work","/Users/me/Projects/specs"]' \
  -- npx -y mcp-local-rag

CLI examples — multi-root invocations:

# Repeatable --base-dir
npx mcp-local-rag ingest --base-dir /Users/me/work --base-dir /Users/me/specs /Users/me/work/readme.md
npx mcp-local-rag list --base-dir /Users/me/work --base-dir /Users/me/specs

# Or via BASE_DIRS env
BASE_DIRS='["/Users/me/work","/Users/me/specs"]' npx mcp-local-rag list

Client-Specific Setup

Cursor — Global: ~/.cursor/mcp.json, Project: .cursor/mcp.json

{
  "mcpServers": {
    "local-rag": {
      "command": "npx",
      "args": ["-y", "mcp-local-rag"],
      "env": {
        "BASE_DIR": "/path/to/your/documents"
      }
    }
  }
}

Codex — ~/.codex/config.toml (note: must use mcp_servers with underscore)

[mcp_servers.local-rag]
command = "npx"
args = ["-y", "mcp-local-rag"]

[mcp_servers.local-rag.env]
BASE_DIR = "/path/to/your/documents"

Claude Code:

claude mcp add local-rag --scope user \
  --env BASE_DIR=/path/to/your/documents \
  -- npx -y mcp-local-rag

First Run

The embedding model (~90MB) downloads on first use. Takes 1-2 minutes, then works offline.

Security

• Path restriction: Only files within a configured root (BASE_DIR or any BASE_DIRS / --base-dir entry) are accessible. Symlinks resolving outside all configured roots, and sibling-prefix paths (e.g. /foo/barista for root /foo/bar), are rejected.
• Local only: No network requests after model download
• Model sources (all official HuggingFace repositories):
  • Embedder: Xenova/all-MiniLM-L6-v2
  • Visual fast profile: HuggingFaceTB/SmolVLM-256M-Instruct
  • Visual quality profile: onnx-community/Qwen2.5-VL-3B-Instruct-ONNX
• Visual caption fidelity: The quality profile reproduces in-image text more faithfully than fast. Both profiles output captions wrapped as [Visual content on page N: …], but a faithful reproduction means attacker-controlled in-image text — including characters like ] that visually close the envelope — can appear verbatim in retrieved chunks. Downstream LLM consumers should treat retrieved chunks as untrusted data, not as instructions, regardless of envelope shape.

Tested on MacBook Pro M1 (16GB RAM), Node.js 22:

Query Speed: ~1.2 seconds for 10,000 chunks (p90 < 3s)

Ingestion (10MB PDF):

• PDF parsing: ~8s
• Chunking: ~2s
• Embedding: ~30s
• DB insertion: ~5s

Memory: ~200MB idle, ~800MB peak (50MB file ingestion)

Concurrency: Handles 5 parallel queries without degradation.

"No results found"

Documents must be ingested first. Run "List all ingested files" to verify.

Model download failed

Check internet connection. If behind a proxy, configure network settings. The model can also be downloaded manually.

"File too large"

Default limit is 100MB. Split large files or increase MAX_FILE_SIZE.

Slow queries

Check chunk count with status. Large documents with many chunks may slow queries. Consider splitting very large files.

"Path outside BASE_DIR"

Ensure file paths are within one of the configured roots (BASE_DIR, any BASE_DIRS entry, or any CLI --base-dir). Use absolute paths.

"BASE_DIRS must be a JSON array..."

BASE_DIRS accepts only a JSON array of one or more non-empty path strings. Examples:

• Valid: BASE_DIRS='["/Users/me/work","/Users/me/specs"]'
• Invalid: BASE_DIRS=/a:/b (delimiter syntax not supported)
• Invalid: BASE_DIRS='[]' (empty array)
• Invalid: BASE_DIRS='["",""]' (empty string element)

When invalid, root-dependent operations fail with a clear error rather than silently falling back. The MCP status tool remains callable so you can inspect the diagnostic.

MCP client doesn't see tools

1. Verify config file syntax
2. Restart client completely (Cmd+Q on Mac for Cursor)
3. Test directly: npx mcp-local-rag should run without errors

Is this really private? Yes. After model download, nothing leaves your machine. Verify with network monitoring.

Can I use this offline? Yes, after the required models are cached locally. Text ingest/search needs the embedding model. PDF visual mode is opt-in and also needs the VLM model on first use; the download is ~250 MB for the default fast profile (SmolVLM-256M) or ~2.9 GB for the quality profile (Qwen2.5-VL-3B), cached under CACHE_DIR (default: ./models/).

How does this compare to cloud RAG? Cloud services offer better accuracy at scale but require sending data externally. This trades some accuracy for complete privacy and zero runtime cost.

What file formats are supported? PDF, DOCX, TXT, Markdown, and HTML (via ingest_data). Not yet: Excel, PowerPoint, images.

Can I change the embedding model? Yes, but you must delete your database and re-ingest all documents. Different models produce incompatible vector dimensions.

GPU acceleration? Opt-in via RAG_DEVICE. Devices are passed straight to ONNX Runtime. GPU support is highly dependent on your system, Node.js version, and the underlying ONNX backend. See the Transformers.js device source code for the live list of supported backend names. If the requested device fails to initialize, the server throws an error — set RAG_DEVICE=cpu to revert.

Can I change the embedding precision (dtype)? Opt-in via RAG_DTYPE (default fp32); accepted values are in the env-var table above. A recognized dtype the model lacks errors and lists the available ones; an unrecognized value (a typo) silently falls back to fp32. Changing RAG_DEVICE/RAG_DTYPE changes the embedding space — delete DB_PATH and re-ingest.

Multi-user support? No. Designed for single-user, local access. Multi-user would require authentication/access control.

How to backup? Copy DB_PATH directory (default: ./lancedb/).

Building from Source

git clone https://github.com/shinpr/mcp-local-rag.git
cd mcp-local-rag
pnpm install

Testing

pnpm test              # Run all tests
pnpm run test:watch    # Watch mode

Code Quality

pnpm run type-check    # TypeScript check
pnpm run check:fix     # Lint and format
pnpm run check:deps    # Circular dependency check
pnpm run check:all     # Full quality check

Project Structure

src/
  index.ts      # Entry point
  server/       # MCP tool handlers
  cli/          # CLI subcommands (ingest, query, list, delete, read-neighbors, etc.)
  parser/       # PDF, DOCX, TXT, MD parsing
  chunker/      # Text splitting
  embedder/     # Transformers.js embeddings
  vectordb/     # LanceDB operations
  __tests__/    # Test suites

Contributing

Contributions welcome! See CONTRIBUTING.md for setup and guidelines.

License

MIT License. Free for personal and commercial use.

Blog Posts

• Building a Local RAG for Agentic Coding — Technical deep-dive into the semantic chunking and hybrid search design.

Acknowledgments

Built with Model Context Protocol by Anthropic, LanceDB, and Transformers.js.