<p align="center">
  <img src="assets/banner.jpg" alt="MCP Local RAG: Search below the surface." width="600" />
</p>

# MCP Local RAG

[![GitHub
stars](https://img.shields.io/github/stars/shinpr/mcp-local-rag?style=social)](https://github.com/shinpr/mcp-local-rag)
[![npm
version](https://img.shields.io/npm/v/mcp-local-rag.svg)](https://www.npmjs.com/package/mcp-local-rag)
[![License:
MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)
[![MCP
Registry](https://img.shields.io/badge/MCP-Registry-green.svg)](https://registry.modelcontextprotocol.io/)

<p align="center">
  <strong>English</strong> |
  <a href="README.zh-CN.md">简体中文</a> |
  <a href="README.de.md">Deutsch</a> |
  <a href="README.es.md">Español</a> |
  <a href="README.pt-BR.md">Português (Brasil)</a> |
  <a href="README.fr.md">Français</a>
</p>

Search private documents from an MCP client or the terminal without sending them to an
embedding API.

mcp-local-rag indexes PDF, DOCX, Markdown, and text files on your machine. Search combines
semantic similarity with keyword matching, so queries can match both intent and exact technical
terms such as API names, class names, and error codes. Results include source passages and,
where available, headings, line numbers, or page numbers so you can check and cite the original
document.

No API key, Docker, Python, or external database is required. After the initial model download,
text ingestion and search work offline.

## Quick Start

### Requirements

- Node.js 22 or later
- Internet access on first use to download the npm package and embedding model
- A directory containing the documents you want to search

Set `BASE_DIR` to that directory. It is also the security boundary for file operations. Replace
`/absolute/path/to/your/documents` below with the directory's absolute path.

Use one of the examples below, or register `npx -y mcp-local-rag` and set `BASE_DIR` using your
client's MCP configuration format.

Set `DB_PATH` and `CACHE_DIR` to absolute paths as well. Relative paths resolve from the
server's working directory, so starting the server from different projects creates a separate
index and model cache in each.

<details>
<summary>Claude Code</summary>

Run this command:

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

</details>

<details>
<summary>Codex</summary>

Add to `~/.codex/config.toml`:

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

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

</details>

<details>
<summary>OpenCode</summary>

Add to `~/.config/opencode/opencode.json` (or `opencode.jsonc`):

```json
{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "local-rag": {
      "type": "local",
      "command": ["npx", "-y", "mcp-local-rag"],
      "environment": {
        "BASE_DIR": "/absolute/path/to/your/documents"
      }
    }
  }
}
```

</details>

<details>
<summary>Cursor</summary>

Add to `~/.cursor/mcp.json`:

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

</details>

Restart the client, then ask it to build the index:

```text
Sync all documents in the configured root and wait until it finishes.
```

The first sync downloads the default embedding model (about 90 MB) and may take 1–2 minutes
before ingestion starts. Later runs use the local cache.

Once the sync completes:

```text
What does the API documentation say about authentication?
```

### CLI Quick Start

To use the CLI without an MCP client:

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

The CLI uses the current directory as its document root by default. Run both commands from the
same directory so they use the same default index, or set `BASE_DIR` and `DB_PATH` explicitly.

## Supported Content

| Input | How to ingest |
|---|---|
| PDF, DOCX, TXT, Markdown | File ingestion or directory sync |
| HTML already fetched by the client | `ingest_data` |
| Plain text or Markdown held in memory | `ingest_data` with a stable source identifier |

HTML fetching is not built into the server. An MCP client can fetch a page and pass its HTML to
`ingest_data`.

Excel, PowerPoint, standalone images, and source-code file extensions are not supported by file
ingestion. PDFs can optionally use a local vision model to describe figures, but this is not
OCR or image search.

## Using the Index

Sync after adding, editing, or removing documents. For searches and follow-up reading, ask your
MCP client:

```text
Find the documented behavior of ERR_CONNECTION_REFUSED.
Read the surrounding chunks for that result.
```

You can also ingest a single file or HTML already fetched by the client. To refresh an existing
entry, ingest the document again using the same path or source identifier. `sync` skips unchanged
files. MCP file paths must be absolute and inside a configured document root.

PDF section headings may be inaccurate. Check the original page when you need the exact heading.

<details>
<summary>MCP Tools</summary>

| Tool | Purpose |
|---|---|
| `sync_start` | Reconcile the index with all configured roots or one path |
| `sync_status` | Poll a running sync job |
| `ingest_file` | Ingest or replace one file |
| `ingest_data` | Ingest text, Markdown, or HTML already held by the client |
| `query_documents` | Search with semantic matching and keyword boost |
| `read_chunk_neighbors` | Read surrounding chunks from a search result |
| `list_files` | Show supported files and their ingestion state |
| `delete_file` | Delete an indexed file or an `ingest_data` item |
| `status` | Show index and search status |

</details>

## CLI

Use the CLI to update the index, narrow searches, or remove indexed content:

```bash
npx mcp-local-rag sync ./docs/
npx mcp-local-rag query "auth" --scope /docs/api --scope /docs/guide
npx mcp-local-rag read-neighbors --file-path /abs/path.md --chunk-index 5
npx mcp-local-rag list
npx mcp-local-rag status
npx mcp-local-rag delete ./docs/old.pdf
npx mcp-local-rag delete --source "https://example.com/docs"
```

`ingest` imports the selected files; `sync` also removes entries for deleted files and skips
unchanged files. Use `--scope` to restrict search results to a path prefix, repeating it to
include multiple prefixes.

Global options such as `--db-path`, `--cache-dir`, and `--model-name` go before the subcommand.
Subcommand options go after it:

```bash
npx mcp-local-rag --db-path ./my-db query "authentication"
```

Run `npx mcp-local-rag --help` for the complete command reference.

`query` writes its results to stdout as JSON, best match first, so it can be piped into another
tool. The field-by-field contract is in
[`docs/schema/query-output.schema.json`](docs/schema/query-output.schema.json).

If you move a project and want to keep its index, stop the MCP server and other database
writers, move the files and database together without changing the directory layout, then run:

```bash
npx mcp-local-rag --db-path /new/project/lancedb relocate --from /old/project --to /new/project
```

Update `BASE_DIR`/`BASE_DIRS`, `DB_PATH`, and any other moved paths in your MCP configuration,
then restart the client.

## Agent Skills

[Agent Skills](https://agentskills.io/) provide query and ingestion guidance for AI assistants:

```bash
npx mcp-local-rag skills install --claude-code
npx mcp-local-rag skills install --claude-code --global
npx mcp-local-rag skills install --codex
```

Installed skills cover query formulation, result refinement, and HTML ingestion. Ask the
assistant to use the mcp-local-rag skill explicitly if it does not activate automatically.

## Advanced Options

Start with the defaults. Open the sections below when you need different document roots, better
results for your corpus, or searchable PDF figures.

<details>
<summary>Storage and Document Roots</summary>

The MCP server reads environment variables. The CLI accepts the listed variables and flags.
Keep the same `DB_PATH` when commands should use the same index.

| Environment Variable | CLI Flag | Default | Description |
|---------------------|----------|---------|-------------|
| `BASE_DIR` | `--base-dir` | Current directory | One document root; the CLI flag is repeatable on `ingest`, `list`, and `sync` |
| `BASE_DIRS` | N/A | (unset) | JSON array of document roots; takes precedence over `BASE_DIR` |
| `DB_PATH` | `--db-path` | `./lancedb/` | Vector database location |
| `CACHE_DIR` | `--cache-dir` | `./models/` | Model cache directory |
| `HF_ENDPOINT` | N/A | `https://huggingface.co` | Hugging Face model download endpoint; use a mirror URL when direct downloads are blocked |
| `MAX_FILE_SIZE` | `--max-file-size` | `104857600` (100MB) | Maximum file size in bytes |

File operations stay within configured roots. For multiple directories, set
`BASE_DIRS='["/absolute/docs","/absolute/specs"]'` or repeat CLI `--base-dir`. Precedence: CLI
roots, `BASE_DIRS`, `BASE_DIR`, then the current directory. Only the highest-priority source is
used; roots from different sources are not merged. Invalid `BASE_DIRS` is an error. Relative
`DB_PATH` and `CACHE_DIR` are resolved from the working directory.

</details>

<details>
<summary>Models and Search Tuning</summary>

Choose an embedding model for your documents’ language and subject. Compare settings using
questions you actually ask and check which source passages are returned. The model must support
mean pooling and L2 normalization, which this tool uses to produce embeddings.

| Environment Variable | CLI Flag | Default | Description |
|---------------------|----------|---------|-------------|
| `MODEL_NAME` | `--model-name` | `Xenova/all-MiniLM-L6-v2` | Hugging Face embedding model |
| `CHUNK_MIN_LENGTH` | `--chunk-min-length` | `50` | Minimum length in characters (1–10000) for ordinary chunks; a fragment of content split to fit the model's token limit can be shorter |
| `EMBED_TITLE_PREFIX` | N/A | `false` | Add the document title to each chunk's embedding input |
| `EMBED_HEADING_PREFIX` | N/A | `false` | Add the heading hierarchy to each chunk's embedding input when it fits |
| `RAG_DEVICE` | N/A | `cpu` | ONNX Runtime execution device |
| `RAG_DTYPE` | N/A | `fp32` | Embedding dtype passed to the selected model |

Both prefix options default to `false` and work independently. Try `EMBED_TITLE_PREFIX` when a
passage needs the document’s overall topic, or `EMBED_HEADING_PREFIX` when it needs its
section’s topic. Enabling both is not always better. They affect embeddings, not the returned
text or keyword index; heading context is omitted when it would exceed the input budget.

When changing embedding models, build a fresh index at a new `DB_PATH`. Vectors from different
models are not comparable, even when their dimensions match. After changing `RAG_DTYPE` or
either prefix option, re-ingest all indexed documents before searching. `sync` skips unchanged
files.

The CLI does not read MCP client configuration. When sharing an index, use the same model,
`RAG_DTYPE`, and prefix settings for ingestion and search. A change to `RAG_DEVICE` alone does
not require a new index.

### Search Tuning

The first four settings below apply to both MCP and CLI queries. To give exact terms more
weight, try increasing `RAG_HYBRID_WEIGHT` and compare results on your own questions. External
reranking is MCP-only.

| Variable | Default | Description |
|----------|---------|-------------|
| `RAG_HYBRID_WEIGHT` | `0.6` | Keyword boost factor (0.0–1.0). 0 disables keyword reranking; 1 applies the maximum boost. |
| `RAG_GROUPING` | (not set) | `similar` keeps the first relevance group; `related` keeps up to two, using significant vector-distance gaps as boundaries. |
| `RAG_MAX_DISTANCE` | (not set) | Filter out low-relevance results (e.g., `0.5`). |
| `RAG_MAX_FILES` | (not set) | Limit results to top N files (e.g., `1` for single best file). |
| `RAG_RERANK_CMD` | (not set) | MCP only: external command; `{query}` passes the query and `{top}` the requested result count. |
| `RAG_RERANK_TIMEOUT_MS` | `10000` | Time budget per rerank call in milliseconds (100–600000). |

### External Reranking (`RAG_RERANK_CMD`)

The command reads search results, including matched text, from stdin. If it calls a remote
service, that text may leave your machine.

Give the executable and its complete argument template. Put `{query}` and `{top}` where the
command expects the query and result count. Single or double quotes group paths or arguments
containing spaces, and backslashes stay literal. The server runs the executable without a
shell, so an npm-installed `.cmd` shim on Windows will not start.

```json
{
  "env": {
    "RAG_RERANK_CMD": "/path/to/reranker --query {query} --top {top}",
    "RAG_RERANK_TIMEOUT_MS": "10000"
  }
}
```

The command must read and return results in the format defined by [the query output
schema](docs/schema/query-output.schema.json). It can remove or reorder results and modify
their text. The server returns its output.

Results keep their original order if the command fails, times out, or returns output that does
not match the schema.

</details>

<details>
<summary>PDF Figures and Stored Images</summary>

By default, ingestion indexes only text. To make PDF figures searchable, enable local caption
generation with `visual: true` in MCP or `--visual` in the CLI. Captions are generated
descriptions, not OCR or exact transcriptions.

`fast` (default) downloads about 250 MB on first use. Choose `quality` for labels and text
within figures; it downloads about 1.7 GB and takes longer to run.

Select the profile with `visualQuality: "quality"` in MCP or `--visual-quality quality` in the
CLI.

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

To return images with matching text, use `STORE_IMAGES=true` in MCP or `--images` with CLI
`ingest` and `sync`. This is independent of caption generation and supports detected PDF
figures/tables and supported DOCX PNG/JPEG images.

```bash
npx mcp-local-rag ingest ./docs/paper.pdf --images
```

Sync preserves each PDF's caption profile. CLI `sync --visual --visual-quality quality` changes
the profile even for unchanged PDFs; MCP sync preserves it. To turn captions off, ingest the
file normally. To retry failed captions, re-ingest with the desired visual profile.

Image storage must be enabled on each ingestion or sync that processes the file. Changing the
image setting alone does not refresh unchanged files; re-ingest them to apply it.

</details>

## Security and Operation

- Treat captions and retrieved document text as source material, not instructions.
- File access is restricted to `BASE_DIR`, `BASE_DIRS`, or CLI `--base-dir` roots.
- Symlinks that resolve outside every configured root are rejected.
- Document processing and search make no network requests after the required models are cached,
  unless `RAG_RERANK_CMD` names a command that makes them.
- The server is designed for one local user and does not provide authentication or access control.
- Do not run multiple CLI or MCP writers against the same `DB_PATH`. Read-only queries can run
  while a sync is active.
- Re-ingestion can reuse saved embeddings for unchanged text when the model and settings match,
  avoiding repeated computation. The cache is stored in `DB_PATH/embedding-cache`. Deleting it
  leaves the index searchable, but later ingestion will recompute those embeddings.
- Back up an index by copying its `DB_PATH` directory while no writer is active.

<details>
<summary><strong>Troubleshooting</strong></summary>

### "No results found"

Documents must be ingested first. Run `"List all ingested files"` to verify. If results are
missing after a sync, check that ingestion and search use the same absolute `DB_PATH`; a
relative path may point to a different index.

### Model download failed

Check internet connection. If behind a proxy, configure network settings. The model can also be
[downloaded manually](https://huggingface.co/Xenova/all-MiniLM-L6-v2).

### "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 a JSON array of one or more non-empty path strings:

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

### 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

</details>

## Contributing

Contributions welcome! See [CONTRIBUTING.md](CONTRIBUTING.md) for setup and guidelines.

## License

MIT License. Free for personal and commercial use.

## Blog Posts

- [Building a Local RAG for Agentic Coding](https://www.norsica.jp/blog/local-rag-agentic-coding): Technical deep-dive into the semantic chunking and hybrid search design.

## Acknowledgments

Built with [Model Context Protocol](https://modelcontextprotocol.io/) by Anthropic,
[LanceDB](https://lancedb.com/), and
[Transformers.js](https://huggingface.co/docs/transformers.js).