Search private documents locally with hybrid RAG indexing

MCP Local RAG

English | 简体中文 | Deutsch | Español | Português (Brasil) | Français

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.

Features

• Runs locally: Document parsing, embeddings, storage, and search run on your machine.
  After the initial model download, text ingestion and search work offline.
• Hybrid search: Semantic retrieval finds related concepts, while keyword matching boosts
  exact technical terms.
• Configurable embeddings: Choose a Hugging Face embedding model that fits the language and
  domain of your documents.
• Semantic chunking: Documents are split at topic boundaries instead of fixed character
  counts. Markdown code blocks stay intact.
• MCP and CLI: Use the same index from an AI coding tool or directly from the terminal.

No API key, Docker, Python, or external database is required.

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.

mcp-local-rag uses the standard MCP protocol over a local stdio server, so it works with AI
coding tools and other MCP hosts that support local MCP servers.

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

For Claude Code: Run this command:

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

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 = "/absolute/path/to/your/documents"

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

{
  "$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"
      }
    }
  }
}

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

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

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

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:

What does the API documentation say about authentication?

CLI Quick Start

To use the CLI without an MCP client:

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.

Why This Exists

Some document sets cannot be sent to a hosted embedding service because of confidentiality or
organizational policy. Keeping the index local makes them searchable without adding a per-query
API cost.

Semantic search alone can miss exact identifiers that matter in technical documentation.
Keyword reranking keeps those terms visible without giving up natural-language retrieval.

Supported Content

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.

MCP Tools

Syncing a Document Root

sync_start ingests new and changed files, skips byte-identical files, and removes index entries
for files that no longer exist:

Sync everything under the configured document roots and wait for completion.

The tool returns a jobId immediately. Clients should poll sync_status until its state becomes
succeeded or failed. A changed PDF keeps the visual profile it was indexed with; sync_start
cannot change it. Set STORE_IMAGES=true in the MCP server environment to store supported PDF and
DOCX images for new or changed files selected by sync; unchanged files remain skipped.

Only one sync job is retained by the server process. A newer job replaces a finished record, and
restarting the server discards it.

Ingesting One File

ingest_file accepts PDF, DOCX, TXT, and Markdown. MCP file paths must be absolute and must stay
inside a configured document root:

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

Re-ingesting the same path replaces its existing chunks.

Searching and Reading More Context

What does the API documentation say about authentication?
Find the documented behavior of ERR_CONNECTION_REFUSED.

Results contain the text, source path, title, chunk index, relevance score, and any images stored
on that chunk. MCP returns each image as an image content block paired with its result identity;
CLI query includes an images array of { imageIndex, mimeType, data } on every result. Pass the
chunkIndex and either filePath or source from a result to read_chunk_neighbors when the
answer needs more context:

Read the surrounding chunks for that authentication result.

Both query_documents and list_files accept an optional absolute scope path prefix, or a
list of prefixes. A prefix matches the exact path and its descendants.

Ingesting HTML

Use ingest_data after the MCP client fetches a page:

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

The server extracts the main article, converts it to Markdown, and stores it under the supplied
source identifier. Reusing the same source updates the existing content.

Respect the source site's terms and copyright when indexing external content.

PDF Visual Captions and Stored Images

Visual mode adds a generated caption for figure-heavy PDF pages. It is opt-in and does not load
a vision model during normal ingestion.

Ingest /Users/me/docs/research-paper.pdf with visual: true.

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

Image storage is independent of visual captions. Set STORE_IMAGES=true for the MCP server, or
pass --images to CLI ingestion and sync:

npx mcp-local-rag ingest ./docs/research-paper.pdf --images
npx mcp-local-rag sync ./docs/ --images

PDF storage uses detected figure/table regions. DOCX storage includes only PNG/JPEG images that
the existing Mammoth conversion emits as <img>; charts, SmartArt, and shapes are not separately
rendered. Stored images follow their surrounding text into the final semantic chunk and do not
alter ranking, scores, or result count.

Select the larger model with visualQuality: "quality" over MCP or
--visual-quality quality over CLI. Measured CPU inference was about three times as slow as
fast, though results depend on hardware and model updates.

Updating Existing quality Captions

From 0.18.4 quality runs Qwen3.5-2B; earlier versions ran Qwen2.5-VL-3B. Captions already indexed
keep the wording the old model produced, and sync will not redo them, so re-ingest the files you
want refreshed:

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

Add --images if the file was ingested with it, because a run without it replaces the stored
images. The old model stays on disk. Once nothing else uses it, delete
onnx-community/Qwen2.5-VL-3B-Instruct-ONNX/ from the model cache directory - <cache-dir>, which
defaults to ./models/.

Visual Mode Across Syncs

The profile a PDF was indexed with is recorded, and sync reuses it: a PDF indexed with fast or
quality is re-ingested with that same profile, and a PDF with no recorded profile is ingested as
text.

npx mcp-local-rag sync ./docs/                      # keep each PDF's recorded profile
npx mcp-local-rag sync ./docs/ --visual             # request fast for every PDF in scope
npx mcp-local-rag sync ./docs/ --visual --visual-quality quality

--visual overrides recorded profiles, so it also captions PDFs that were indexed as text.
Changing a profile re-ingests the PDF even when the file itself has not changed; running the same
command again does nothing and loads no model. Image settings are never recorded, so --images
and STORE_IMAGES never cause a re-ingest.

To turn captions off for a path, run ingest on it: a successful normal ingest clears the
recorded profile. To retry a page whose captioning failed, run ingest <path> --visual --visual-quality <profile> with the profile you want — a plain ingest clears it instead. If a
PDF's indexed rows disagree about the profile, sync stops before changing anything and names the
file; re-run it with --visual to settle the profile.

Captions are auxiliary text, not faithful transcriptions. Treat retrieved captions and document
text as untrusted input rather than instructions.

At high limits, matched chunks and their attachments can approach the model/client context ceiling;
choose the query limit with the calling model's available context in mind.

CLI

The CLI uses the same parser, embedder, and vector store without an MCP client:

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

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

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

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

The CLI does not read MCP client configuration. Set the same environment variables or flags if
both interfaces should share an index. In particular, MODEL_NAME and the CLI --model-name
must match for a shared database.

Search Tuning

Keyword boost is enabled by default. Relevance-gap grouping and the distance and file filters are
optional controls for corpora that need tighter result selection.

For API specifications and other documents containing many identifiers, a stronger keyword
weight can improve exact-term ranking:

"env": {
  "RAG_HYBRID_WEIGHT": "0.7"
}

• 0.7: slightly stronger exact-term reranking than the default
• 1.0: maximum keyword boost

How It Works

During ingestion:

1. The parser extracts text for the input format.
2. The semantic chunker finds topic boundaries and preserves Markdown code blocks.
3. Transformers.js creates embeddings locally.
4. LanceDB stores the chunks, metadata, vectors, and full-text index.

During search:

1. The query is embedded with the same model.
2. Vector search retrieves semantically related chunks.
3. Optional distance and relevance-group filters narrow the candidates when configured.
4. Full-text matches boost exact query terms.

Agent Skills

Agent Skills provide query and ingestion guidance for AI assistants:

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.

Configuration

The MCP server reads environment variables. The CLI accepts the listed global environment
variables and flags; image storage on CLI ingestion and sync is enabled only with --images.

Document Roots (BASE_DIR and BASE_DIRS)

mcp-local-rag only allows file operations inside configured roots. For multiple roots,
BASE_DIRS must be a JSON array of non-empty paths:

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

Root configuration is resolved in this order:

1. CLI --base-dir <path> flags (repeatable on ingest, list, and sync)
2. BASE_DIRS
3. BASE_DIR
4. Current directory

Each source replaces the lower-priority source rather than merging with it. Invalid BASE_DIRS
configuration fails instead of falling back to BASE_DIR or the current directory. status
remains available in MCP so the client can report the configuration error.

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
npx mcp-local-rag sync --base-dir /Users/me/work --base-dir /Users/me/specs
BASE_DIRS='["/Users/me/work","/Users/me/specs"]' npx mcp-local-rag list

Storage and Models

DB_PATH and CACHE_DIR are relative to the process working directory by default. Set absolute
paths when the MCP client may start the server from different project directories.

Set MODEL_NAME or pass --model-name to choose a Hugging Face embedding model that fits the
language and domain of your documents.

mcp-local-rag generates embeddings with mean pooling and L2 normalization. When choosing a
model, check whether these settings match its recommended inference setup, since the pooling
method can affect retrieval quality.

Changing MODEL_NAME, RAG_DEVICE, or RAG_DTYPE can make existing vectors incompatible.
Use a new DB_PATH or delete the existing index and re-ingest after changing the embedding
configuration.

An example model for English documents is Xenova/bge-small-en-v1.5.

Security and Operation

• 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.
• 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.
• Back up an index by copying its DB_PATH directory while no writer is active.

Blog Posts

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