Note: This is a personal project meant for my Journal, this is not a general purpose RAG mcp server

search-journal

To install dependencies:

bun install

Index your journal

Start Ollama and make sure the embedding model is available:

ollama pull embeddinggemma-2:270m
bun run embedFiles.ts

This reads Markdown files in ~/notes/Journal, splits them into 512-word chunks, embeds them with Ollama, and stores the text and vectors in journal.sqlite. Re-running the command replaces each scanned file's previous chunks, including clearing chunks for files that are now empty. Deleted files are not automatically removed from the index.

Related MCP server: lore-mcp

Start the MCP server

bun run index.ts

The MCP server uses Streamable HTTP at http://127.0.0.1:8000/mcp. Configure that URL in your MCP client (HTTP transport, not stdio). It exposes a search tool with query and an optional limit (1–10, default 5). It embeds the query and returns the closest indexed passages, including their file, zero-based chunk index, and cosine distance (lower is closer). An empty index returns no results. Indexing is a separate command; the server does not re-embed your journal at startup. Keep Ollama running for searches.

Both commands use the same defaults, configurable through environment variables:

To listen on other interfaces or use a different port:

HOST=0.0.0.0 PORT=8000 bun run index.ts

The server has no authentication and provides access to private journal content. Keep it on localhost unless access is restricted by a firewall or an authenticated reverse proxy. Cross-origin browser requests are rejected. HTTP requests are stateless; no session ID is required.

Use the same embedding model for indexing and search. If changing models, use a fresh database and re-index; existing vectors cannot be mixed with another model.

Docker Compose

Requires Docker with Compose and host networking (native on Linux; Docker Desktop 4.34+ requires opting in to host networking). Ollama runs separately on the host; Compose does not install it or download models.

Make sure your journal directory exists and is readable by the container's non-root user, then build, index, and start:

ollama pull embeddinggemma-2:270m
docker compose build
docker compose run --rm search-journal bun run embedFiles.ts
docker compose up -d

The MCP endpoint is http://127.0.0.1:8000/mcp. Compose uses host networking and binds the server to localhost by default, so no port publishing is needed and host-local Ollama remains reachable without exposing Ollama to the network.

The journal defaults to ~/notes/Journal on the host and is mounted read-only at /journal. Override JOURNAL_DIR, OLLAMA_URL, EMBEDDING_MODEL, HOST, or PORT using your shell environment or a Compose .env file, for example:

JOURNAL_DIR=/absolute/path/to/your/journal
OLLAMA_URL=http://127.0.0.1:11434
EMBEDDING_MODEL=embeddinggemma-2:270m
PORT=8000

SQLite is stored at /data/journal.sqlite in the journal-data named volume, shared by the server and indexing commands. The existing local journal.sqlite is not copied into the image or imported; the commands above create a fresh index. Search results report container paths under /journal.

After changing journal files, re-run the indexing command. To avoid concurrent SQLite writes during indexing, you can stop the server first:

docker compose stop
docker compose run --rm search-journal bun run embedFiles.ts
docker compose up -d
docker compose logs -f search-journal

docker compose down preserves the index; docker compose down -v deletes it. The image build excludes local databases, journal files, .env files, and node_modules. Do not set HOST=0.0.0.0 without restricting access: with host networking this exposes the unauthenticated server on all host interfaces.

Tests

bun test

Tests use temporary databases and mocked embeddings, not your journal or Ollama.