🪄 Searchless-ngx

Stop searching your documents. Start asking them. An Agentic RAG MCP Server for Paperless-ngx.

Less searching. More finding. Searchless-ngx transforms your Paperless-ngx instance from a static, keyword-based archive into an intelligent, conversational agent. By leveraging the Model Context Protocol (MCP) and Agentic RAG, it allows modern LLMs to natively understand, search, filter, and reason over your documents.

🤔 About the Name

If Paperless freed you from the burden of physical paper, Searchless frees you from the burden of manual searching.

• Serverless means you don't manage servers.

• Passwordless means you don't type passwords.

• Searchless means you don't click through filters or skim 20-page PDFs anymore. You just ask your assistant a question, and it does the heavy lifting for you. The -ngx pays homage to the incredible Paperless-ngx project that makes all of this possible.

(Note: Under the hood, the technical service is named paperless-mcp-server to provide optimal contextual grounding for the AI).

✨ Key Features

• Agentic RAG: Equips your LLM with tools to query, filter, and summarize your personal documents.

• Hybrid Search Strategy:

  • Exact Metadata API: Leverage Paperless-ngx's powerful filtering (tags, correspondents, dates) for precise retrieval.

  • Semantic Vector Search: Use ChromaDB and Mistral embeddings (Google Gemini optional) to find documents based on meaning and context (e.g., "software subscriptions", "food receipts").

• Optimized for Open WebUI:

  • Strict JSON Schema: Zero anyOf or null types to ensure 100% compatibility with experimental MCP parsers.

  • Interactive Cards: Search results are presented as beautiful Markdown cards with clickable titles and metadata.

• Read-Only: Zero destructive actions. It uses existing OCR text and never downloads binary PDFs.

• Smart Sync: Startup sync uses a watermark to fetch only new/changed documents from Paperless in seconds. Periodic background sync (default: every 15 min, configurable) keeps the index continuously up to date. Webhook support for real-time ingestion of individual documents. Manual full-sync via POST /sync/all.

• Search Resilience: Proactive fallback strategies ensure the LLM finds documents even when initial filters are too restrictive.

Related MCP server: paperlessngx-mcp

🏗️ Architecture

graph TD
    User([User]) -->|Chat| OWUI(Open WebUI)
    OWUI -->|MCP Streamable HTTP| MCP(FastAPI MCP Server)
    MCP -->|Mistral / Gemini Embeddings| Chroma[(ChromaDB)]
    MCP -->|"Read-Only API (metadata, content, sync)"| Paperless(Paperless-ngx)
    MCP -->|Paginated Cache| Cache[(In-Memory Metadata)]

    External([Paperless Workflow / External]) -->|POST /webhook/sync| MCP
    Timer([Periodic Sync\nevery 15 min]) -->|bulk_sync_documents| MCP

🚀 Setup & Installation

1. Prerequisites

• Docker & Docker Compose

• Paperless-ngx instance

• A Mistral API Key (default provider) or a Google Gemini API Key (alternative)

2. Environment Configuration

Copy .env.example to .env and configure:

cp .env.example .env

Provider selection (Mistral / Google / OpenAI / Ollama)

LLM_PROVIDER is the umbrella default (Mistral: mistral-embed embeddings + mistral-small-latest for fuzzy matching). Set it to google, openai or ollama to switch. openai and ollama share one OpenAI-compatible client, so any OpenAI-compatible endpoint (vLLM, TEI, LM Studio, …) works by pointing OPENAI_BASE_URL / OLLAMA_BASE_URL at it.

Best-of-breed: the embedding and chat axes can be split independently via EMBEDDING_PROVIDER and CHAT_PROVIDER (each falls back to LLM_PROVIDER). Google's gemini-embedding-001 leads multilingual retrieval benchmarks, so a common combination is Google embeddings + Mistral (EU) chat:

EMBEDDING_PROVIDER=google   # stronger retrieval
CHAT_PROVIDER=mistral       # EU-hosted chat
# both MISTRAL_API_KEY and GEMINI_API_KEY required

⚠️ Switching the embedding provider requires rebuilding the vector DB. Embeddings from different models live in different vector spaces (and dimensions: Mistral = 1024, Gemini = 3072), so the existing index is unusable. Wipe the chroma_data volume and let the startup sync re-embed everything:

docker compose down
docker volume rm $(docker compose config --volumes | grep chroma)
docker compose up -d

Fully local & private: Qwen3 embeddings + Mistral Small 24B on Ollama

Run everything on your own machine (e.g. a MacBook Pro M4 with ≥32 GB RAM). Install Ollama natively on the host — Docker on macOS has no GPU access, so the containers should call out to the host's Ollama via host.docker.internal.

# 1. On the host: install Ollama (macOS)
brew install ollama

# 2. Pull the models (pin explicit tags — see table below) and expose Ollama to containers
ollama pull qwen3-embedding:0.6b
ollama pull mistral-small:24b
OLLAMA_HOST=0.0.0.0 ollama serve        # bind so Docker can reach it

# 3. In .env
LLM_PROVIDER=ollama
OLLAMA_BASE_URL=http://host.docker.internal:11434/v1
EMBEDDING_MODEL=qwen3-embedding:0.6b
CHAT_MODEL=mistral-small:24b

Embedding model options (always pin a tag — bare qwen3-embedding / :latest = 8B):

No API key is needed for ollama. Embeddings stay 100 % on-device — nothing leaves your machine. (You can also mix, e.g. EMBEDDING_PROVIDER=ollama for private local embeddings while keeping CHAT_PROVIDER=mistral for the cloud chat.)

💡 The embedding dimension is baked into the ChromaDB collection. Switching tags (e.g. :0.6b → :8b) changes the dimension and requires a vector-DB rebuild.

Model loading & memory: Ollama loads a model into memory on first use and automatically unloads it after 5 minutes idle (OLLAMA_KEEP_ALIVE, default 5m) to free RAM. Since the periodic sync runs every 15 min, the embedding model is unloaded between runs and reloaded on each sync — for qwen3-embedding:0.6b that reload is negligible (well under a second), so the default 5m is a good fit and needs no change. If you'd rather keep it permanently warm, set OLLAMA_KEEP_ALIVE=-1 on the host (costs ~0.6 GB RAM for the 0.6B model).

3. Docker Compose

Start the agent and Open WebUI:

docker compose up -d

4. Connect to Open WebUI

1. Open http://localhost:8080.

2. Go to Settings > Connections > MCP Servers.

3. Preferred Method: Click the import button and select scripts/webui-connection.json.

4. Manual Method: Add a new server with type MCP Streamable HTTP and URL http://mcp-server:8001/mcp.

For detailed Open WebUI instructions, see WEBUI_SETUP.md.

💡 Usage Examples

Listing Documents

• "List the last 5 documents from Amazon." (Uses exact metadata search)

• "Show me my most recent invoices." (Uses empty query to fetch by date)

Conceptual Search (Semantic)

• "Find all software subscriptions I have." (Finds "Netflix", "Adobe", "Microsoft" even if "subscription" isn't in the title)

• "Where are my food receipts from my last trip to Berlin?" (Combines location context with document meaning)

Data Extraction

• "How much did I spend on mobility in February 2024?" (LLM iterates through scouter/train invoices and calculates the sum)

• "Summarize the cancellation terms for my gym contract." (LLM uses get_document_details to read the full OCR text)

🛠️ Development & Testing

Diagnostic Tools

Use the raw protocol checker to verify the server's output:

docker exec paperless-mcp-server python scripts/test_mcp_raw.py

Test Coverage

Run the test suite using uv:

uv run pytest

👤 Author

Developed and maintained by Dr. Henning Dickten (@hensing).

⚖️ License

Licensed under the GPLv3.