DlightRAG

DlightRAG is a production ready multimodal RAG service built on LightRAG. It offers superior context intelligence, great accuracy with citation / highlight grounding, and unified interfaces for REST, Web, MCP, and Python SDK clients. It is designed for developers, seasoned users and teams who need a reliable RAG core service with cutting edge features integrated into their workflows and products.

Status: Python 3.14. Storage: PostgreSQL 18 ecosystem. License: Apache-2.0.

Architecture At A Glance

Clients
  -> REST / Web / MCP / SDK adapters
  -> RAGServiceManager
  -> RAGService
  -> LightRAG main
  -> PostgreSQL 18 storage ecosystem

DlightRAG has one unified production RAG path: LightRAG provides fusional one-hop graph traversal and vector retrieval. DlightRAG adds product-layer metadata governance, hybrid BM25 sparse retrieval, fused visual-vector alignment, orchestration, citations, highlighting and standardized interfaces. The full runtime and code-layer view is in docs/architecture.md.

Choose Your Deployment Path

Do not install a parser into the DlightRAG app container. Configure one parser_sidecars.mineru or parser_sidecars.docling block; DlightRAG derives LightRAG routing automatically. If both blocks are present, MinerU takes priority. On macOS, keep MinerU as a native host process for MLX/MPS, or use the optional Docling Compose profile.

Quick Start

Prerequisites. Install Docker + Compose (runs the API and PostgreSQL), uv (builds the isolated MinerU sidecar environment), plus git and make. DlightRAG targets Python 3.14; uv installs it for you (uv python install 3.14), so a system Python is not required for the Docker path.

# Install uv — macOS/Linux (see the uv docs for the Windows PowerShell command)
curl -LsSf https://astral.sh/uv/install.sh | sh

# Install make if you don't have it (git is usually already present):
#   macOS          xcode-select --install       # or: brew install make
#   Debian/Ubuntu  sudo apt-get install -y make
#   Fedora/RHEL    sudo dnf install -y make
#   Windows        no POSIX make on Windows — use WSL2 (Ubuntu): install Docker
#                  Desktop with the WSL2 backend, then inside WSL2 follow the
#                  Debian/Ubuntu line above. uv, make, and Python 3.14 all live
#                  inside WSL2 (a Linux environment), not on Windows.

The Docker Quick Start does not require uv sync — DlightRAG itself runs inside containers, and make mineru-install builds its own isolated .venv-mineru. Run uv sync only for the Native API Variant or development.

One-command setup (recommended)

From a fresh clone, an interactive wizard configures your models, selects MinerU (local/official) or Docling (bundled/external), brings up the stack, and ends with a clickable Web UI link:

git clone https://github.com/hanlianlu/dlightrag.git
cd dlightrag
uv run prerequisite_setup.py

It writes config.yaml and .env for you (with timestamped backups) and is safe to re-run. The wizard does not preserve the checked-in model choices: the minimum path writes only llm.default plus embedding, while the custom path replaces role-specific LLM blocks with the roles you choose. Prefer the manual steps below if you'd rather configure everything by hand.

Manual setup

1. Clone the repo and create a secrets file:

git clone https://github.com/hanlianlu/dlightrag.git
cd dlightrag
cp .env.example .env

Fill secrets in .env:

DLIGHTRAG_LLM__DEFAULT__API_KEY=...
DLIGHTRAG_EMBEDDING__API_KEY=...
DLIGHTRAG_LLM__ROLES__EXTRACT__API_KEY=...
DLIGHTRAG_LLM__ROLES__KEYWORD__API_KEY=...
DLIGHTRAG_LLM__ROLES__QUERY__API_KEY=...
DLIGHTRAG_LLM__ROLES__VLM__API_KEY=...
DLIGHTRAG_RERANK__API_KEY=...

These match the checked-in config.yaml, which configures DeepSeek extract and keyword roles, host-local Unsloth query and VLM roles, plus Voyage reranking. The Unsloth roles use their own query/VLM API key entries. Role overrides are atomic: an omitted or blank key falls back to the complete default model configuration instead of combining its endpoint with the default key. If you remove role-specific model blocks or switch rerank back to chat_llm_reranker, the corresponding role/rerank keys can be omitted. Reserve api_key: null for a genuinely unauthenticated endpoint.

Normal behavior lives in config.yaml: model names, parser sidecar settings, metadata schema, retrieval breadth, auth mode, Langfuse behavior, and deployment endpoints. Deep config reference is in docs/configuration.md.

2. Choose one parser. The checked-in config uses a native MinerU sidecar:

cp .env.mineru.example .env.mineru
make mineru-install
make mineru-api

make mineru-api serves http://127.0.0.1:8210 by default and blocks in the current terminal. Its config block uses http://host.docker.internal:8210, which is reachable from the DlightRAG containers.

To run MinerU in the background instead (launchd on macOS, systemd --user on Linux/WSL2), use the service targets — one command starts both the API backend and the Gradio WebUI. The WebUI reuses that same backend automatically, so it needs no extra setup and never loads a second copy of the models:

make mineru-service-install   # install + start at login
make mineru-service-status    # also: -start / -stop / -logs / -uninstall

The WebUI opens at http://127.0.0.1:7860 (unauthenticated — keep it on loopback). Set MINERU_GRADIO_ENABLE=false in .env.mineru to manage the API backend alone.

Alternatively, comment/remove the MinerU block in config.yaml, enable the commented Docling block, and start the optional official CPU image:

docker compose --profile docling up -d

The image follows quay.io/docling-project/docling-serve-cpu:latest. Set a different official image with DOCLING_SERVE_IMAGE; external Docling users set the block's endpoint instead and do not enable the profile.

3. Start DlightRAG and PostgreSQL:

docker compose up -d
docker compose ps

This starts:

4. Open the Web UI:

http://localhost:8100/web/

Upload documents or images from the Files panel, then ask a question.

Native API Variant

Use this when the API process should run on the host while PostgreSQL stays in Docker:

docker compose up -d postgres
uv sync
DLIGHTRAG_PARSER_SIDECARS__MINERU__LOCAL_ENDPOINT=http://127.0.0.1:8210 \
  uv run dlightrag-api

Native runs can ingest host paths directly because the API process sees the same filesystem as your shell. The checked config is Docker-first, so the command overrides its Docker host alias with the native loopback endpoint. A native Docling deployment likewise sets its active block endpoint to http://127.0.0.1:5001.

Use DlightRAG

Web

The Web UI is served by the REST API at /web/. It supports workspace selection, file/folder upload, durable principal-scoped conversations and current-turn images, citations, source panels, and semantic highlights. The Web-only conversation lifecycle provides New chat, select, rename, delete, and reload persistence with 30-day inactivity retention. Search in: All authorized workspaces is the answer default; the independent Files in selector remains a single-workspace file-management target.

REST, MCP, and Python answer/retrieve calls remain stateless. Answer calls accept an optional caller-supplied history of prior turns for multi-turn follow-ups, but never a Web conversation ID or server-stored history: the client owns conversation storage and re-sends the turns it wants on each request. Web current-image admission is configurable and defaults to three images at 15 MiB each. Durable current-turn images always have priority over any historical image selection; historical images that miss a transport slot contribute their stored text descriptions rather than raw pixels.

REST

REST ingest starts durable background jobs. Poll the job endpoint for status.

curl -X POST http://localhost:8100/ingest \
  -H "Content-Type: application/json" \
  -d '{"source_type": "local", "path": "report.pdf"}'

curl http://localhost:8100/ingest/jobs/<job_id>

curl -X POST http://localhost:8100/answer \
  -H "Content-Type: application/json" \
  -d '{"query": "What are the key findings?", "stream": false}'

All SDK, REST, MCP, Web contracts and response shapes are in docs/interfaces.md.

Python SDK

uv add dlightrag

import asyncio
import os

from dlightrag import IngestSpec, RAGServiceManager
from dlightrag.config import DlightragConfig, EmbeddingConfig, LLMConfig, ModelConfig


async def main() -> None:
    workspace = "research_notes"
    config = DlightragConfig(
        workspace=workspace,
        working_dir="./dlightrag_storage/sdk_demo",
        llm=LLMConfig(
            default=ModelConfig(
                provider="openai",  # protocol family: openai | anthropic | gemini (vendor via base_url)
                model="gpt-4.1-mini",
                api_key=os.environ["OPENAI_API_KEY"],
                temperature=0.2,
            )
        ),
        embedding=EmbeddingConfig(
            provider="openai_compatible",
            model="text-embedding-3-large",
            api_key=os.environ["OPENAI_API_KEY"],
            base_url="https://api.openai.com/v1",
            dim=3072,
        ),
    )
    manager = await RAGServiceManager.acreate(config)
    try:
        await manager.aingest(
            workspace,
            IngestSpec(source_type="local", path="./docs"),
        )
        answer = await manager.aanswer("What are the key findings?", workspace=workspace)
        print(answer.answer)
    finally:
        await manager.aclose()


asyncio.run(main())

config.yaml is optional for SDK users; constructor values take precedence.

MCP

Use stdio when an agent starts DlightRAG as a subprocess:

{
  "mcpServers": {
    "dlightrag": {
      "command": "uvx",
      "args": ["dlightrag-mcp", "--env-file", "/absolute/path/to/.env"]
    }
  }
}

Use streamable HTTP when multiple clients connect to a running service:

DLIGHTRAG_MCP_TRANSPORT=streamable-http \
DLIGHTRAG_MCP_HOST=127.0.0.1 \
dlightrag-mcp

MCP tools include retrieve, answer, ingest, get_ingest_job, list_files, delete_files, list_workspaces, create_workspace, and delete_workspace, plus get_capabilities for image and metadata-filter capability discovery. See docs/interfaces.md for the authoritative tool-result contract.

Core Concepts

Workspaces. A workspace is the primitive isolation unit for indexed data, metadata, jobs, files, and queries. Query calls can target one workspace or federate across multiple workspaces.

Ingestion sources. Local files, Web uploads, S3, Azure Blob, public/signed HTTPS URLs, and SDK AsyncDataSource connectors flow through the same ingest contract. Web and REST uploads are staged under DlightRAG's managed working_dir/inputs/<workspace>/ tree, then copied into the workspace input root as retained local sources. Upload batch staging under __uploads__/ is cleaned by the durable ingest job after the handoff.

Source downloads. Every successful ingest remains downloadable, whether or not DlightRAG retains a local copy. source_uri is stable provenance; download_uri is the durable S3, Azure, or queryless public HTTPS locator used when retain_source_file is false. Signed HTTPS fetch URLs need either a separate durable locator or retention. A non-retained custom SDK connector must provide SourceDocument.download_uri (or download_uri_for_key). DlightRAG rejects a document before materialization when this contract cannot be met; it never silently changes the caller's retention choice.

Runtime storage. Docker Compose stores working_dir in the dlightrag_data named volume mounted at /app/dlightrag_storage; the host ./dlightrag_storage directory is only used by native, non-Docker runs.

Metadata. Declare filterable custom fields once in configuration. Ingest calls pass values. Request-level metadata is the batch default; manifest or SourceDocument metadata overlays it per document.

Retrieval and answers. DlightRAG uses LightRAG mix as the base retrieval mode, then adds metadata filtering, BM25, optional direct image retrieval, RRF fusion, reranking, answer packing, citations, and optional semantic highlights. The detailed mechanism is in docs/retrieval-answer.md.

Observability. Langfuse tracing is optional. Non-secret SDK behavior is set in docs/configuration.md. To run the bundled local Langfuse stack (make langfuse-up) and view traces, see docs/operations.md.

Security Model

Local loopback development can use auth_mode: none. Shared or exposed deployments should enable auth:

DlightRAG verifies bearer tokens and can enforce workspace/action access control. It does not issue OAuth tokens or manage users. Use an external IdP or gateway for login and token issuance. Full guidance is in docs/security.md.

Operations And Development

Use docs/operations.md for the full stop/rebuild/restart sequence and maintenance safety notes.

Development setup:

uv sync
cd frontend && npm ci && cd ..
make hooks

Verification:

make ci          # lint + type-check + architecture + unit tests
make ci-full     # above + integration tests
make ci-e2e      # above + E2E smoke

Frontend checks after editing frontend/:

cd frontend
npm run typecheck
npm run build
npm run lint:css

npm run build writes the browser bundle to src/dlightrag/web/static/generated/; commit those generated files with frontend changes.

Evaluation with RAGAS is documented in docs/evaluation.md.

Documentation Map

• docs/architecture.md - runtime ownership, storage topology, and code layering.
• docs/interfaces.md - SDK, REST, MCP, and Web contracts.
• docs/security.md - auth, JWT/JWKS, IdP boundaries, and access control.
• docs/configuration.md - configuration precedence, fields, and defaults.
• docs/retrieval-answer.md - retrieval, filters, BM25, fusion, rerank, answers, citations, and highlights.
• docs/postgresql.md - PostgreSQL requirements and tuning.
• docs/operations.md - maintenance commands and recovery workflows.
• docs/evaluation.md - RAGAS evaluation workflow.
• docs/web-theme-design.md - current Web appearance state, tokens, interaction, and accessibility design.
• LightRAG API Server docs - upstream parser routing and external parser contracts.
• MinerU Docker deployment docs - Linux/WSL2 Docker support and macOS warning.
• Docling Serve - official Docling HTTP service and container images.

License

Apache License 2.0. See LICENSE.

Built by HanlianLyu. Contributions welcome.