dlightrag 1.5.17

Multimodal RAG with knowledge graph and contextual intelligence

DlightRAG

DlightRAG is a multimodal RAG service built on LightRAG main. LightRAG owns document parsing, staged ingest, chunks, document status, vectors, and the knowledge graph. DlightRAG adds product-layer metadata governance, PostgreSQL BM25, direct image-vector alignment, answer orchestration, citations, REST, Web, SDK, and MCP interfaces.

Status: Python 3.12. Storage: PostgreSQL 18 with pgvector, Apache AGE, pg_textsearch, and pg_jieba. 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 production RAG path: DlightRAG stages sources and metadata, LightRAG performs parser/KG/vector/chunk/doc-status work, and DlightRAG adds metadata filtering, BM25, direct visual retrieval, cited answers, jobs, and interfaces. The full runtime and code-layer view is in docs/architecture.md.

Choose Your Deployment Path

Path	Use this when	PostgreSQL	Parser endpoint	Security
Local Docker	Developer machine, Web UI, smoke tests	Compose PG18	Host-native MinerU-compatible sidecar	auth_mode: none on loopback
Native API	API process runs on host, PostgreSQL stays in Docker	Compose PG18	Any reachable MinerU-compatible endpoint	Local or explicit auth
Shared service	Remote users, agents, team workspace	Managed or self-hosted PG18	Official MinerU API or independent parser service	simple or jwt
Enterprise	Multi-user internal product	Managed PG18	Independently operated parser service	jwt + JWKS, optional claim access control

Do not install MinerU into the DlightRAG app container. DlightRAG consumes the MinerU-compatible HTTP endpoint that LightRAG expects. On macOS, keep MinerU as a native host process so MLX/MPS acceleration is available. On Linux GPU, run MinerU as an independent service/router or use the official API.

Quick Start

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=...

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. Install and start a native MinerU sidecar if one is not already running:

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. Docker Compose does not run MinerU; it maps the host-native endpoint into DlightRAG containers as http://host.docker.internal:8210.

3. Start DlightRAG and PostgreSQL:

docker compose up -d
docker compose ps
curl http://localhost:8100/health

This starts:

Service	Purpose	Host port
dlightrag-api	REST API + Web UI	127.0.0.1:8100
dlightrag-mcp	MCP streamable HTTP server	127.0.0.1:8101
lightrag-gui	Upstream LightRAG graph browser	127.0.0.1:9621
postgres	PG18 + pgvector + AGE + pg_textsearch + pg_jieba	5432

4. Open the Web UI:

http://localhost:8100/web/

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

Use DlightRAG

Web

The Web UI is served by the REST API at /web/. It supports workspace selection, file/folder upload, chat, session image memory, citations, source panels, and semantic highlights.

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",
                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.create(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.close()


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, ingest_job_status, list_files, delete_files, list_workspaces, create_workspace, and delete_workspace.

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.

Source retention. Remote source files are transient by default for S3, Azure Blob, URL, and SDK connectors. Set retain_remote_source_files: true, or pass retain_source_file: true on one ingest call, when fetched files should be kept under the workspace input root.

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. If Langfuse keys are absent, tracing is a no-op. Configuration is documented in docs/configuration.md.

Security Model

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

Mode	Use case
simple	One shared bearer token
jwt	Externally issued signed tokens
jwt + JWKS	OIDC-style issuers with key rotation
jwt + jwt_claims access control	Workspace/action permissions from verified claims

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

Offline vector maintenance:

uv run dlightrag-rebuild-vdb --target check
uv run dlightrag-rebuild-vdb --target all --yes

Use docs/operations.md for rebuild 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

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, and citations.
• docs/postgresql.md - PostgreSQL requirements and tuning.
• docs/operations.md - maintenance commands and recovery workflows.
• docs/evaluation.md - RAGAS evaluation workflow.
• LightRAG API Server docs - upstream parser routing and MinerU official API contract.
• MinerU Docker deployment docs - Linux/WSL2 Docker support and macOS warning.

License

Apache License 2.0. See LICENSE.

Built by HanlianLyu. Contributions welcome.