docs/LightRAG-API-Server.md
The LightRAG Server is designed to provide a Web UI and API support. The Web UI facilitates document indexing, knowledge graph exploration, and a simple RAG query interface. LightRAG Server also provides an Ollama-compatible interface, aiming to emulate LightRAG as an Ollama chat model. This allows AI chat bots, such as Open WebUI, to access LightRAG easily.
LightRAG v1.5.x adds the new file-processing pipeline, parser routing, multimodal analysis, role-specific LLM/VLM configuration, JSON entity extraction, and several provider/storage changes. Review the v1.5.0rc2 release notes before upgrading a production instance.
LIGHTRAG_PARSER=*:legacy-F
ENTITY_TYPES is no longer supported. Use ENTITY_TYPE_PROMPT_FILE instead, with a YAML profile stored under PROMPT_DIR/entity_type (PROMPT_DIR defaults to ./prompts). A sample template is available at prompts/samples/entity_type_prompt.sample.yml.LIGHTRAG_PARSER) or filename hints affects newly uploaded files. To switch an existing document to another parser engine, delete that document and upload it again.CHUNK_*) affects documents enqueued after the server restarts. Reprocess older documents if you want their stored chunk_options snapshot to match the new settings.i/t/e) requires parsed sidecars plus VLM_PROCESS_ENABLE=true. Existing documents can be reprocessed to run VLM analysis on available sidecars; switching extraction engines still requires delete + re-upload.The release that layers MAX_REQUEST_BODY_BYTES turns it on by default at 1 MiB, where it used to be off and to cover three ingestion routes only. Two things change for clients:
/query*, /api/chat, /api/generate and everything else that is neither an upload nor a text insert. /documents/text and /documents/texts keep a 50 MiB ceiling, and /documents/upload derives its own from MAX_UPLOAD_SIZE, so bulk ingestion is unaffected. Set MAX_REQUEST_BODY_BYTES to any positive value to govern every non-upload route with it, or 0 to turn every ceiling off.top_k / chunk_top_k at most 1000, and the max_*_tokens budgets at most 1,000,000. Clients that relied on unbounded top_k or on multi-megabyte queries need adjusting. These are not configurable by design.Neither change affects a deployment that was already sizing its requests sensibly; both bound how much work one unauthenticated request can ask the server to do.
The release that introduces PIPELINE_SCHEDULING_PAGE_SIZE, MAX_PENDING_DOCUMENTS and MAX_UNACKED_MANUAL_RETRIES (see env.example) also changes the concurrency protocol writers use to coordinate through shared state. It is a one-time, in-place upgrade that writes no marker and no protocol version, so the storage cannot detect a stale writer for you. The requirement is therefore operational:
Stop every old writer before starting a new one against the same storage and workspace. A rolling restart that leaves one old worker — or one old instance sharing the same Redis/PostgreSQL workspace — running is the failure case, not a slower upgrade.
Three things an old writer cannot honour:
/documents/reprocess_failed no longer resets FAILED rows inline. It publishes an intent, freezes ingestion, waits for the pipeline to go idle, and only then rewrites FAILED→PENDING page by page with no worker running. An old writer does not read the freeze flag, so it keeps enqueueing into a window the reset assumes is exclusive.created_at is now the immutable (created_at, id) keyset cursor, written as a UTC ISO-8601 timestamp. Rows an old writer stamps in another format sort inconsistently against it, and a keyset page can then skip or repeat documents.Recommended sequence:
/health, scheduling.drain_waiting_on_workers is false and scheduling.drain_pending_enqueues is 0 when nothing is in flight.No data migration is required. The first sweep after startup is a strict full sweep, so a document an old writer left mid-flight — a row stuck in PARSING/ANALYZING/PROCESSING with no worker behind it — is picked up and reprocessed on its own. If a run genuinely cannot be drained, stopping mid-run is still safe for the same reason; what is not safe is starting the old version again afterwards.
After starting, check the log for a strict-capability warning. All five built-in doc_status backends (JSON, Redis, PostgreSQL, MongoDB, OpenSearch) have every capability. A third-party backend may not, and each gap fails closed rather than degrading quietly: admission answers 503, the source-conflict endpoints answer 501, and a scan keeps re-examining a stale FAILED stub. Startup names each missing capability and what it costs, and the authenticated /health reports the same under capabilities. Set PIPELINE_REQUIRE_STRICT_STORAGE_READS=true to turn those gaps into a startup failure instead. There is no equivalent knob for bounded paging: the paging and typed source-resolution methods are abstract, so a backend without them cannot be constructed at all.
### Install LightRAG Server as tool using uv (recommended)
uv tool install "lightrag-hku[api]"
### Or using pip
# python -m venv .venv
# source .venv/bin/activate # Windows: .venv\Scripts\activate
# pip install "lightrag-hku[api]"
# Clone the repository
git clone https://github.com/HKUDS/lightrag.git
# Change to the repository directory
cd lightrag
# Bootstrap the development environment (recommended)
make dev
source .venv/bin/activate # Activate the virtual environment (Linux/macOS)
# Or on Windows: .venv\Scripts\activate
# make dev installs the test toolchain plus the full offline stack
# (API, storage backends, and provider integrations), then builds the frontend.
# Run make env-base or copy env.example to .env before starting the server.
# Equivalent manual steps with uv
# Note: uv sync automatically creates a virtual environment in .venv/
uv sync --extra test --extra offline
source .venv/bin/activate # Activate the virtual environment (Linux/macOS)
# Or on Windows: .venv\Scripts\activate
# Or using pip with virtual environment
# python -m venv .venv
# source .venv/bin/activate # Windows: .venv\Scripts\activate
# pip install -e ".[test,offline]"
# Build front-end artifacts
cd lightrag_webui
bun install --frozen-lockfile
bun run build
cd ..
LightRAG necessitates the integration of both an LLM (Large Language Model) and an Embedding Model to effectively execute document indexing and querying operations. Prior to the initial deployment of the LightRAG server, it is essential to configure the settings for both the LLM and the Embedding Model.
LightRAG supports these LLM backends:
LightRAG supports these embedding backends:
It is recommended to use environment variables to configure the LightRAG Server. There is an example environment variable file named env.example in the root directory of the project. Please copy this file to the startup directory and rename it to .env. After that, you can modify the parameters related to the LLM and Embedding models in the .env file. It is important to note that the LightRAG Server will load the environment variables from .env into the system environment variables each time it starts. LightRAG Server will prioritize the settings in the system environment variables to .env file.
Since VS Code with the Python extension may automatically load the .env file in the integrated terminal, please open a new terminal session after each modification to the .env file.
If you need to configure different LLMs/VLMs for entity extraction, keyword extraction, final answers, or multimodal analysis, see the Role-Specific LLM/VLM Configuration Guide.
Here are some examples of common settings for LLM and Embedding models:
LLM_BINDING=openai
LLM_MODEL=gpt-4o
LLM_BINDING_HOST=https://api.openai.com/v1
LLM_BINDING_API_KEY=your_api_key
EMBEDDING_BINDING=ollama
EMBEDDING_BINDING_HOST=http://localhost:11434
EMBEDDING_MODEL=bge-m3:latest
EMBEDDING_DIM=1024
# EMBEDDING_BINDING_API_KEY=your_api_key
When targeting Google Gemini, set
LLM_BINDING=gemini, choose a model such asLLM_MODEL=gemini-flash-latest, and provide your Gemini key viaLLM_BINDING_API_KEY(orGEMINI_API_KEY).
LLM_BINDING=ollama
LLM_MODEL=mistral-nemo:latest
LLM_BINDING_HOST=http://localhost:11434
# LLM_BINDING_API_KEY=your_api_key
### Ollama Server context length (Must be larger than MAX_TOTAL_TOKENS+2000)
OLLAMA_LLM_NUM_CTX=16384
EMBEDDING_BINDING=ollama
EMBEDDING_BINDING_HOST=http://localhost:11434
EMBEDDING_MODEL=bge-m3:latest
EMBEDDING_DIM=1024
# EMBEDDING_BINDING_API_KEY=your_api_key
Important Note: The embedding model and asymmetric embedding configuration must be determined before document indexing, and the same settings must be used during the query phase. For certain storage solutions (e.g., PostgreSQL), the vector dimension must be defined upon initial table creation. When changing the embedding model, embedding dimension,
EMBEDDING_ASYMMETRIC, query/document prefixes, or provider task behavior, clear the existing LightRAG workspace/vector data and re-index the source files.
LightRAG uses symmetric embeddings by default. Query/document asymmetric embeddings are enabled only when EMBEDDING_ASYMMETRIC=true is explicitly set.
jina, gemini, and voyageai use provider parameters (task / task_type / input_type) and should not use query/document prefixes.openai, azure_openai, and ollama require both EMBEDDING_QUERY_PREFIX and EMBEDDING_DOCUMENT_PREFIX. Use NO_PREFIX for a side that should intentionally have no prefix.For the full validation rules and examples, see Asymmetric Embedding Configuration.
Instead of editing env.example by hand, you can use the interactive setup wizard to generate a configured .env and, when needed, docker-compose.final.yml:
make env-base # Required first step: LLM, embedding, reranker
make env-storage # Optional: storage backends and database services
make env-server # Optional: server port, auth, and SSL
make env-security-check # Optional: audit the current .env for security risks
For a full description of every target and what each flow does, see docs/InteractiveSetup.md.
The setup wizards update configuration only; run make env-security-check separately to audit the
current .env for security risks before deployment.
The LightRAG Server supports two operational modes:
lightrag-server
lightrag-gunicorn --workers 4
When starting LightRAG, the current working directory must contain the .env configuration file. It is intentionally designed that the .env file must be placed in the startup directory. The purpose of this is to allow users to launch multiple LightRAG instances simultaneously and configure different .env files for different instances. After modifying the .env file, you need to reopen the terminal for the new settings to take effect. This is because each time LightRAG Server starts, it loads the environment variables from the .env file into the system environment variables, and system environment variables have higher precedence.
During startup, configurations in the .env file can be overridden by command-line parameters. Common command-line parameters include:
--host: Server listening address (default: 0.0.0.0)--port: Server listening port (default: 9621)--timeout: LLM request timeout (default: 150 seconds)--log-level: Log level (default: INFO)--working-dir: Database persistence directory (default: ./rag_storage)--input-dir: Directory for uploaded files (default: ./inputs)--workspace: Workspace name, used to logically isolate data between multiple LightRAG instances (default: empty)--api-prefix: Reverse-proxy path prefix exposed to browsers, also configurable with LIGHTRAG_API_PREFIX--rerank-binding: Rerank provider (null, cohere, jina, or aliyun)Set LIGHTRAG_API_PREFIX or --api-prefix when one host serves multiple LightRAG instances behind a reverse proxy. Either forwarding style works: the proxy may strip the site prefix before forwarding to the backend, or forward the request unchanged.
LIGHTRAG_API_PREFIX=/site01
lightrag-server --port 9621
The backend passes this value to FastAPI as root_path and injects the same runtime prefix into the WebUI. The WebUI is always mounted at /webui inside the server, so one frontend build can serve any prefix. See Single-Server Multi-Site Deployment for full Nginx, Docker, and Kubernetes examples.
WHITELIST_PATHSis written without the prefix. Its entries are internal route paths, exactly as the routes are declared. The mount prefix is removed before matching, in both forwarding styles, so withLIGHTRAG_API_PREFIX=/site01the shipped defaultWHITELIST_PATHS=/health,/api/*is already correct and exempts/site01/healthas the browser sees it. Writing the browser-visible form (WHITELIST_PATHS=/site01/health) matches nothing and makes those paths require authentication.
Using Docker Compose is the most convenient way to deploy and run the LightRAG Server.
docker-compose.yml file from the LightRAG repository into your project directory..env file: Duplicate the sample file env.exampleto create a customized .env file, and configure the LLM and embedding parameters according to your specific requirements.docker compose up
# If you want the program to run in the background after startup, add the -d parameter at the end of the command.
You can get the official docker compose file from here: docker-compose.yml. For historical versions of LightRAG docker images, visit this link: LightRAG Docker Images. For more details about docker deployment, please refer to DockerDeployment.md.
If you are new to LightRAG, start with the smallest working configuration and add capabilities only after the previous step is healthy:
The full env.example file remains the complete configuration reference and is used by the make env-* setup wizard. The snippets below intentionally show only the values that matter for each step.
Use this path when you want the WebUI and API running first, with no external database, parser service, or local model service. Create .env next to docker-compose.yml with a minimal OpenAI-compatible configuration:
###########################
### Server Configuration
###########################
PORT=9621
WEBUI_TITLE='My First LightRAG KB'
WEBUI_DESCRIPTION='Simple and Fast Graph Based RAG System'
OLLAMA_EMULATING_MODEL_TAG=latest
########################################
### Document processing configuration
########################################
SUMMARY_LANGUAGE=English
ENTITY_EXTRACTION_USE_JSON=true
LIGHTRAG_PARSER=*:native-teP,*:legacy-R
VLM_PROCESS_ENABLE=false
###########################################################################
### LLM Configuration
###########################################################################
LLM_BINDING=openai
LLM_BINDING_HOST=https://api.openai.com/v1
LLM_BINDING_API_KEY=your_api_key
LLM_MODEL=gpt-5-mini
KEYWORD_LLM_MODEL=gpt-5-nano
QUERY_LLM_MODEL=gpt-5
#######################################################################################
### Embedding Configuration (do not change after the first file is processed)
#######################################################################################
EMBEDDING_BINDING=openai
EMBEDDING_BINDING_HOST=https://api.openai.com/v1
EMBEDDING_BINDING_API_KEY=your_api_key
EMBEDDING_MODEL=text-embedding-3-large
EMBEDDING_DIM=3072
EMBEDDING_TOKEN_LIMIT=8192
EMBEDDING_SEND_DIM=false
EMBEDDING_USE_BASE64=true
# Overlap (in tokens) the embedding hard fallback borrows from the previous
# chunk's tail when a chunk still exceeds EMBEDDING_TOKEN_LIMIT after
# chunking. Independent of CHUNK_OVERLAP_SIZE. Default 100; 0 disables it.
# EMBEDDING_CHUNK_OVERLAP_TOKEN_SIZE=100
############################
### Data storage selection
############################
LIGHTRAG_KV_STORAGE=JsonKVStorage
LIGHTRAG_DOC_STATUS_STORAGE=JsonDocStatusStorage
LIGHTRAG_GRAPH_STORAGE=NetworkXStorage
LIGHTRAG_VECTOR_STORAGE=NanoVectorDBStorage
Replace the model IDs with models available in your provider account when needed. Start the service and verify it before uploading documents:
docker compose up -d
curl http://localhost:9621/health
Then open the WebUI at http://localhost:9621/webui, upload a small text or DOCX file, wait for indexing to finish, and run a hybrid or mix query.
Reranking is a query-time improvement. Enabling, disabling, or changing the reranker usually does not require re-indexing existing documents.
For Cohere's official hosted rerank service:
RERANK_BINDING=cohere
RERANK_MODEL=rerank-v3.5
RERANK_BINDING_HOST=https://api.cohere.com/v2/rerank
RERANK_BINDING_API_KEY=your_cohere_api_key
For a local vLLM reranker that exposes a Cohere-compatible API:
RERANK_BINDING=cohere
RERANK_MODEL=BAAI/bge-reranker-v2-m3
RERANK_BINDING_HOST=http://localhost:8000/rerank
RERANK_BINDING_API_KEY=your_rerank_api_key_here
If LightRAG itself runs inside Docker and the reranker runs on the host, use a host-reachable address such as host.docker.internal instead of localhost. If the setup wizard creates the vLLM service, it injects the internal Compose service URL into docker-compose.final.yml for you.
Use this after the basic document flow works. The MinerU official API avoids running a local parser service, but MINERU_API_TOKEN must be configured before the LightRAG server starts. The VLM role must use a provider/model that supports image input.
LIGHTRAG_PARSER=*:native-iteP,*:mineru-iteP,*:legacy-R
VLM_PROCESS_ENABLE=true
VLM_LLM_MODEL=gpt-5-mini
MINERU_API_MODE=official
MINERU_API_TOKEN=your_mineru_api_token
MINERU_OFFICIAL_ENDPOINT=https://mineru.net
MINERU_MODEL_VERSION=vlm
MINERU_IS_OCR=false
This routing uses the built-in native parser for supported DOCX files, MinerU for other MinerU-supported files such as PDFs and images, and legacy as the fallback. The i, t, and e options enable VLM analysis for image, table, and equation sidecars when the parser produces them.
For official mode, Docker does not need a host-loopback MinerU endpoint. The container only needs outbound network access to MINERU_OFFICIAL_ENDPOINT.
For a local GPU-backed deployment, let the wizard generate .env and docker-compose.final.yml instead of hand-writing every service block:
make env-base
Recommended answers:
yes to Run embedding model locally via Docker (vLLM)?.cuda for the embedding device.yes to Run rerank service locally via Docker?, and choose cuda for the rerank device.Then configure storage:
make env-storage
Recommended storage choices:
LIGHTRAG_KV_STORAGE=PGKVStorageLIGHTRAG_DOC_STATUS_STORAGE=PGDocStatusStorageLIGHTRAG_VECTOR_STORAGE=MilvusVectorDBStorageLIGHTRAG_GRAPH_STORAGE=MemgraphStorageyes to run PostgreSQL, Milvus, and Memgraph locally via Docker.cuda for Milvus if your host has NVIDIA GPU support and the NVIDIA Container Toolkit is installed.Finally configure server-facing settings and validate the result:
make env-server
make env-validate
make env-security-check
docker compose -f docker-compose.final.yml up -d
Before exposing this deployment, configure authentication, API keys, and SSL in make env-server. The generated .env stays host-usable; container-only service names and Docker-specific overrides are written into docker-compose.final.yml.
Important rules before processing production data:
LIGHTRAG_PARSER affects only newly uploaded files. Delete and upload an existing document again if you want it processed by a different parser route.When using Nginx as a reverse proxy in front of LightRAG Server, you need to configure client_max_body_size for the /documents/upload endpoint to handle large file uploads. Without this configuration, Nginx will reject files larger than 1MB (the default limit) with a 413 Request Entity Too Large error before the request reaches LightRAG.
Recommended Configuration:
server {
listen 80;
server_name your-domain.com;
# Global default: 8MB for LLM queries with long context
client_max_body_size 8M;
# Upload endpoint: 100MB for large file uploads
location /documents/upload {
client_max_body_size 100M;
proxy_pass http://localhost:9621;
proxy_set_header Host $host;
proxy_set_header X-Real-IP $remote_addr;
proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
proxy_set_header X-Forwarded-Proto $scheme;
# Increase timeouts for large file uploads
proxy_read_timeout 300s;
proxy_send_timeout 300s;
}
# Streaming endpoints: LLM response streaming
location ~ ^/(query/stream|api/chat|api/generate) {
gzip off; # Disable compression for streaming responses
proxy_pass http://localhost:9621;
proxy_set_header Host $host;
proxy_set_header X-Real-IP $remote_addr;
proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
proxy_set_header X-Forwarded-Proto $scheme;
# Long timeout for LLM generation
proxy_read_timeout 300s;
}
# Other endpoints
location / {
proxy_pass http://localhost:9621;
proxy_set_header Host $host;
proxy_set_header X-Real-IP $remote_addr;
proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
proxy_set_header X-Forwarded-Proto $scheme;
}
}
Key Points:
MAX_UPLOAD_SIZE in your .env file. The default MAX_UPLOAD_SIZE is 100MB.gzip off) for streaming endpoints to ensure real-time response delivery. LightRAG automatically sets X-Accel-Buffering: no header to disable response buffering.proxy_read_timeout and proxy_send_timeout accordingly.Content-Length header firstenv.example):
MAX_REQUEST_BODY_BYTES bounds the raw body of every route, counted as it streams through ASGI. Unlike MAX_UPLOAD_SIZE (which bounds one uploaded file after multipart parsing), it also stops a body that understates or omits its Content-Length, answering 413 before the whole body is read. It is layered, because routes differ by orders of magnitude in what they legitimately carry:
| Route | Ceiling |
|---|---|
ordinary routes (/query, /api/chat, ...) | MAX_REQUEST_BODY_BYTES, default 1 MiB |
/documents/text, /documents/texts | 50 MiB, built in, when MAX_REQUEST_BODY_BYTES is not set |
/documents/upload | MAX_UPLOAD_SIZE + 1 MiB of multipart overhead |
Setting MAX_REQUEST_BODY_BYTES to any positive value makes it govern every non-upload route, ingestion included — including when that value happens to equal the 1 MiB default, which is the behaviour this knob had before the tiers existed. Setting it to 0 turns off every ceiling, including the derived upload one, and the server warns at startup.
Input field ceilings apply to the model-facing fields of /query*, /api/chat and /api/generate: 64 KiB per query or prompt, 32 KiB per message, 128 KiB of model-facing text per request, 128 messages, and upper bounds on top_k / chunk_top_k (1000) and the max_*_tokens budgets (1,000,000). These are fixed rather than configurable — a limit that keeps an unauthenticated caller from choosing how much CPU the server spends is worth nothing if it can be misconfigured away. /query* answers 422 for an over-limit field (FastAPI's own validation response); /api/* answers 413.
MAX_TEXTS_PER_REQUEST bounds how many texts one /documents/texts request may carry, answering 413 before any per-text storage lookup. It bounds the fan-out of a single request, so — unlike the capacity limit below — it is not a "retry later" condition: an oversized batch never fits and must be split.
MAX_PENDING_DOCUMENTS bounds how many documents may be active (PENDING/PARSING/ANALYZING/PROCESSING) or reserved by an in-flight request. Over capacity the server answers 429 with a Retry-After header and a detail naming the current count, the requested count and the capacity — refused before the body is transferred. /documents/scan and manual retries exceed the cap on purpose; the documents they create make ordinary uploads wait.
Official LightRAG Docker images are fully compatible with offline or air-gapped environments. If you want to build up your own offline environment, please refer to Offline Deployment Guide.
There are two ways to start multiple LightRAG instances. The first way is to configure a completely independent working environment for each instance. This requires creating a separate working directory for each instance and placing a dedicated .env configuration file in that directory. The server listening ports in the configuration files of different instances cannot be the same. Then, you can start the service by running lightrag-server in the working directory.
The second way is for all instances to share the same set of .env configuration files, and then use command-line arguments to specify different server listening ports and workspaces for each instance. You can start multiple LightRAG instances in the same working directory with different command-line arguments. For example:
# Start instance 1
lightrag-server --port 9621 --workspace space1
# Start instance 2
lightrag-server --port 9622 --workspace space2
The purpose of a workspace is to achieve data isolation between different instances. Therefore, the workspace parameter must be different for different instances; otherwise, it will lead to data confusion and corruption.
When launching multiple LightRAG instances via Docker Compose, simply specify unique WORKSPACE and PORT environment variables for each container within your docker-compose.yml. Even if all instances share a common .env file, the container-specific environment variables defined in Compose will take precedence, ensuring independent configurations for each instance.
Configuring an independent working directory and a dedicated .env configuration file for each instance can generally ensure that locally persisted files in the in-memory database are saved in their respective working directories, achieving data isolation. By default, LightRAG uses all in-memory databases, and this method of data isolation is sufficient. However, if you are using an external database, and different instances access the same database instance, you need to use workspaces to achieve data isolation; otherwise, the data of different instances will conflict and be destroyed.
The command-line workspace argument and the WORKSPACE environment variable in the .env file can both be used to specify the workspace name for the current instance, with the command-line argument having higher priority. Here is how workspaces are implemented for different types of storage:
JsonKVStorage, JsonDocStatusStorage, NetworkXStorage, NanoVectorDBStorage, FaissVectorDBStorage.RedisKVStorage, RedisDocStatusStorage, MilvusVectorDBStorage, MongoKVStorage, MongoDocStatusStorage, MongoVectorDBStorage, MongoGraphStorage, PGGraphStorage.QdrantVectorDBStorage uses shared collections with payload filtering for unlimited workspace scalability.workspace field to the tables for logical data separation: PGKVStorage, PGVectorStorage, PGDocStatusStorage.Neo4JStorage, MemgraphStorageOpenSearchKVStorage, OpenSearchDocStatusStorage, OpenSearchGraphStorage, OpenSearchVectorDBStorageTo maintain compatibility with legacy data, the default workspace for PostgreSQL is default and for Neo4j is base when no workspace is configured. For all external storages, the system provides dedicated workspace environment variables to override the common WORKSPACE environment variable configuration. These storage-specific workspace environment variables are: REDIS_WORKSPACE, MILVUS_WORKSPACE, QDRANT_WORKSPACE, MONGODB_WORKSPACE, POSTGRES_WORKSPACE, NEO4J_WORKSPACE, MEMGRAPH_WORKSPACE, OPENSEARCH_WORKSPACE.
The LightRAG Server can operate in the Gunicorn + Uvicorn preload mode. Gunicorn's multiple worker (multiprocess) capability prevents document indexing tasks from blocking RAG queries. CPU-heavy document extraction tools should be deployed as external services so they do not block the API process.
Though LightRAG Server uses one worker to process the document indexing pipeline, with the async task support of Uvicorn, multiple files can be processed in parallel. The bottleneck of document indexing speed mainly lies with the LLM. If your LLM supports high concurrency, you can accelerate document indexing by increasing the concurrency level of the LLM. Below are several environment variables related to concurrent processing, along with their default values:
### Number of worker processes, not greater than (2 x number_of_cores) + 1
WORKERS=2
### Number of parallel files to process in one batch
MAX_PARALLEL_INSERT=3
### Max concurrent requests to the LLM (MAX_ASYNC is still accepted as a deprecated alias)
MAX_ASYNC_LLM=4
On macOS, Gunicorn multi-worker mode also requires the Objective-C fork-safety override to be present before the Python process starts. Do not rely on .env for this variable; .env is loaded after Python startup and is too late for the Objective-C runtime:
export OBJC_DISABLE_INITIALIZE_FORK_SAFETY=YES
lightrag-gunicorn --workers 2
Create your service file lightrag.service from the sample file: lightrag.service.example. Modify the start options the service file:
# Set environment to your Python virtual environment
Environment="PATH=/home/netman/lightrag-xyj/venv/bin"
WorkingDirectory=/home/netman/lightrag-xyj
# ExecStart=/home/netman/lightrag-xyj/venv/bin/lightrag-server
ExecStart=/home/netman/lightrag-xyj/venv/bin/lightrag-gunicorn
The ExecStart command must be either
lightrag-gunicornorlightrag-server; no wrapper scripts are allowed. This is because service termination requires the main process to be one of these two executables.
Install LightRAG service. If your system is Ubuntu, the following commands will work:
sudo cp lightrag.service /etc/systemd/system/
sudo systemctl daemon-reload
sudo systemctl start lightrag.service
sudo systemctl status lightrag.service
sudo systemctl enable lightrag.service
We provide Ollama-compatible interfaces for LightRAG, aiming to emulate LightRAG as an Ollama chat model. This allows AI chat frontends supporting Ollama, such as Open WebUI, to access LightRAG easily.
After starting the lightrag-server, you can add an Ollama-type connection in the Open WebUI admin panel. And then a model named lightrag:latest will appear in Open WebUI's model management interface. Users can then send queries to LightRAG through the chat interface. You should install LightRAG as a service for this use case.
Open WebUI uses an LLM to do the session title and session keyword generation task. So the Ollama chat completion API detects and forwards OpenWebUI session-related requests directly to the underlying LLM. Screenshot from Open WebUI:
The default query mode is mix if you send a message (query) from the Ollama interface of LightRAG. You can select query mode by sending a message with a query prefix.
A query prefix in the query string can determine which LightRAG query mode is used to generate the response for the query. The supported prefixes include:
/local
/global
/hybrid
/naive
/mix
/bypass
/context
/localcontext
/globalcontext
/hybridcontext
/naivecontext
/mixcontext
For example, the chat message /hybrid What's LightRAG? will trigger a hybrid mode query for LightRAG. A chat message without a query prefix will trigger a mix mode query by default.
/bypass is not a LightRAG query mode; it will tell the API Server to pass the query directly to the underlying LLM, including the chat history. So the user can use the LLM to answer questions based on the chat history. If you are using Open WebUI as a front end, you can just switch the model to a normal LLM instead of using the /bypass prefix.
/context is also not a LightRAG query mode; it will tell LightRAG to return only the context information prepared for the LLM. You can check the context if it's what you want, or process the context by yourself.
When using LightRAG for content queries, avoid combining the search process with unrelated output processing, as this significantly impacts query effectiveness. User prompt is specifically designed to address this issue — it does not participate in the RAG retrieval phase, but rather guides the LLM on how to process the retrieved results after the query is completed. We can append square brackets to the query prefix to provide the LLM with the user prompt:
/[Use mermaid format for diagrams] Please draw a character relationship diagram for Scrooge
/mix[Use mermaid format for diagrams] Please draw a character relationship diagram for Scrooge
By default, the LightRAG Server can be accessed without any authentication. We can configure the server with an API Key or account credentials to secure it.
LIGHTRAG_API_KEY=your-secure-api-key-here
WHITELIST_PATHS=/health,/api/*
Health check and Ollama emulation endpoints are excluded from API Key check by default. For security reasons, remove
/api/*fromWHITELIST_PATHSif the Ollama service is not required./healthstays whitelisted as a liveness probe but only returns its full configuration to authenticated callers — unauthenticated requests get liveness signals only.Entries are internal route paths, never prefixed. A
/*suffix matches on path-segment boundaries, so/api/*covers/apiand everything under/api/and nothing else. IfLIGHTRAG_API_PREFIXis set, do not include it here: the prefix is removed before matching, soWHITELIST_PATHS=/healthexempts/site01/healthandWHITELIST_PATHS=/site01/healthexempts nothing. See Path Prefix and Multi-Site WebUI.
The API key is passed using the request header X-API-Key. Below is an example of accessing the LightRAG Server via API:
curl -X 'POST' \
'http://localhost:9621/documents/scan' \
-H 'accept: application/json' \
-H 'X-API-Key: your-secure-api-key-here-123' \
-d ''
LightRAG API Server implements JWT-based authentication using the HS256 algorithm. To enable secure access control, the following environment variables are required:
# For jwt auth
AUTH_ACCOUNTS='admin:{bcrypt}$2b$12$replace-with-generated-hash,user1:pass456'
TOKEN_SECRET='your-key'
TOKEN_EXPIRE_HOURS=4
Passwords without a prefix are treated as plaintext. To store a bcrypt password, prefix the generated hash with {bcrypt}. The easiest way to generate a value that can be pasted directly into AUTH_ACCOUNTS is:
lightrag-hash-password --username admin
The command prompts for the password and prints an admin:{bcrypt}... entry ready to paste into .env.
Currently, only the configuration of an administrator account and password is supported. A comprehensive account system is yet to be developed and implemented.
If Account credentials are not configured, the Web UI will access the system as a Guest. Therefore, even if only an API Key is configured, all APIs can still be accessed through the Guest account, which remains insecure. Hence, to safeguard the API, it is necessary to configure both authentication methods simultaneously.
Although the server can be configured with both an API key and account credentials, a single request should send either
X-API-KeyorAuthorization: Bearer <token>— not both. When both headers are present, theAuthorizationtoken is validated first; if it is invalid or expired the request is rejected with401 Invalid tokeneven when a validX-API-Keyis also supplied.
Azure OpenAI API can be created using the following commands in Azure CLI (you need to install Azure CLI first from https://docs.microsoft.com/en-us/cli/azure/install-azure-cli):
# Change the resource group name, location, and OpenAI resource name as needed
RESOURCE_GROUP_NAME=LightRAG
LOCATION=swedencentral
RESOURCE_NAME=LightRAG-OpenAI
az login
az group create --name $RESOURCE_GROUP_NAME --location $LOCATION
az cognitiveservices account create --name $RESOURCE_NAME --resource-group $RESOURCE_GROUP_NAME --kind OpenAI --sku S0 --location swedencentral
az cognitiveservices account deployment create --resource-group $RESOURCE_GROUP_NAME --model-format OpenAI --name $RESOURCE_NAME --deployment-name gpt-4o --model-name gpt-4o --model-version "2024-08-06" --sku-capacity 100 --sku-name "Standard"
az cognitiveservices account deployment create --resource-group $RESOURCE_GROUP_NAME --model-format OpenAI --name $RESOURCE_NAME --deployment-name text-embedding-3-large --model-name text-embedding-3-large --model-version "1" --sku-capacity 80 --sku-name "Standard"
az cognitiveservices account show --name $RESOURCE_NAME --resource-group $RESOURCE_GROUP_NAME --query "properties.endpoint"
az cognitiveservices account keys list --name $RESOURCE_NAME -g $RESOURCE_GROUP_NAME
The output of the last command will give you the endpoint and the key for the OpenAI API. You can use these values to set the environment variables in the .env file.
# Azure OpenAI Configuration in .env:
LLM_BINDING=azure_openai
LLM_BINDING_HOST=your-azure-endpoint
LLM_MODEL=your-model-deployment-name
LLM_BINDING_API_KEY=your-azure-api-key
### API version is optional, defaults to latest version
AZURE_OPENAI_API_VERSION=2024-08-01-preview
### If using Azure OpenAI for embeddings
EMBEDDING_BINDING=azure_openai
EMBEDDING_MODEL=your-embedding-deployment-name
The API Server can be configured in two ways (highest priority first):
Most of the configurations come with default settings; check out the details in the sample file: .env.example. Storage configuration should also be set through environment variables or the .env file.
LightRAG supports binding to various LLM backends:
LightRAG supports binding to various Embedding backends:
Use environment variables LLM_BINDING or CLI argument --llm-binding to select the LLM backend type. Use environment variables EMBEDDING_BINDING or CLI argument --embedding-binding to select the Embedding backend type.
Bedrock ignores LLM_BINDING_API_KEY and EMBEDDING_BINDING_API_KEY. Use SigV4 credentials through the AWS credential chain, or set the process-level AWS_BEARER_TOKEN_BEDROCK environment variable before startup for Bedrock API key / bearer-token auth:
LLM_BINDING=bedrock
LLM_BINDING_HOST=DEFAULT_BEDROCK_ENDPOINT
LLM_MODEL=us.amazon.nova-lite-v1:0
AWS_REGION=us-west-2
# Use the AWS credential chain, or set AWS_ACCESS_KEY_ID/AWS_SECRET_ACCESS_KEY,
# or set AWS_BEARER_TOKEN_BEDROCK before starting the server.
Asymmetric embedding is explicit opt-in. Set EMBEDDING_ASYMMETRIC=true only when the selected embedding backend supports either provider task parameters or task prefixes. See Asymmetric Embedding Configuration before changing these settings, because existing data must be cleared and files re-indexed after any change.
For LLM and embedding configuration examples, please refer to the env.example file in the project's root directory. To view the complete list of configurable options for OpenAI and Ollama-compatible LLM interfaces, use the following commands:
lightrag-server --llm-binding openai --help
lightrag-server --llm-binding ollama --help
lightrag-server --llm-binding gemini --help
lightrag-server --embedding-binding ollama --help
lightrag-server --embedding-binding gemini --help
Please use OpenAI-compatible method to access LLMs deployed by OpenRouter or vLLM/SGLang. You can pass additional parameters to OpenRouter or vLLM/SGLang through the
OPENAI_LLM_EXTRA_BODYenvironment variable to disable reasoning mode or achieve other personalized controls.
Set the max_tokens to prevent excessively long or endless output loop during the entity relationship extraction phase for Large Language Model (LLM) responses. The purpose of setting max_tokens parameter is to truncate LLM output before timeouts occur, thereby preventing document extraction failures. This addresses issues where certain text blocks (e.g., tables or citations) containing numerous entities and relationships can lead to overly long or even endless loop outputs from LLMs. This setting is particularly crucial for locally deployed, smaller-parameter models. Max tokens value can be calculated by this formula: LLM_TIMEOUT * llm_output_tokens/second (i.e. 240s * 50 tokens/s = 12000, max_tokens should smaller than 12000)
# For vLLM/SGLang doployed models, or most of OpenAI compatible API provider
OPENAI_LLM_MAX_TOKENS=9000
# For Ollama Deployed Modeles
OLLAMA_LLM_NUM_PREDICT=9000
# For OpenAI o1-mini or newer modles
OPENAI_LLM_MAX_COMPLETION_TOKENS=9000
The server can use different models for different stages without changing client APIs. Four roles are supported:
| Role | Purpose |
|---|---|
EXTRACT | Entity/relation extraction and merge summaries |
KEYWORD | Query keyword generation before retrieval |
QUERY | Final answers, bypass queries, and Ollama-compatible chat responses |
VLM | Multimodal analysis for images, tables, equations, and similar sidecar items |
If a role is not configured, it inherits the base LLM_* settings. Minimal same-provider example:
LLM_BINDING=openai
LLM_MODEL=gpt-5-mini
LLM_BINDING_HOST=https://api.openai.com/v1
LLM_BINDING_API_KEY=your_api_key
EXTRACT_LLM_MODEL=gpt-5-mini
KEYWORD_LLM_MODEL=gpt-5-nano
QUERY_LLM_MODEL=gpt-5
VLM_LLM_MODEL=gpt-5-mini
Recommended models by role:
EXTRACT: Entity-relation extraction runs on every chunk, so a fast, cost-effective mainstream model is enough — a non-thinking model (reasoning/thinking mode disabled) is strongly recommended. E.g. GPT-5.6-luna, Claude Haiku, or Gemini-mini (hosted), or DeepSeek-V4-lite / Kimi in China. For local deployment, Qwen3-30B-A3B-Instruct is a reasonable minimum.QUERY: Writes the final answer from long, noisy context, so choose a stronger model than EXTRACT to maximize answer quality; a thinking-capable model is fine here.KEYWORD: A lightweight, latency-sensitive step that must use a non-thinking model to keep query latency low; a fast model comparable to EXTRACT is enough.VLM: Any mainstream multimodal model with image-input support works; for local deployment, consider Qwen3.6-35B-A3B.BAAI/bge-m3 for embeddings and BAAI/bge-reranker-v2-m3 for reranking.Within an acceptable latency and cost budget, prefer the highest-scoring model available (per public benchmarks/leaderboards).
For cross-provider rules, provider-specific options such as QUERY_OPENAI_LLM_REASONING_EFFORT, role-level Bedrock SigV4 credentials, and queue behavior, see Role-Specific LLM/VLM Configuration Guide.
The parser can produce sidecars for drawings/images, tables, and equations. Analysis of a modality requires the document's process_options to contain the matching flag — i for images, t for tables, e for equations — and the corresponding sidecar to exist.
VLM_PROCESS_ENABLE gates images only. Tables and equations are analyzed by the EXTRACT role and run regardless of this switch, so *:native-teP works without any VLM configured. With i enabled and the VLM unavailable, an image that survives the pre-filters (file present, raster format, both sides at least VLM_MIN_IMAGE_PIXEL) fails the document rather than being skipped — it lands in FAILED with error_msg "VLM analysis required but VLM role is not available".
Current vision-capable providers are openai, azure_openai, gemini, bedrock, ollama, and anthropic; lollms is rejected for VLM use. Typical configuration:
VLM_PROCESS_ENABLE=true
VLM_LLM_BINDING=openai
VLM_LLM_MODEL=gpt-4o
VLM_LLM_BINDING_HOST=https://api.openai.com/v1
VLM_LLM_BINDING_API_KEY=your_vlm_api_key
VLM_MAX_IMAGE_BYTES=5242880
SURROUNDING_LEADING_MAX_TOKENS=2000
SURROUNDING_TRAILING_MAX_TOKENS=2000
The surrounding-context budgets control how much nearby text is included in VLM and extraction prompts for a multimodal item. Parser and per-file option examples are in Document and Chunk Processing.
Entity extraction is controlled by the base or EXTRACT role LLM. Important server-side options:
ENTITY_EXTRACTION_USE_JSON: request JSON-structured extraction output. In v1.5 this is recommended for reliability, but it can increase latency.ENTITY_TYPE_PROMPT_FILE: file-name-only YAML profile for entity type guidance and examples. The file is loaded from PROMPT_DIR/entity_type; do not pass an absolute path here.MAX_EXTRACT_INPUT_TOKENS: maximum token budget for one extraction input context.MAX_EXTRACTION_RECORDS: per-response cap for total entity and relationship records.MAX_EXTRACTION_ENTITIES: per-response cap for entity records.Example:
ENTITY_EXTRACTION_USE_JSON=true
ENTITY_TYPE_PROMPT_FILE=entity_type_prompt.yml
PROMPT_DIR=/opt/lightrag/prompts
MAX_EXTRACT_INPUT_TOKENS=20480
MAX_EXTRACTION_RECORDS=100
MAX_EXTRACTION_ENTITIES=40
If an old .env still contains ENTITY_TYPES, remove it before startup. The server fails fast because this variable has been replaced by prompt profiles.
LightRAG uses 4 types of storage for different purposes:
Each storage type offers multiple implementations. By default, LightRAG Server uses in-memory databases with data persisted to the WORKING_DIR directory. This is suitable for quickly evaluating the project but is not recommended for production. The implementations currently available for each storage type are listed below:
| Storage Type | Available Implementations (Default First) |
|---|---|
| KV_STORAGE | JsonKVStorage, RedisKVStorage, PGKVStorage, MongoKVStorage, OpenSearchKVStorage |
| VECTOR_STORAGE | NanoVectorDBStorage, MilvusVectorDBStorage, PGVectorStorage, FaissVectorDBStorage, QdrantVectorDBStorage, MongoVectorDBStorage, OpenSearchVectorDBStorage |
| GRAPH_STORAGE | NetworkXStorage, Neo4JStorage, PGTableGraphStorage, PGGraphStorage, MongoGraphStorage, MemgraphStorage, OpenSearchGraphStorage |
| DOC_STATUS_STORAGE | JsonDocStatusStorage, RedisDocStatusStorage, PGDocStatusStorage, MongoDocStatusStorage, OpenSearchDocStatusStorage |
For production deployments, PostgreSQL, MongoDB, or OpenSearch can provide all four storage types through a single backend. You can also select a specialized database for each storage type, such as Milvus or Qdrant for vector storage and Neo4j or Memgraph for graph storage.
PostgreSQL Graph Storage — prefer PGTableGraphStorage: For new PostgreSQL deployments, PGTableGraphStorage is the recommended GRAPH_STORAGE implementation and supersedes PGGraphStorage. It keeps the entity-relation graph in ordinary tables — JSONB properties plus B-tree indexes — instead of going through Apache AGE, which brings two practical advantages:
PGGraphStorage requires the Apache AGE extension, which most managed PostgreSQL services (Amazon RDS, Cloud SQL, Supabase, Neon) do not offer — so the graph layer frequently could not run on the same database as the other three storage types. PGTableGraphStorage runs on any stock PostgreSQL 14+ and creates the tables it needs during initialize(). For a Docker deployment this means the official pgvector/pgvector:pg18 image is sufficient; the AGE-bundled gzdaniel/postgres-for-rag:pg18-age-pgvector image is only needed by PGGraphStorage.agtype, and get_knowledge_graph uses a frontier-capped BFS bounded by max_nodes. From the measurements published with PR #3103 (PostgreSQL 18, an 8k-node / ~40k-edge graph, both backends VACUUM ANALYZEd before measuring): get_knowledge_graph p50 39 ms vs 1,099 ms (~28×), bulk graph load 3.0 s vs 434 s, mixed-workload throughput 1,431 vs 73 RPS.Both implementations read the same POSTGRES_* environment variables, but they store the graph in different places — PGTableGraphStorage in its own lightrag_graph_nodes / lightrag_graph_edges tables, PGGraphStorage inside an AGE graph. Switching an existing deployment is therefore not an in-place migration: the previously extracted graph will not be visible and the documents have to be re-indexed (the LLM cache can be carried over — see LLM Cache Migration Between Storage Types below). PGGraphStorage remains supported for deployments already running on AGE.
The environment variables required at startup for each storage implementation are listed below. Implementations not listed require no additional configuration and rely only on file persistence under WORKING_DIR.
| Storage Implementation | Required Environment Variables |
|---|---|
PGKVStorage / PGVectorStorage / PGGraphStorage / PGTableGraphStorage / PGDocStatusStorage | POSTGRES_USER, POSTGRES_PASSWORD, POSTGRES_DATABASE (plus POSTGRES_HOST and POSTGRES_PORT) |
Neo4JStorage | NEO4J_URI, NEO4J_USERNAME, NEO4J_PASSWORD |
MongoKVStorage / MongoVectorDBStorage / MongoGraphStorage / MongoDocStatusStorage | MONGO_URI, MONGO_DATABASE (MongoVectorDBStorage requires a MongoDB deployment that supports Atlas Search / Vector Search) |
RedisKVStorage / RedisDocStatusStorage | REDIS_URI |
MilvusVectorDBStorage | MILVUS_URI, MILVUS_DB_NAME |
QdrantVectorDBStorage | QDRANT_URL (QDRANT_API_KEY is optional) |
MemgraphStorage | MEMGRAPH_URI |
OpenSearchKVStorage / OpenSearchVectorDBStorage / OpenSearchGraphStorage / OpenSearchDocStatusStorage | OPENSEARCH_HOSTS |
The WORKSPACE environment variable isolates data for multiple LightRAG instances on the same backend (valid characters are a-z, A-Z, 0-9, and _). Each storage backend also provides a backend-specific override such as POSTGRES_WORKSPACE or NEO4J_WORKSPACE. These overrides are retained only for compatibility with legacy configurations; under normal circumstances, use WORKSPACE consistently.
The table above lists only the connection parameters required at startup. Each storage implementation also provides many optional tuning environment variables, including connection pool sizes, SSL settings, sharding thresholds for batch writes and deletions, and vector index parameters. For the complete list and default values, see the repository's root-level env.example, where the variables are grouped by storage backend and include detailed comments.
Milvus Index Configuration: LightRAG now supports configurable index types for Milvus vector storage (AUTOINDEX, HNSW, HNSW_SQ, IVF_FLAT, etc.) through environment variables. HNSW_SQ requires Milvus 2.6.8+ and provides significant memory savings. For the complete configuration options, see MilvusConfigurationGuide.md.
You can select the storage implementation by configuring environment variables. For instance, prior to the initial launch of the API server, you can set the following environment variable to specify your desired storage implementation:
LIGHTRAG_KV_STORAGE=PGKVStorage
LIGHTRAG_VECTOR_STORAGE=PGVectorStorage
LIGHTRAG_GRAPH_STORAGE=PGTableGraphStorage
LIGHTRAG_DOC_STATUS_STORAGE=PGDocStatusStorage
You cannot change storage implementation selection after adding documents to LightRAG. Data migration from one storage implementation to another is not supported yet. For further information, please read the sample .env.example file.
The dev-lancedb development branch provides community-contributed LanceDB storage implementations for all four storage types: key-value (KV), vector, graph, and document status. The dev-nebula-graph development branch provides a community-contributed Nebula graph storage implementation. Developers who need these storage options are welcome to try them and help improve them.
When switching the storage implementation in LightRAG, the LLM cache can be migrated from the existing storage to the new one. Subsequently, when re-uploading files to the new storage, the pre-existing LLM cache will significantly accelerate file processing. For detailed instructions on using the LLM cache migration tool, please refer to README_MIGRATE_LLM_CACHE.md
| Parameter | Default | Description |
|---|---|---|
--host | 0.0.0.0 | Server host |
--port | 9621 | Server port |
--working-dir | ./rag_storage | Working directory for RAG storage |
--input-dir | ./inputs | Directory containing uploaded/input documents |
--timeout | 150 | Gunicorn worker timeout and fallback request timeout |
--max-async | 4 | Maximum concurrent LLM operations |
--log-level | INFO | Logging level (DEBUG, INFO, WARNING, ERROR, CRITICAL) |
--verbose | False | Verbose debug output, effective with debug logging |
--key | None | API key for authentication |
--ssl | False | Enable HTTPS |
--ssl-certfile | None | Path to SSL certificate file, required if --ssl is enabled |
--ssl-keyfile | None | Path to SSL private key file, required if --ssl is enabled |
--workspace | "" | Default workspace for storage isolation |
--api-prefix | "" | Reverse-proxy path prefix, also configurable with LIGHTRAG_API_PREFIX |
--workers | 1 | Gunicorn worker count |
--llm-binding | ollama | LLM binding type (lollms, ollama, openai, openai-ollama, azure_openai, bedrock, gemini) |
--embedding-binding | ollama | Embedding binding type (lollms, ollama, openai, azure_openai, bedrock, jina, gemini, voyageai) |
--rerank-binding | null | Rerank binding type (null, cohere, jina, aliyun) |
Reranking query-recalled chunks can significantly enhance retrieval quality by re-ordering documents based on an optimized relevance scoring model. LightRAG currently supports the following rerank providers:
v2/rerank endpoint. As vLLM provides a Cohere-compatible reranker API, all reranker models deployed via vLLM are also supported.The rerank provider is configured via the .env file. Below is an example configuration for a rerank model deployed locally using vLLM:
RERANK_BINDING=cohere
RERANK_MODEL=BAAI/bge-reranker-v2-m3
RERANK_BINDING_HOST=http://localhost:8000/rerank
RERANK_BINDING_API_KEY=your_rerank_api_key_here
Here is an example configuration for utilizing the Reranker service provided by Aliyun (gte-rerank-* and qwen3-vl-rerank, which use the nested input/parameters payload format):
RERANK_BINDING=aliyun
RERANK_MODEL=gte-rerank-v2
RERANK_BINDING_HOST=https://dashscope.aliyuncs.com/api/v1/services/rerank/text-rerank/text-rerank
RERANK_BINDING_API_KEY=your_rerank_api_key_here
Aliyun
qwen3-rerankseries: Unlikegte-rerank-*andqwen3-vl-rerank, theqwen3-rerankmodels use a flat, Cohere-style payload ({"model", "query", "documents", "top_n", ...}), return top-levelresults, and are served from a different, Cohere-compatible endpoint —/compatible-api/v1/reranks, not the.../text-rerank/text-rerankpath used above. Because the format is identical to standard Cohere, configure them withRERANK_BINDING=cohere(notaliyun); no dedicated binding is needed. Replace{WorkspaceId}and the region with your own (see the Aliyun Text Rerank API docs):
RERANK_BINDING=cohere
RERANK_MODEL=qwen3-rerank
RERANK_BINDING_HOST=https://{WorkspaceId}.cn-beijing.maas.aliyuncs.com/compatible-api/v1/reranks
RERANK_BINDING_API_KEY=your_rerank_api_key_here
Reranker calls have their own concurrency and timeout controls:
MAX_ASYNC_RERANK=4
RERANK_TIMEOUT=30
MAX_ASYNC_RERANK falls back to MAX_ASYNC_LLM when unset (MAX_ASYNC is still accepted as a deprecated alias). RERANK_TIMEOUT has an independent default because reranker requests are usually shorter than LLM generation requests. For comprehensive reranker configuration examples, including Cohere-compatible chunking options and Jina/Aliyun endpoints, refer to the env.example file.
Reranking can be enabled or disabled on a per-query basis.
The /query and /query/stream API endpoints include an enable_rerank parameter, which is set to true by default, controlling whether reranking is active for the current query. To change the default value of the enable_rerank parameter to false, set the following environment variable:
RERANK_BY_DEFAULT=False
By default, the /query and /query/stream endpoints return references with only reference_id and file_path. For evaluation, debugging, or citation purposes, you can request the actual retrieved chunk content to be included in references.
The include_chunk_content parameter (default: false) controls whether the actual text content of retrieved chunks is included in the response references. This is particularly useful for:
Important: The content field is an array of strings, where each string represents a chunk from the same file. A single file may correspond to multiple chunks, so the content is returned as a list to preserve chunk boundaries.
Example API Request:
{
"query": "What is LightRAG?",
"mode": "mix",
"include_references": true,
"include_chunk_content": true
}
Example Response (with chunk content):
{
"response": "LightRAG is a graph-based RAG system...",
"references": [
{
"reference_id": "1",
"file_path": "/documents/intro.md",
"content": [
"LightRAG is a retrieval-augmented generation system that combines knowledge graphs with vector similarity search...",
"The system uses a dual-indexing approach with both vector embeddings and graph structures for enhanced retrieval..."
]
},
{
"reference_id": "2",
"file_path": "/documents/features.md",
"content": [
"The system provides multiple query modes including local, global, hybrid, and mix modes..."
]
}
]
}
Notes:
include_references=true. Setting include_chunk_content=true without including references has no effect.content as a single concatenated string. Now it returns an array of strings to preserve individual chunk boundaries. If you need a single string, join the array elements with your preferred separator (e.g., "\n\n".join(content)).The examples below are reference snippets for tuning existing deployments. For a first run, follow Progressive Setup Recipes instead of copying the entire env.example file by hand.
### Server Configuration
# HOST=0.0.0.0
PORT=9621
WORKERS=2
# LIGHTRAG_API_PREFIX=/site01
### Settings for document indexing
ENTITY_EXTRACTION_USE_JSON=true
# ENTITY_TYPE_PROMPT_FILE=entity_type_prompt.yml
# MAX_EXTRACT_INPUT_TOKENS=20480
# MAX_EXTRACTION_RECORDS=100
# MAX_EXTRACTION_ENTITIES=40
SUMMARY_LANGUAGE=Chinese
MAX_PARALLEL_INSERT=3
LIGHTRAG_PARSER=*:native-teP,*:legacy-R
# CHUNK_R_SEPARATORS=["\n\n","\n","。","！","？","；","，"," ",""]
# CHUNK_P_SIZE=2000
### LLM Configuration (Use valid host. For local services installed with docker, you can use host.docker.internal)
TIMEOUT=150
MAX_ASYNC_LLM=4
LLM_BINDING=openai
LLM_MODEL=gpt-4o-mini
LLM_BINDING_HOST=https://api.openai.com/v1
LLM_BINDING_API_KEY=your-api-key
KEYWORD_LLM_MODEL=gpt-4o-mini
QUERY_LLM_MODEL=gpt-4o
### Optional VLM configuration for documents using i/t/e process options
VLM_PROCESS_ENABLE=false
# VLM_LLM_MODEL=gpt-4o
# VLM_MAX_IMAGE_BYTES=5242880
# SURROUNDING_LEADING_MAX_TOKENS=2000
# SURROUNDING_TRAILING_MAX_TOKENS=2000
### Optional reranker configuration
RERANK_BINDING=null
# MAX_ASYNC_RERANK=4
# RERANK_TIMEOUT=30
### Embedding Configuration (Use valid host. For local services installed with docker, you can use host.docker.internal)
# see also env.ollama-binding-options.example for fine tuning ollama
EMBEDDING_MODEL=bge-m3:latest
EMBEDDING_DIM=1024
EMBEDDING_BINDING=ollama
EMBEDDING_BINDING_HOST=http://localhost:11434
# Optional asymmetric embedding for prefix-based models:
# EMBEDDING_ASYMMETRIC=true
# EMBEDDING_QUERY_PREFIX="search_query: "
# EMBEDDING_DOCUMENT_PREFIX="search_document: "
# Use NO_PREFIX for a side that should intentionally have no prefix.
### For JWT Auth
# AUTH_ACCOUNTS='admin:{bcrypt}$2b$12$replace-with-generated-hash,user1:pass456'
# TOKEN_SECRET=your-key-for-LightRAG-API-Server-xxx
# TOKEN_EXPIRE_HOURS=48
# LIGHTRAG_API_KEY=your-secure-api-key-here-123
# WHITELIST_PATHS=/api/*
# WHITELIST_PATHS=/health,/api/*
v1.5 introduces a staged document pipeline. Files first go through a content extraction engine, optional multimodal analysis, text chunking, and then entity/relation extraction unless the file disables knowledge graph construction.
Keep v1.4-compatible behavior:
LIGHTRAG_PARSER=*:legacy-F
Recommended starting point without external parser services:
LIGHTRAG_PARSER=*:native-teP,*:legacy-R
This uses the built-in native parser for supported files, enables table/equation sidecar analysis options for those files, uses paragraph semantic chunking where possible, and falls back to legacy extraction plus recursive chunking for other files.
Full multimodal setup with the MinerU official API and a VLM:
LIGHTRAG_PARSER=*:native-iteP,*:mineru-iteP,*:legacy-R
VLM_PROCESS_ENABLE=true
VLM_LLM_MODEL=gpt-4o
MINERU_API_MODE=official
MINERU_API_TOKEN=your_mineru_api_token
MINERU_OFFICIAL_ENDPOINT=https://mineru.net
MINERU_MODEL_VERSION=vlm
MINERU_IS_OCR=false
Use DOCLING_ENDPOINT=http://localhost:5001 when routing files to docling.
LIGHTRAG_PARSER defines default extraction rules by file extension. Rules are matched left to right and can be separated by commas or semicolons:
LIGHTRAG_PARSER=pdf:mineru-R,docx:native-ietP,*:legacy-R
Supported engines:
| Engine | Use case |
|---|---|
legacy | Original extraction behavior. Good for compatibility and simple text-like files. |
native | Built-in structured parser, currently focused on .docx and LightRAG Document sidecars. |
mineru | External MinerU parser for PDFs, Office files, and images. Requires MINERU_API_MODE plus MINERU_LOCAL_ENDPOINT or MINERU_API_TOKEN. |
docling | External docling-serve parser for PDFs, Office files, Markdown/HTML, and images. Requires DOCLING_ENDPOINT. |
Filename hints override the default rule for one uploaded file:
paper.[mineru-iteP].pdf
memo.[native-R!].docx
notes.[-R].md
The /documents/upload and /documents/scan paths honor filename hints and LIGHTRAG_PARSER. The /documents/text and /documents/texts endpoints insert already-provided text and currently use fixed chunking on the server path.
Processing options are appended after the engine with a hyphen, or supplied alone in a filename hint with [-OPTIONS].
| Option | Meaning |
|---|---|
i | Run VLM analysis for image/drawing sidecars when present |
t | Run VLM analysis for table sidecars when present |
e | Run VLM analysis for equation sidecars when present |
! | Skip entity/relation extraction and graph writes; chunk vectors are still stored |
F | Fixed token chunking, the legacy chunking method |
R | Recursive character chunking with configurable separator cascade |
V | Semantic vector chunking; oversize chunks are re-split by R |
P | Paragraph semantic chunking for structured LightRAG Document content; falls back to R when structured content is unavailable |
At most one of F, R, V, and P should be selected for a file. Chunker parameters are configured with CHUNK_SIZE, CHUNK_OVERLAP_SIZE, and strategy-specific variables such as CHUNK_R_SEPARATORS, CHUNK_V_BREAKPOINT_THRESHOLD_TYPE, CHUNK_P_SIZE, and CHUNK_P_OVERLAP_SIZE. These values are read at server startup and stored as a per-document chunk_options snapshot when a document is enqueued.
The V strategy's sentence splitter is the one chunker parameter that cannot be set per request: CHUNK_V_SENTENCE_SPLIT_REGEX (or the SDK's addon_params) is the only way to change it. /documents/text and /documents/texts reject a sentence_split_regex key inside chunking.params with HTTP 422. A caller-supplied pattern is applied to that same request's text, and CPython's regex engine holds the GIL while backtracking, so a pattern such as (a+)+$ can freeze an entire worker process — see GHSA-32jh-39m7-8x84. A value already stored in a document's chunk_options snapshot is discarded at processing time as well (logged at WARNING), so a pattern persisted by an older build cannot freeze the worker after an upgrade.
The R strategy's separator cascade is bounded to 64 entries of at most 256 characters each, wherever it comes from; the built-in cascade is 9. A request body over the limit is rejected with HTTP 422. A non-HTTP configured value is converged and logged once when cached: CHUNK_R_SEPARATORS at configuration load, and a supplied or replaced addon_params['chunker'] immediately (a compatible nested in-place mutation is handled at its first enqueue). The normalized value is then reused for later documents, corrected in place so a caller-held reference to the nested recursive_character dict still applies. Direct SDK calls and per-document snapshots persisted before the bound existed retain their stored value and are converged silently at execution, so one stale value cannot warn once per document. A separators value that is neither a list/tuple nor None is not converged at all — the key is dropped, with its own warning, because bounding a bare string would silently turn it into 64 single-character separators. Converging is not the same as shortening: an entry over 256 characters is dropped, while a list over 64 entries is truncated to 64 (keeping the trailing char-level "" sentinel when present). A lone 300-character separator therefore disappears rather than matching its first 256, and the split points come from the fallback cascade — see the pipeline spec for which fallback applies where.
For the full routing syntax, supported extensions, parser cache behavior, chunker configuration, concurrency rules, and Python SDK differences, see File Processing Pipeline Specification. For the P strategy details, see Paragraph Semantic Chunking. To debug parser output before indexing a file, see Parser Debug CLI.
MAX_PARALLEL_INSERT controls how many files are processed in parallel. MAX_ASYNC_LLM (deprecated alias: MAX_ASYNC) controls concurrent LLM calls, including extraction, merging, query keyword generation, and final answer generation. Optional staged-pipeline variables such as MAX_PARALLEL_PARSE_NATIVE, MAX_PARALLEL_PARSE_MINERU, MAX_PARALLEL_PARSE_DOCLING, and MAX_PARALLEL_ANALYZE can be used for parser-heavy deployments.
Uploads and text inserts can be accepted while the processing loop is busy; the running loop is nudged to pick up the new pending work. Destructive jobs such as document clear/delete and the classification phase of /documents/scan still reject concurrent enqueues to protect storage consistency. Failed files can be reprocessed from the WebUI or by triggering /documents/scan.
All supported backends (lollms, ollama, openai / OpenAI-compatible, azure_openai, bedrock, and gemini) expose the same LightRAG REST API surface. When the API Server is running, visit:
You can test the API endpoints using the provided curl commands or through the Swagger UI interface. Make sure to:
The /health endpoint reports operational state and selected configuration, including role LLM configuration, LLM/embedding/rerank queue status, workspace/storage workspace mapping, VLM enablement, rerank enablement, and pipeline busy/scanning/destructive status. It always returns HTTP 200 so it stays usable as a liveness probe, but the configuration and operational diagnostics are returned only to authenticated callers (valid JWT or X-API-Key). Unauthenticated callers receive only liveness signals (status, auth_mode, core_version, api_version, pipeline_busy/pipeline_active, and the WebUI title/availability fields — all of which are also exposed by the unauthenticated /auth-status endpoint or are plain booleans). Provide credentials to retrieve the full payload, e.g. curl -H "X-API-Key: <key>" http://localhost:9621/health.
LightRAG implements asynchronous document indexing to enable frontend monitoring and querying of document processing progress. Upon uploading files or inserting text through designated endpoints, a unique Track ID is returned to facilitate real-time progress monitoring.
API Endpoints Supporting Track ID Generation:
/documents/upload/documents/text/documents/textsDocument Processing Status Query Endpoint:
/documents/track_status/{track_id}This endpoint provides comprehensive status information including: