feat: Finish implement workspace isolation in lightrag_server

[sillydong]

Description

In this PR #2369, it proposed a way to get workspace from request header. It focused on pipeline status and in-mem storage. For the routers, it does not cover. All requests sent to server is still in the hardcoded workspace in env. In this new PR, they are fixed.

• There will be a dict of LightRAG instances and a dict of DocumentManager instances for different workspace kept in memory.
• All routes (document/graph/ollama/query) supports get workspace from header.
• WebUI supports switch workspace by manually input and update. User can easily change workspace in webpage.

Related Issues

[Reference any related issues or tasks addressed by this pull request.]
Issue #1289
Issue #2373
Discussion #1016

Changes Made

[List the specific changes made in this pull request.]

• create create_doc_manager and create_rag in lightrag_server and pass to routers
• get workspace from request header, use create_doc_manager and create_rag to create instances for given workspace
• add input and button in webui to allow user change workspace
• send workspace with request in header LIGHTRAG-WORKSPACE

Checklist

• Changes tested locally
• Code reviewed
• Documentation updated (if necessary)
• Unit tests added (if applicable)

Additional Notes

[Add any additional notes or context for the reviewer(s).]

---

Note

Route all API requests to per-workspace LightRAG/DocumentManager instances via LIGHTRAG-WORKSPACE header and add WebUI control to switch workspaces.

• Backend (FastAPI):
  • Implement per-workspace caching with create_rag(request) and create_doc_manager(request); derive workspace from LIGHTRAG-WORKSPACE header (fallback to default).
  • Update all routers (documents, graph, query, ollama) to accept Request and fetch workspace-scoped instances; change include_router usage to pass factory functions.
  • Adjust lifespan to pre-create default instances and finalize all cached RAG storages on shutdown; expose current workspace in /health; set default via set_default_workspace.
• WebUI:
  • Add workspace input and update action in header; persist to localStorage and reload.
  • Send LIGHTRAG-WORKSPACE with all requests (axios interceptor and streaming fetch).

^{Written by Cursor Bugbot for commit 4d2d781. This will update automatically on new commits. Configure here.}

[sillydong]

@codex review

[chatgpt-codex-connector]

To use Codex here, create a Codex account and connect to github.

[sillydong]

Hi @danielaskdd I proposed a PR based on your change to add workspace support for all APIs. Please help review.

[LarFii]

@cursor review

[cursor]

Bug: Race condition when creating cached LightRAG instances

The create_rag and create_doc_manager functions check and modify rag_cache and doc_manager_cache without any synchronization. When multiple concurrent requests arrive for the same workspace, each can pass the cache check (if workspace in rag_cache) before any has finished initializing and storing the instance. This causes multiple LightRAG instances to be created for the same workspace, with the expensive initialize_storages() and check_and_migrate_data() running in parallel. The last one to finish overwrites others in the cache, potentially leaving orphaned resources. An asyncio.Lock per-cache (or per-workspace keyed lock) is needed to make the check-and-create pattern atomic.

Additional Locations (1)

• lightrag/api/lightrag_server.py#L340-L354

 

[francis2tm]

Hey @danielaskdd , is there something I can do to help merge this PR?