Is it possible to change the graph_storage from an existing working_dir?

[hungphongtrn]

I have already initialized a storage with "Neo4JStorage" using the following snippet:

rag = LightRAG(
    working_dir=LIGHTRAG_NEO4J_WORKING_DIR,
    embedding_func=openai_embed,
    llm_model_func=gpt_4o_mini_complete,
    graph_storage="Neo4JStorage",
    enable_llm_cache=False,
    enable_llm_cache_for_entity_extract=False,
)

Is it possible for me to use the existing data in LIGHTRAG_NEO4J_WORKING_DIR, and initialize a storage with other graph_storage options such as NetworkXStorage or MemgraphStorage? Thank you in advance.

[onestardao]

It looks like your question is actually related to Problem #7 (storage mismatch between backends) — each graph_storage backend writes its own format, so you can’t just swap Neo4JStorage to NetworkXStorage or MemgraphStorage using the same working_dir.

If you want, I can share a short guide (with link) that shows how to handle this case and what options you have to migrate cleanly. Just let me know.

[RedEye1605]

No, you cannot directly switch the graph_storage backend on an existing working_dir and expect it to work. Each storage backend writes data in its own format:

• NetworkXStorage → writes to local JSON/pickle files in working_dir
• Neo4JStorage → stores graph data in Neo4j database
• MemgraphStorage → stores in Memgraph
• PostgresStorage → stores in PostgreSQL

The working_dir structure is specific to each backend. For example, NetworkX stores the graph as graph_chunk_entity_relation.graphml while Neo4j stores it in the database directly. They are not interchangeable.

Migration options:

1. Re-ingest: The cleanest approach is to re-ingest your source documents with the new storage backend. This ensures the graph is properly constructed.

2. Export and rebuild: If you need to preserve entity/relationship data:

   • Use rag.export_data() to get a CSV of your graph data
   • Initialize a new LightRAG instance with the new graph_storage
   • Re-ingest the original documents

3. If you want to keep the vector store but change only the graph storage: This is tricky since the working_dir couples all storage together. You'd need to copy the vector store files (e.g., the embedding index files) to a new working directory and initialize with the new graph storage, but this is not officially supported and may cause inconsistencies.