Skip to content

llamaindex

Reads a StorageContext.persist directory: docstore.json and a SimpleVectorStore JSON (--vector-store, default default__vector_store.json; named stores use <name>__vector_store.json). Plain JSON, no pickle, no LlamaIndex import. Verified with indexes built and persisted by real llama_index.core.

  • Each node becomes a record: id = node id, text, vector from embedding_dict (or the node's own embedding), node metadata, plus ref_doc_id, relationships (llamaindex_source, _previous, _next, _parent, _child), llamaindex_node_type, start_char_idx, end_char_idx, mimetype, and any excluded-metadata key lists. creation_date / last_modified_date metadata set the timestamps.
  • Source Documents are records only when they have no chunks (--documents auto|always|never).
  • Ids present only in the vector store (external-style layouts) are imported with their metadata, and text from _node_content when present.
  • Limit: the format is whole-file JSON, and LlamaIndex loads it whole too, so memory is proportional to the persisted index.