Questions on connecting adjacent chunks and on subsequent invocations of indexing

[Kekuaka]

Hello! Thank you for sharing your work. I've read the paper and then tested the code from this repo. I like it.
Could you please clarify few questions.

1. The paper does not state that adjacent chunks (passages) are connected by the graph edges, although this is implemented in the code within the add_adjacent_passage_edges() method. Due to this, the prepared dataset should contain numbered chunks (passages). Is this a crucial feature of the indexing and the graph construction method, or is it optional? When the adjacent chunks come from different documents should they be connected by the graph edge?

2. When the code is executed, the graph edges are doubled in subsequent invocations of the index() method with empty passage list. The simplest test is based on the run.py script. I think this is not expected behavior.
   ...
   rag_model = LinearRAG(global_config=config)
   questions,passages = load_dataset(args.dataset_name)
   rag_model.index(passages)
   rag_model.index([])

2025-12-30 14:22:25,591 - INFO - Using retrieval method: BFS Iteration
Indexed 11 passages
Graph vertices: 11
Graph edges: 10
Indexed 0 passages
Graph vertices: 11
Graph edges: 20

[LuyaoZhuang]

Thanks for your interest in our work!

Regarding Q1: Connecting adjacent chunks is an optional trick, not a requirement. When all chunks come from different documents, this operation is unnecessary. The purpose is to preserve contextual integrity, which aligns with LinearRAG's design philosophy.

Regarding Q2: This can be fixed by adding a conditional check when adding edges. We'll update the code.