Clarification about Graph Structure Retrieval implementation

[joaomr7]

Hello RAKG authors,

I am currently reproducing the RAKG framework as part of my undergraduate thesis on document-level knowledge graph construction and retrieval.

While examining the current public repository, I noticed a possible difference between the methodology described in the paper and the available implementation.

The paper and the repository README describe two retrieval components:

1. Corpus Retrospective Retrieval;
2. Graph Structure Retrieval.

In particular, the paper describes:

• vectorizing an initial knowledge graph;
• retrieving related entities and their relationship networks;
• providing both retrieved text and graph information to the LLM;
• merging the generated knowledge graph with the initial knowledge graph.

However, in the execution path I examined, the method get_target_kg_sigle appears to retrieve information only from document sentences.

When the entity-centered extraction chain is invoked in the get_target_kg_sigle method, the related_kg argument is set to "none":

result = chain.invoke({
"text": chunk_text,
"target_entity": entity_dic[entity_id].get("name"),
"related_kg": "none"
})

I was unable to identify the implementation responsible for:

• vectorizing the initial knowledge graph;
• retrieving related nodes, edges, or subgraphs;
• providing the retrieved relationship network to the extraction prompt;
• merging the generated graph with the initial knowledge graph.

Could you please clarify:

1. Is Graph Structure Retrieval implemented elsewhere in the repository?
2. Is a specific configuration, script, or execution path required to enable it?
3. Was this component used in the experiments reported in the paper?
4. Is the public repository the complete implementation used for the paper?

If this functionality is available in another branch, version, or internal implementation, any additional details, pseudocode, or documentation would be very helpful for reproducing the method.

Thank you for your time.