[Question]:Cross-chunk relations & multi-hop retrieval design choices in LightRAG

[SangwookBaek]

Do you need to ask a question?

• I have searched the existing question and discussions and this question is not already answered.
• I believe this is a legitimate question, not just a bug or feature request.

Your Question

Hi LightRAG team 👋,

First, thanks for releasing such a clean and practical RAG framework—its dual-level retrieval has been working great for us!

While reviewing the indexing flow I noticed that

• entity / relation extraction is performed within each chunk only;
• during graph construction, entities deduplicated by D(·) become shared nodes, but no new edges are created between entities that never co-occur in the same chunk.

I’d love to understand the design rationale and its impact on deeper reasoning:

1. Why omit explicit cross-chunk edge creation?

   • Is the main goal to avoid graph explosion, reduce LLM calls, minimise hallucination risk, or something else?

2. Multi-hop coverage

   • In practice, does the current dual-level retrieval (low-level 1-hop + high-level keyword expansion) capture >2-hop relations reliably when the entities never appeared together?
   • Have you benchmarked long-range queries that depend on such paths (e.g. linking characters introduced in different chapters of a novel)?

3. Possible extensions

   • Would the project welcome an option to add lightweight cross-chunk edges, such as:

     • coreference-resolved “same entity” links
     • “same paragraph / same section” contextual edges
     • LLM-inferred relations between high-frequency entity pairs across chunks

   • Prompting an LLM on a vector‑based top‑k chunk set to infer cross‑chunk relations
     Any insights, papers, or code pointers would be greatly appreciated.
     Thanks again for the excellent work!

Best regards,
SangwookBaek

Additional Context

No response

[eric2323223]

Upvote this question.

[Eco-nom]

Reducing the chunk size and increasing the overlap size may help, but it won't be enough on its own.

[SangwookBaek]

Thank you for your detailed explanation.
While exploring this further, I also realized that simply adjusting chunk size and overlap isn’t really an effective solution — it seems more like a workaround than a robust fix.
Your insights confirmed my impression that the current chunk-based design, while practical, inevitably limits cross-chunk reasoning and relation coverage.
I appreciate your clarification and will definitely look deeper into the options you suggested.

[onestardao]

this is a super sharp observation — you're actually brushing up against one of the core edge cases in chunk-based RAG systems:

by skipping cross-chunk edge creation, the system implicitly assumes all semantics are chunk-local.

this works fine for low-entropy queries or shallow facts — but collapses hard on anything that requires distributed semantic resolution, like latent co-reference, symbolic links, or sequential narrative arcs.

your mention of "LLM-inferred links across high-frequency entity pairs" is especially important — many systems skip this due to cost or complexity, but without it, you often end up with confident completions that are logically disconnected across sessions.

curious if you've tested this on anything like character arcs or evolving definitions (e.g. legal terms across a contract corpus)? you'd be surprised how easily the current design masks drift under the hood.