Hi, thank you for releasing GraphRAG-Benchmark.
I noticed that BAAI/bge-large-en-v1.5 supports at most 512 input tokens, while some pipelines can generate longer chunks. For example, LightRAG uses chunk_token_size=1200, but truncates text to about 400 tokens before embedding.
Does this mean that only the prefix of a long chunk is represented in the vector index?
If so, relevant information located after the truncation point may not affect retrieval similarity, even if the full chunk text is retained for later generation.
Could you clarify:
- Is truncating overlong chunks before embedding the intended behavior?
- How should information beyond the 512-token limit be handled?
I would greatly appreciate any guidance or recommended configuration for using BGE-large safely while preserving retrieval coverage.
Hi, thank you for releasing GraphRAG-Benchmark.
I noticed that
BAAI/bge-large-en-v1.5supports at most 512 input tokens, while some pipelines can generate longer chunks. For example, LightRAG useschunk_token_size=1200, but truncates text to about 400 tokens before embedding.Does this mean that only the prefix of a long chunk is represented in the vector index?
If so, relevant information located after the truncation point may not affect retrieval similarity, even if the full chunk text is retained for later generation.
Could you clarify:
I would greatly appreciate any guidance or recommended configuration for using BGE-large safely while preserving retrieval coverage.