Find answers from the community

Updated 2 months ago

Do we have any native implementation in

Do we have any native implementation in LlamaIndex to help with "Lost in the middle: Performance degrades when models must access relevant information in the middle of long contexts". I find this implementation in LangChain https://api.python.langchain.com/en/latest/document_transformers/langchain.document_transformers.long_context_reorder.LongContextReorder.html ? Thank you

6 comments

I think @Logan M can help you in this. 🦇 🔦

@xrt nah we don't. All it does is re-order the retrieved chunks, so you could implement that easily with a custom node-postprocessor
https://gpt-index.readthedocs.io/en/stable/core_modules/query_modules/node_postprocessors/usage_pattern.html#custom-node-postprocessor

Low-key though, I don't know if it's actually helpful. You aren't reducing the size of the context, you still have stuff "in the middle" -- it's just re-ordering it.

Haystack actually released this first. I'm not convinced that it's helpful though

When this issue is visible, there is a limit after we see this happening ? or issue is the same if we send 4 or 12 chunks ?

Thanks, I will test.

Add a reply