Migrating from chromadb to qdrant for RAG creates inconsistencies

[JINO-ROHIT]

I have been using llamaindex for RAG with chromadb, and the search metric was the L2 squared.

When I migrate my data to qdrant, the retrieved nodes are completely different despite following the same chunking strat. I noticed that Euclidean distance is the closest metric to L2 squared that takes a root. Even using this gives different results.

Is this entirely due to the search metric? Or is there something else going on?

Thanks

[generall]

Hey @JINO-ROHIT, if you compute the metric values manually, which version does it match?

[JINO-ROHIT]

@generall it matches the chromadb version for L2 squared

[generall]

could you please share an example?