The community member has a hybrid Qdrant database with sparse and dense vectors. They notice that the query scores are very weak, rarely exceeding 0.5, even when the answer is correct. They ask the community for suggestions on how to properly set up the query_engine parameters (similarity_top_k, alpha, etc.).
Another community member responds that the scores are relative to the model and not absolute measurements. They provide the example of OpenAI embeddings, where similarities are usually between 0.7 and 0.9 for any query, and that it can be different for other embedding models and methods.
Hello Guys, I have an hybrid Qdrant database (i.e sparse and dense vectors). When querying, I notice that the scores are very weak, even if the answer is correct. it rarely exceeds 0.5. Do you have any idea how can I properly set up the query_engine parameters ? (similarity_top_k, alpha,...) ?
imo scores are relative to the model, and not absolute measurements
openai embeddings for example, the similarities are usually between 0.7 and 0.9 for any query. For other embeddin models and methods, it can be different
