I'm working with a vector database (qdrant) and using semantic search via embeddings (Sentence Transformers). My use case involves queries like: "Give animals other than cats" or "List inventions except electrical ones" Since vector search is based on semantic similarity, including "cats" or "electrical" in the prompt still increases their similarity score — even though I want to exclude those concepts. The vector database returns results that include the excluded concept (e.g., cats), because it's semantically close to the query due to the presence of the term in the prompt. What are the best practices for handling negative search queries (like "not", "except", or "other than") in a vector search system? I want to know: Can any vector databases natively support negative filters on embeddings? How to post-process or pre-process queries to make exclusion work better? Is there a standard way to do this using hybrid search or metadata filtering? Should I rely on LLMs (e.g., GPT-4) to re-rank or filter after the search? Continue reading...
