RAG & Retrieval · Ground answers in evidence
How does metadata filtering interact with vector search?
Short answer
Filters restrict which vectors are eligible before or during search. Pre-filtering is safer for permissions but can hurt ANN recall if the eligible set is tiny; post-filtering risks leaking or returning too few results.
Metadata filters express hard constraints: tenant, document type, date range, access group. They are not ranking signals. Where the filter is applied matters. Pre-filtering limits the search space, which is correct for security but can break an ANN index's assumptions. Post-filtering searches everything and discards disallowed results, which is simpler but can leak through side channels and often returns too few results.
- Apply permission and tenant filters before content reaches the prompt, always.
- Use an index or store that supports filtered search efficiently.
- Watch recall when filters are highly selective.
- Treat filters as data, not as instructions inside the prompt.
Follow-ups you should expect
Interviewers use these to check whether the first answer was understanding or memorisation.
- 1Why can a highly selective filter reduce ANN recall?
- 2How would you debug a filtered search that returns nothing?
- 3Would you denormalise metadata into the vector store?
Say it in the interview
Metadata filters are hard constraints, especially for permissions. I apply them before the model sees anything, using a store that supports filtered vector search efficiently. The danger is that a very selective filter can shrink the candidate pool and hurt ANN recall, so I measure both correctness and recall. Filtering after generation is not security; by then the model has already read the content.
Use this as a model, not a script. Saying it in your own words is what makes it survive follow-up pressure.
Related questions
Reading is step one
Now say it out loud.
Understanding an answer and delivering it under pressure are different skills. Rehearse this topic in a live RAG & Retrieval mock interview and get scored feedback on depth, clarity and confidence.