[poll] Post-filter or pre-filter? And have you measured what it costs you?

[akash-coded]

Calibration poll. The answer is well known. The second question is the one worth asking.

Documents carry permissions. A user may only see a subset.

Where does the filter go?

• A — After retrieval. Retrieve k, drop what they may not see
• B — Inside the retrieval query, as a predicate
• C — Both — pre-filter for the exact path, post-filter as a safety net
• D — Depends on filter selectivity

If you answered A

You are in good company and it is wrong twice, not once. It collapses k unpredictably — ask for
10, get 3, with no signal it happened — and the number of results is a function of content the
user cannot see, which is an inference channel.

If you answered B or C

Second question, and this is the real one: what is your recall against exhaustive search at 1%
filter selectivity? ANN benchmarks almost never filter and real systems always do. Pre-filtering
an ANN graph by removing nodes can disconnect the region containing the answer. Almost nobody has
this number for their own stack.

How to vote. The API cannot create a native poll widget, so vote with a reaction on this post — 👍 A · 🎉 B · 😄 C · 🚀 D — and reply if your answer needs a condition attached. The reveal is posted as a reply in about a week, never by editing this body: editing would destroy the record of what you were actually asked.