# Pre-Filtering, Partial Indexes and Partitions — Vector Databases

Source: https://www.geekswithgeeks.com/en/vector-databases/f-strategies

> Choose a physical design for multi-tenant and metadata filters.

## Match the design to the filter

Choose by how selective and how stable the filter is. **Few huge tenants**: a **partial index per tenant** (`CREATE INDEX ... USING hnsw (...) WHERE tenant = 'acme'`) or a **table partitioned by tenant** gives each tenant its own small, accurate index, at the cost of more indexes to build and maintain. **Many small tenants**: one index with a tenant column and iterative scans or an exact scan per tenant (a few thousand rows scan quickly). **Low-selectivity filters** (for example language = en matching 80% of rows): post-filtering with some over-fetching works. **Date ranges**: partition by time and search only the relevant partitions. Always verify with `EXPLAIN` and a recall measurement, since the planner's choice depends on statistics. In a dedicated engine, create a **payload index** on every field you filter on and prefer engines that evaluate filters during graph traversal.

## Filter design cheat sheet

Starting points; verify with EXPLAIN and measured recall.

```text
Filter shape                                   Good design
few very large tenants                          partial index or partition per tenant
many small tenants (< ~10k rows each)           one index + iterative scan, or exact scan per tenant
low selectivity (matches most rows)             plain index + post-filter, over-fetch a little
range by time (recent data matters)             time partitions; search only the relevant ones
hard security boundary                          row-level security / mandatory tenant predicate in every query
```

## Verify the plan, not the assumption

After changing indexes or partitions, run EXPLAIN and re-measure recall. Planner choices depend on statistics.

**Quiz:** For a few very large tenants, which design gives each tenant an accurate index?

- [ ] Filtering in the UI
- [ ] One index with no tenant column
- [ ] No indexes at all
- [x] A partial index (or partition) per tenant

*Answer:* A partial index (or partition) per tenant. Each index then contains only rows that can match.