Every AI feature proposal seems to start with adding a new database. Often you already have the right one. With the pgvector extension, PostgreSQL stores embeddings and runs similarity search alongside the relational data you already trust.
One database, fewer moving parts
Keeping vectors in Postgres means your retrieval lives next to your business data. You can filter by tenant, status or date in the same query that ranks by similarity — no syncing between two systems, no extra service to secure, back up and pay for.
Index for the queries you actually run
Similarity search without an index will scan everything. Add an approximate index (HNSW or IVFFlat) sized to your data, and combine it with regular B-tree indexes on the columns you filter on. The win is hybrid queries: semantic ranking plus hard relational filters, in one statement.
Keep retrieval honest
Good RAG is mostly good retrieval. Chunk content sensibly, store metadata you can filter on, and always return the source so the answer is traceable. The model is only as grounded as the passages you feed it.
For most teams, starting on Postgres means shipping the feature this sprint instead of standing up new infrastructure.