Memector
Memector is a meme-template search built by Nicholas Hopper that fuses up to five ranking channels with Reciprocal Rank Fusion in pgvector.
What Memector does and how it is built
Searches 1,859 templates by fusing CLIP text-to-image, MiniLM text, Postgres keyword search and two sketch channels with Reciprocal Rank Fusion.
Measured recall@1 of 94% and recall@5 of 100% on 50 vague queries that do not contain the template's name, with median latency around 100 ms.
Raised candidate depth per channel from 80 to 150 after finding that a result one place past a channel's cutoff scored zero there, which moved recall@5 from 98% to 100%.