What it does

bertsearch is a reference setup, not a library. You bring a pretrained BERT model, feed documents through it to generate 768-dimensional embeddings, store those vectors in Elasticsearch’s dense_vector field, and query via a small Flask UI. The README walks through a job-search demo: CSV in, vectors out, browser open.

The interesting bit

The heavy lifting is the plumbing. The project wires together three containers—Elasticsearch, a BERT inference service, and a Flask frontend—so you don’t have to. The actual “search” is Elasticsearch’s cosine-similarity on dense vectors; BERT just supplies the embeddings. It’s a clean illustration of how to bolt semantic search onto a traditional engine without rewriting either.

Key highlights

• Pre-baked Docker Compose setup: docker-compose up and you’re running
• Supports Google’s original pretrained BERT models (Base/Large, cased/uncased, multilingual, Chinese)
• Uses Elasticsearch’s native dense_vector type—no custom plugin needed
• Includes end-to-end scripts: create index, vectorize CSV, bulk-index, search
• Flask UI served on localhost:5000 for manual queries

Caveats

• README warns BERT container needs >8GB RAM; this will not run happily on a small laptop
• No mention of GPU support, model quantization, or approximate nearest-neighbor indexing—scaling beyond toy data is an exercise for the reader
• Last commit activity and BERT model links suggest this targets the 2018-era Google BERT release, not modern successors

Verdict

Good for developers who want a concrete, working skeleton for semantic search with Elasticsearch and need to convince a team (or themselves) that the pieces fit. Skip it if you need production-scale vector search, modern models, or a maintained framework—this is a proof of concept that stayed one.