Lightweight Code Retrieval Models

MiniLM-based sentence-transformer models (22M/33M parameters) fine-tuned for domain-specific code retrieval achieving 97% Recall@10

Lightweight sentence-transformer embedding models (512-dim) for code search, small enough to run fast over a whole codebase.

  • Fine-tuned a MiniLM-based model for code-to-code and text-to-code retrieval.
  • Built for low-latency similarity search in RAG and developer tooling.
  • Released on Hugging Face, ready to drop into an embedding pipeline or vector database.

Results: 97% Recall@10 and 95% MRR@10 on internal benchmarks.

Models (public on Hugging Face):