Entity linking that won't melt your GPU budget

What it does

ReLiK is a two-stage information extraction system: a retriever fetches relevant documents, then a reader extracts entities and relations from them. It handles both entity linking (mapping mentions to Wikipedia IDs) and relation extraction (subject-predicate-object triples) through a single unified pipeline. The project ships multiple model sizes, from a “thicc” extra-large variant down to a deliberately tiny one, all loadable via HuggingFace’s from_pretrained.

The interesting bit

The “academic budget” framing in the title is doing real work here. Most NLP pipelines this capable assume you have A100s to spare; ReLiK’s smallest entity-linking model is explicitly pitched as tiny and fast, and the README details CPU-only FAISS installation paths alongside GPU options. The retriever-reader split itself is classic, but the emphasis on keeping both stages lightweight enough for commodity hardware is the actual contribution.

Key highlights

• Pre-trained models for entity linking, relation extraction, or both combined (“Closed Information Extraction”)
• Multiple size tiers: small, large, and extra-large variants with different speed/accuracy tradeoffs
• HuggingFace integration with from_pretrained loading and a dedicated model collection
• Optional FAISS backend for retrieval, with separate CPU and GPU install paths
• FastAPI/Ray serving support via optional pip install relik[serve]
• Colab notebooks and a live HuggingFace Space for trying without installing

Caveats

• The README is truncated mid-example in the source, so some usage details for retriever-only and reader-only modes are incomplete
• FAISS GPU support requires conda installation from specific channels, not PyPI, which adds friction
• The model zoo is somewhat sprawling; the README lists overlapping entries (two different “Small for Entity Linking” variants) without clarifying differences

Verdict

Worth a look if you need entity linking or relation extraction in resource-constrained settings — academic labs, small cloud instances, or edge deployment. Skip it if you’re already running massive models on heavy hardware and don’t care about efficiency; you probably have better options.