Instructor Embedding

// Overview

Instructor is an instruction-finetuned text embedding model. Instead of training a separate embedder for each job, you write a short natural-language instruction (for example, "Represent the Science title:") and prepend it to your text. The same model then produces embeddings tuned to that task and domain without any extra finetuning.

It is aimed at developers and researchers who need embeddings for several different tasks — retrieval, classification, clustering, or text similarity — and don't want to maintain one model per task. You install a small Python package, load a pretrained checkpoint from Hugging Face, and call a single encode function.

As an embedding model, it slots into the same place as any other sentence encoder in a search or RAG pipeline. The difference is the instruction prefix, which lets you steer the output vectors toward a specific task and subject area.

// What it does

• Instruction-conditioned embeddings: prepend a task instruction to steer the output vectors, no finetuning required
• One model covers many tasks — classification, retrieval, clustering, and text evaluation — across domains like science and finance
• Simple Python API: load a checkpoint and call a single encode function
• Pretrained checkpoints published on Hugging Face (for example, hkunlp/instructor-large)
• encode supports batching, progress bars, numpy or PyTorch tensor output, device selection, and optional normalization
• Reports state-of-the-art results across 70 diverse embedding tasks in the accompanying paper

// Getting started

Install the package, load a pretrained checkpoint, then call encode with instruction/text pairs.

pip install InstructorEmbedding

from InstructorEmbedding import INSTRUCTOR
model = INSTRUCTOR('hkunlp/instructor-large')

sentence = 'Dynamical Scalar Degree of Freedom in Horava-Lifshitz Gravity'
instruction = 'Represent the Science title:'
embedding = model.encode([[instruction, sentence]])
print(embedding)

Commands and code are distilled from the project's own documentation — always check the official repo for the latest.

// When to use it

• Power a retrieval or RAG pipeline by encoding queries and documents with a retrieval-specific instruction
• Generate features for text classification by representing each input with a domain-aware instruction
• Cluster documents (for example, by topic) using embeddings steered toward the clustering task
• Measure semantic similarity between texts with instruction-conditioned embeddings and dot-product or cosine scoring

// How Instructor Embedding compares

Instructor Embedding alongside other open-source embedding models & inference tools AI/TLDR tracks, ranked by GitHub stars.