Why do some 2026 embedding models accept an 'instruction' parameter at query time?

Attempt it

Models like Voyage v3 and the E5-Instruct family accept an optional 'instruction' or 'task' parameter at embedding time. Explain what this is, how it differs from the simple 'query:'/'passage:' prefix, and when it helps.

Free · 2 AI evals / day

Quick reference

TL;DR

An instruction parameter lets you describe the task at embedding time, so one model can serve many retrieval objectives without per-task fine-tuning.

Memory aid

Sign in for the mnemonic that makes this stick. See Pro pricing.

Easy to grasp

Imagine asking a librarian to fetch books. If you just hand them a topic, they grab anything on it. If instead you say I need beginner cookbooks, not chef textbooks for a restaurant kitchen, they pick differently. That extra sentence you whisper is the instruction. The librarian learned, over thousands of past visits, that the same book lives on a different mental shelf depending on why someone asked for it. So the same paragraph of text gets a different filing card depending on what task you announced at the door. Same library, same book, but a different shelf gets highlighted depending on why you came in.

Key concepts

Embeddings

Detailed answer & concept explanation

~4 min read

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.

Sign in for 5 free unlocks/day or go Pro for unlimited. See Pro pricing.

Quick scan · 2–4 min

Everything important, quickly.

6-8 min: what an instruction is + how training instills the behavior + three scenarios where it helps + index/query contract risk + 2026 vendor landscape.

Real-world examples

Real products, models, and research that use this idea.

Voyage v3 and Voyage-large-3 ship with task instructions for retrieval, similarity, and clustering.
E5-Mistral-Instruct (intfloat) accepts arbitrary instructions and tops several MTEB tracks in 2026.
BGE-en-icl and BGE-large-en-v1.5-instruct from BAAI support per-call instructions.
NV-Embed-v2 from NVIDIA uses instructions to switch between retrieval and similarity modes.

Interviewer follow-ups

What an interviewer would ask next. Try answering before peeking at the approach.

QHow would you A/B test two candidate instructions before shipping?

Sign in or go Pro to see the approach for this follow-up.

QWhat happens if a downstream team writes their own instruction without coordinating?

Sign in or go Pro to see the approach for this follow-up.

QWhy might a model that supports instructions still underperform a task-specific fine-tune?

Sign in or go Pro to see the approach for this follow-up.

Don't say this

Red flags and common mistakes that signal junior thinking. Click to expand.

Most common mistake

Confusing the embedding instruction with an LLM system prompt. The instruction steers the vector projection at encode time, not what the generator says later.

Sign in or go Pro to see all red flags and common mistakes. See Pro pricing.

60-second cheat sheet

The night-before-the-interview bullets. Scan these on the way to the call.

Sign in for 5 unlocks/day or go Pro for unlimited revision sheets. See Pro pricing.

References

Primary sources. Skim if you want the original framing.