multilingual-e5-large-instruct
by intfloat
The text embedding model Mixpeek runs by default, in four extractors
intfloat/multilingual-e5-large-instructOverview
Every piece of text that enters Mixpeek without a custom configuration is embedded by this model. It backs text_extractor@v1, the text side of document_graph_extractor@v1, the transcription, OCR and description indexes of multimodal_extractor@v1 and @v2, and the text index of web_scraper@v1. If you have run a Mixpeek collection over anything with words in it, you have used it.
It is XLM-RoBERTa-large fine-tuned for retrieval with an instruction prefix, which is why the query and the passage are encoded differently: a query gets a task description in front of it, a passage does not. Getting that prefix wrong is the most common way to lose several points of recall without any error appearing anywhere.
Why this one and not an English-only model that scores higher on English: it covers 100 languages, and a retrieval layer that silently degrades on a customer's French or Japanese content is a worse failure than a couple of nDCG points on BEIR. The numbers for that trade are in the FAQ below rather than asserted.
Architecture
XLM-RoBERTa-large encoder, 24 layers, 1024-dimensional hidden state, 559,890,432 parameters, 250,002-token vocabulary. max_position_embeddings is 514, so the usable window is 512 tokens including the two special tokens; longer text has to be chunked before encoding. Output is a 1024-dimensional mean-pooled vector, L2-normalized, compared with cosine.
Mixpeek SDK Integration
import { Mixpeek } from "mixpeek";
const mx = new Mixpeek({ apiKey: "API_KEY" });
// Managed, and this one is not a stand-in: text_extractor@v1 runs exactly this
// model. The collection gets a vector index named text_extractor_v1_embedding
// at 1024 dimensions, and nothing has to run on your hardware.
const collection = await mx.collections.create({
namespace_id: "my-namespace",
collection_name: "transcripts",
source: { type: "bucket", bucket_ids: ["bkt_your_bucket"] },
feature_extractor: {
feature_extractor_name: "text_extractor",
version: "v1",
parameters: { chunk_size: 512, chunk_overlap: 64 },
},
});Capabilities
- 1024-dimensional dense text embeddings across roughly 100 languages
- Instruction-prefixed queries, so query and passage encode asymmetrically
- Cross-lingual retrieval: an English query against Japanese passages in one index
- 512-token window, which is the chunk size to design around
Use Cases on Mixpeek
Benchmarks
| Dataset | Metric | Score | Source |
|---|---|---|---|
| MTEB retrieval, 15 datasets | Mean nDCG@10 | 52.47 | Computed from the model card's own model-index, 2026-09-01 |
| MTEB TRECCOVID | nDCG@10 | 82.00 | model-index, intfloat/multilingual-e5-large-instruct |
| MTEB QuoraRetrieval | nDCG@10 | 89.15 | model-index, intfloat/multilingual-e5-large-instruct |
| MTEB SciFact | nDCG@10 | 71.85 | model-index, intfloat/multilingual-e5-large-instruct |
Performance
intfloat publishes no latency or throughput figures. Mixpeek has not published its own for this model either, so none are quoted here rather than estimated from parameter count.
Common Pipeline Companions
Frequently Asked Questions
Does Mixpeek actually run multilingual-e5-large-instruct, or just support it?
It runs it. GET /v1/discovery/extractors lists it as the inference service behind four extractors: text_extractor@v1 (vector index text_extractor_v1_embedding, 1024-d), document_graph_extractor@v1 (document_graph_extractor_v1_text_embedding), multimodal_extractor v1 and v2 for their transcription, OCR and description indexes, and web_scraper@v1 for page text. No inference of your own is involved.
Is it better than bge-large-en-v1.5 for English retrieval?
No, and the gap is worth knowing before you pick. On the 15 MTEB retrieval datasets both model cards report, bge-large-en-v1.5 averages 54.29 nDCG@10 against 52.47, and wins 9 of the 15. E5 is ahead on TRECCOVID (82.00 against 74.82), FiQA (47.71 against 45.02), NQ, Quora, Touche and CQADupstack; BGE is ahead on FEVER (87.18 against 77.99), ClimateFEVER, ArguAna, DBPedia, HotpotQA, MSMARCO, NFCorpus, SCIDOCS and SciFact. The reason Mixpeek defaults to E5 is that bge-large-en-v1.5 is English-only, and a retrieval layer that quietly degrades on a customer's French or Japanese content costs more than two points of nDCG.
Do I need the instruction prefix, and what happens if I skip it?
Queries need it, passages do not. The format is "Instruct: {task description}\nQuery: {query}" on the query side only. Skipping it does not error; it encodes the query in a slightly different region of the space than the model was trained to put queries in, and you lose recall with nothing to indicate why. If you run through text_extractor@v1 the prefix is handled for you.
What is the real context limit?
512 tokens. The config declares max_position_embeddings 514, and two of those are the special tokens, so 512 is the usable window. Longer text has to be chunked before encoding, and text_extractor@v1 exposes chunk_size and chunk_overlap for exactly that. A chunk that overruns is truncated silently, so the tail of a long document simply does not get indexed.
Can I change the embedding model on a collection later?
Not without a migration. text_extractor@v1 takes an embedding_model parameter, but its own description is explicit: dimensions are fixed at namespace creation, so switching to a model with a different width means re-indexing. Choose the width before you load, not after. The parameter selects from Mixpeek's central embedding registry rather than accepting an arbitrary Hugging Face id.
Specification
Research Paper
Multilingual E5 Text Embeddings: A Technical Report
arxiv.orgBuild a pipeline with multilingual-e5-large-instruct
Add this model to a processing pipeline alongside other extractors. Combine with retrieval stages for end-to-end search.
Run it on your own data, free