How to Use multilingual-e5-large-instruct for Semantic Search

Keyword search works well when users know exactly what they are looking for. Enterprise search is rarely that simple.

Employees may search using different terminology, languages, abbreviations, or incomplete questions. This is where semantic search becomes useful. Instead of matching exact words, semantic search compares the meaning of a query with the meaning of stored content.

The multilingual-e5-large-instruct model from intfloat is designed for this type of text embedding task. The model supports multilingual applications and produces 1024-dimensional embeddings.

Step 1: Install the Model

A simple implementation can use Sentence Transformers:

from sentence_transformers import SentenceTransformer

‍

model = SentenceTransformer(

    "intfloat/multilingual-e5-large-instruct"

)

You can then convert documents into vectors.

documents = [

    "Employees can request annual leave through the HR portal.",

    "Expense claims must be submitted within 30 days."

]

‍

doc_embeddings = model.encode(documents)

These embeddings can be stored in a vector database.

Step 2: Encode the User Query

Multilingual E5 Large Instruct is slightly different from a basic embedding model because queries should contain an instruction describing the retrieval task.

For example:

Instruct: Retrieve HR policy information

Query: How many vacation days can I take?

The model documentation recommends this instruction format and notes that instructions are important for retrieval performance.

Step 3: Search Your Vector Database

When a user submits a question, generate its embedding and compare it with the document vectors using cosine similarity or another supported distance metric.

The system can then return the most relevant documents.

A production architecture might look like:

User Query → E5 Embedding → Vector Search → Reranking → Relevant Documents → LLM

This architecture works particularly well for RAG applications.

What About Multilingual Search?

One of the biggest advantages of multilingual embeddings is that documents and queries do not necessarily have to use the same language.

For example, an employee could ask a question in Spanish while the relevant company document is written in English. A multilingual embedding model can help bring these semantically related pieces of content closer together in vector space.

That makes multilingual-e5-large-instruct particularly interesting for global organisations with distributed teams.

What Should You Evaluate?

Do not judge semantic search purely by whether the model generates embeddings successfully.

Measure retrieval precision, recall, latency, multilingual performance, relevance of top-k results, and performance on your actual business vocabulary.

Also test different chunk sizes and metadata filters. A strong embedding model cannot compensate for poorly structured source data.

Conclusion

Multilingual E5 Large Instruct provides a solid foundation for semantic search, but the embedding model is only one component of the system.

The real performance comes from combining the model with good document preparation, vector indexing, retrieval logic, reranking, and continuous evaluation.

‍