Intfloat: E5-Large-v2

E5-Large-v2 is intfloat’s twenty-four-layer English text embedding model with a 1024-dimensional output. It is designed to represent queries and passages numerically for retrieval and other embedding tasks. The E5 preparation convention is part of its training: the input text carries a role prefix, and the encoded token representations are pooled into one vector for that text.

Keep the query and passage roles explicit

For asymmetric retrieval, start the question with “query:” and the corpus text with “passage:”. These roles distinguish what the user is looking for from the material that might answer it. For a symmetric comparison, paraphrase search, clustering or classification features, the maker recommends “query:” for the texts being embedded.

A practical collection can index English support explanations. Prepare a question such as “query: What should I check before replacing the connector?” and pair it with source passages describing the checks and exceptions. The passage text keeps its corresponding prefix. Store the source identifier and original wording beside each vector so the result can return the material a user needs to read.

The prefixes are not decorative labels to remove from one side of the task. The manufacturer states that omitting them degrades retrieval performance because the model was trained with that preparation. Apply the chosen role consistently when building the corpus and when later embedding a query.

Mean pooling and normalised vectors

The official encoding example averages the token representations while excluding padding positions with the attention mask. It then normalises the resulting vector. A dot product between normalised query and passage vectors supplies cosine similarity for ordering candidates. This mean-pooling recipe belongs to E5-Large-v2; a first-token pooling recipe from a different family is a separate preparation choice.

The 1024 output values form the embedding, not a written answer or a direct category label. A search stage compares those values with compatible corpus representations, and a later stage uses the selected passage text. If vectors serve as clustering or classifier features, retain the association between each vector and the input record whose features it represents.

English text within the window

The model card specifies English-only text and truncation of long inputs to at most 512 tokens. Keep each document passage coherent inside that window, including the role prefix and useful source context. A long manual should be prepared as sections whose titles, conditions and instructions can be understood when retrieved separately.

For example, a clause describing a setting may need the definition of that setting and an exception in the next sentence. Preserve that small unit together rather than splitting purely at a character count. Keep the parent document and source version with the passage so an update replaces the correct material.

Read the ranking as retrieval evidence

The maker explains that the ordering of cosine scores matters more than their absolute values. A score should not be interpreted as a calibrated probability that a passage answers the question. Check known answer-bearing passages and nearby irrelevant material when choosing how many results to retain or where to filter a candidate set.

The useful output is a corpus or query representation, a semantic comparison signal or numerical features for another task. This exact English Large v2 checkpoint remains separate from E5-Base-v2 and Multilingual-E5-Large. Matching dimensions alone do not make their vectors a documented common embedding space.