Intfloat: E5-Base-v2

E5-Base-v2 is intfloat’s twelve-layer English embedding model with a 768-dimensional representation. It turns prepared text into one numerical vector for retrieval, semantic comparison or use as features in another task. The Base v2 checkpoint is distinct from the twenty-four-layer, 1024-dimensional Large v2 edition and the separately trained multilingual E5 models.

Use the trained role prefixes

E5 expects a role at the beginning of the input. For passage retrieval, a question starts with “query:” and its possible answer-bearing source starts with “passage:”. The manufacturer recommends “query:” on both sides for symmetric tasks such as semantic similarity and paraphrase retrieval. It also recommends that prefix when the vectors will be features for clustering or classification.

For example, a feedback collection can contain several differently worded requests for the same improvement. Encode those records with the recommended symmetric preparation, then compare or cluster the resulting vectors. Keep the original wording and record identifier with each representation so a cluster can be examined as actual customer messages rather than an unexplained group of numbers.

For a search collection, prepare questions and source passages in their corresponding roles. A useful query is “query: Which setting changes the notification frequency?” The corpus passage retains the explanation and its relevant conditions. The prefix is part of the model’s trained preparation; the maker notes that omitting it reduces performance.

Pool and compare the representation

The official example computes an attention-mask-aware average of the token representations, excluding padding, then normalises the resulting vector. A dot product between normalised embeddings provides cosine similarity. The 768 output values describe one text; they are not an answer, a class label or a pairwise reranker score.

In a clustering workflow, another stage forms groups from those vectors. In a classification workflow, another stage uses them as input features. Retaining that distinction makes the result easier to interpret: the embedding represents the record, while the downstream task determines a group or label. Use one model and preparation convention consistently for the representations being compared.

Work within English and 512 tokens

The model card says this edition works only for English and truncates long text to at most 512 tokens. Prepare coherent passages inside that span, including the required role prefix. A long review, report or manual should be divided around meaningful sections rather than embedded as an oversized input whose tail is lost.

Keep a heading with the sentences it introduces, and preserve the condition or exception that changes the meaning of a rule. For short records such as feedback, retain the complete message when it fits. A source identifier and version kept beside the vector let a later collection update replace the correct record.

Interpret similarity for the task

The maker explains that relative cosine-score order is more important than an absolute value. A filtering threshold should be assessed on the collection’s own examples, and vector similarity is not a calibrated probability of a matching label or correct answer. Inspect nearby records and known mismatches when deciding how the vectors will be used.

The useful deliverable is an English search index, a semantic comparison set or numerical text features paired with their source records. This Base v2 profile supplies its own dimensions, pooling and role preparation; another E5 edition’s vector shape or language coverage does not establish an interchangeable representation.