BAAI: bge-base-en-v1.5

bge-base-en-v1.5 is BAAI’s base-sized English dense embedding model. It converts each text into a 768-dimensional numerical representation for retrieval and semantic comparison. The model belongs to the BGE v1.5 line, which addresses the earlier family’s similarity-score distribution and improves retrieval when no instruction is supplied. Its output is a vector used to compare texts or find relevant passages.

Prepare queries and passages

For a short query searching longer passages, BAAI recommends prefixing the query with “Represent this sentence for searching relevant passages:”. The corpus passages do not receive that instruction. For other tasks, the original text can be embedded directly. The maker notes that v1.5 works without an instruction as well, so the preparation should be chosen around the task rather than treating the prefix as the only valid input.

A practical search collection can contain help articles, product explanations or technical instructions. A query such as “How do I change the export format?” should retrieve the passage which describes that action. Keep the section heading, relevant conditions and actual source text together. Store the passage’s identifier and location alongside its vector so a search result can return something the user can read or a later answer stage can cite.

Vector format and comparison

The documented representation uses the final hidden state of the first, or CLS, token as the sentence embedding. The manufacturer’s examples normalise the vector and compare normalised embeddings with a dot product, equivalent to cosine similarity. The 768 dimensions are the size of this model’s output representation; they are not a selectable context length or a text-generation allowance.

Use the same model and preparation convention for the query and corpus vectors being compared. A practical check is to retrieve known answer-bearing passages for a small set of actual user questions, then inspect the returned text. A compatible vector shape lets an index store the representation, while the source passage establishes what the result actually says.

Input length and document preparation

The model’s documented sequence length is 512 tokens. A long source should therefore be divided into coherent, answer-bearing passages before embedding. Keep definitions and exceptions with the statement they qualify, and preserve the connection from a passage to its parent document. The text to be embedded includes any instruction or heading you add, so a compact preparation leaves more room for the useful source content.

For a report or manual, choose boundaries around a section or a small group of related paragraphs. A passage that only contains a title or a fragment without its subject may be difficult to interpret when retrieved. Retain the original text and a source version so an index update can replace the corresponding material when the document changes.

Scores and useful output

BAAI cautions that absolute similarity values are not universal thresholds for deciding that two texts mean the same thing. The relative ranking matters, and a filtering threshold should be selected using the distribution on the task’s own data. A high vector similarity is a retrieval signal rather than verification of the passage’s factual content.

The useful deliverable is an English passage index, a set of query vectors or numerical features for semantic comparison. A separate reranker can refine retrieved candidates, and a separate generation stage can write an answer from their text. This exact base English v1.5 model remains distinct from BGE Large, Chinese editions and BGE-M3’s multilingual and hybrid retrieval profiles.