BAAI: bge-large-en-v1.5

bge-large-en-v1.5 is BAAI’s large English dense embedding edition. It maps a text to a 1024-dimensional vector for passage retrieval and semantic comparison. Version 1.5 is the maker’s update aimed at a more useful similarity distribution and improved retrieval without an instruction. Large remains a distinct checkpoint from Base, Small, Chinese editions and the multilingual BGE-M3 model.

Set up passage retrieval

For a short search query against longer documents, the maker recommends the query instruction “Represent this sentence for searching relevant passages:”. Add it to the query side; corpus passages keep their own source text without that instruction. In other uses, direct text is sufficient. BAAI also documents retrieval without the prefix in v1.5 and recommends choosing the preparation that works best on the actual task.

A practical knowledge collection can index the sections of an English technical manual. A question about which setting changes the output format should return the section explaining that setting and its conditions. Retain the heading, relevant explanation and source location with each passage. The model provides the numerical representation; the returned source text is what a reader or a later RAG stage uses.

Representation and matching

The manufacturer’s direct encoding example uses the final hidden state of the first CLS token as the pooled embedding and then normalises the vector. Comparing normalised vectors with a dot product gives cosine similarity. Each passage can therefore be prepared once and stored for comparison with incoming query vectors, while changed source material receives an updated representation.

The 1024 dimensions specify the output vector length. They differ from the 768-dimensional Base representation and do not establish a compatible shared space with another checkpoint. Keep the model and preparation convention associated with the stored index. A change in the embedding edition calls for checking the corpus and query representations together.

Fit the source into the window

The documented sequence length is 512 tokens. Prepare a long manual as coherent sections rather than one oversized input. Keep a condition or exception beside the rule it qualifies, and include enough context for a retrieved section to make sense on its own. An added query instruction also occupies input tokens, so the model’s window concerns the full prepared input.

A useful passage boundary can include a heading and the paragraphs answering one user question. Preserve the parent-document identifier and version outside the vector so a result remains traceable to the correct source. For a comparison task, use the actual texts to compare rather than replacing them with unrelated summaries that remove the distinction you need to find.

Interpret the score

The maker warns against treating absolute similarity as a universal rule for semantic equivalence. Relative score order is central to retrieval, and a threshold should be calibrated on the task’s own text distribution. Review candidate passages for questions with known answers and for questions whose answer is absent. That reveals whether a returned section provides the required information rather than merely resembling the topic.

bge-large-en-v1.5 supplies English dense vectors. A separate reranker can refine candidate order, and a separate language model can compose an answer from the selected passages. The useful output from this edition is a searchable representation tied to its source text, with the exact Large v1.5 model choice and input preparation retained.