Sentence Transformers: all-MiniLM-L6-v2

all-MiniLM-L6-v2 is a Sentence Transformers sentence and short-paragraph encoder. It maps each text to a 384-dimensional dense vector for retrieval, clustering and sentence similarity. The model was fine-tuned from the MiniLM-L6-H384-uncased backbone using contrastive learning on a billion-pair sentence collection.

Prepare concise, meaningful records

Choose the sentence or paragraph that should appear in a search result. For an FAQ collection, retain the question and the short explanation that distinguishes its topic. A passage about resetting a forgotten password should keep the reset procedure, rather than being reduced to an isolated word such as “Account.”

For a long article, index coherent short sections and keep their source locations. Retrieve candidate sections by comparing the query’s vector with the stored representations, then show the original passages so the reader can inspect the actual instructions.

Obtain the sentence representation

The maker’s direct encoding recipe averages token representations with the attention mask, excluding padding positions from the mean. It then applies L2 normalization. Those steps produce the sentence embedding used for comparison; individual token representations from the underlying transformer serve a different role.

Each output has 384 components. Use the same checkpoint, pooling and normalization when building the collection and encoding later queries. Cosine similarity can compare the resulting vectors. A similarity value is useful for ranking candidates under that setup, while the original text determines whether a candidate answers the question.

Length and edition boundaries

The published default truncates inputs longer than 256 word pieces. This tokenizer unit does not correspond directly to a normal word or character. Keep important conditions inside the indexed unit, and split a lengthy procedure at meaningful boundaries before encoding it.

Training sequences were limited to 128 tokens. That training setting is separate from the 256-word-piece default used for encoding. For unusually long or specialised passages, use representative query-and-passage examples to assess the retrieval behavior rather than assuming that fitting the input boundary guarantees equal treatment of every detail.

The exact L6-v2 edition is an English embedding profile. The L12-v2 and paraphrase-MiniLM editions have separate model identities and configurations. Their common MiniLM name or 384-dimensional output does not establish interchangeable vectors.

Search, similarity and grouping

For duplicate-question review, retrieve similar questions and compare the actual wording before merging records. “Can I reset my password?” and “Can an administrator reset another user’s password?” concern related subjects but different permissions. Preserve that distinction in the encoded text and in the review stage.

For clustering, use the embeddings as numerical inputs to a grouping process, then inspect representative records before naming a group. This model supplies vectors rather than written cluster labels or a generated FAQ answer. Its useful result is a compact representation of a short source passage that remains linked to the original record.