Sentence Transformers: all-MiniLM-L12-v2

all-MiniLM-L12-v2 is a Sentence Transformers encoder for sentences and short paragraphs. It produces one 384-dimensional dense vector per text for semantic search, information retrieval, clustering and sentence similarity. The model uses the MiniLM-L12-H384-uncased backbone and was fine-tuned with contrastive learning on a billion-pair sentence collection.

Encode the unit you want to retrieve

Use a sentence or a short paragraph that expresses one useful idea. For a collection of product descriptions, keep the feature and the condition that distinguishes the item: “The adapter connects a wired headset to a USB-C port” is more informative than “Audio accessory.” Store the record identifier and original description beside its vector.

A search query can then be compared with the encoded descriptions. Inspect the returned candidates for exact requirements such as connector type or device compatibility before presenting a final choice. The vector represents the text’s meaning; the surrounding search application retrieves the corresponding records.

Mean pooling and normalized vectors

The published encoding example averages token representations with the attention mask, so padding positions do not contribute to the mean. It then applies L2 normalization. These steps are part of the reference recipe for obtaining the sentence vector, rather than an instruction to take an arbitrary token from the underlying transformer.

Keep the model, pooling and normalization configuration consistent between indexed texts and new queries. The resulting 384-component representations can be compared using cosine similarity. Keep a record of that configuration with the collection; a different 384-dimensional encoder does not become compatible simply because its vectors have the same length.

Default input length and training scope

The maker specifies default truncation beyond 256 word pieces. This is a tokenizer limit, not a count of ordinary words or characters. For a longer record, choose coherent short sections and preserve the heading or identifying detail needed to interpret each one.

The model’s training procedure used sequences of up to 128 tokens. That training setting is separate from the published 256-word-piece default at encoding time. A passage that fits the default can still differ substantially from the short sentence material used during training, so assess long or specialised descriptions with representative examples.

Similarity and clustering workflows

For a set of support messages, embeddings can supply numerical inputs to a grouping stage. Review several messages from each proposed group and those near its boundary before assigning a human-readable label. For duplicate-content review, retrieve similar records and compare their actual wording; two passages may discuss the same subject while giving different instructions.

This exact L12-v2 edition is an English sentence-embedding profile. It is distinct from all-MiniLM-L6-v2 and the paraphrase-MiniLM checkpoints. Shared family names do not transfer their sequence configuration or pooling recipe. Use its compact dense vectors together with source records and task-specific evaluation, rather than treating a similarity value as a verified answer or a final editorial decision.