Sentence Transformers: multi-qa-mpnet-base-dot-v1

multi-qa-mpnet-base-dot-v1 is a Sentence Transformers model designed to retrieve text passages for a question or query. Its training used about 215 million question–answer pairs. Each input becomes a 768-dimensional dense vector, with CLS pooling and dot-product scoring defining this exact edition.

Prepare a question and candidate passages

Write the query as the information need and prepare corpus entries as readable text passages. For example, a question about changing a delivery address can be compared with separate paragraphs covering address changes, order cancellation and delivery tracking. Keep the relevant conditions in the address-change passage, including whether the order has already been dispatched.

Encode the query and passages with this same model. The published example does this directly, without a query/document prefix, and sorts the passages by their dot-product scores. Keep each vector attached to its original passage and document identifier so the ranked result can show the material that was actually indexed.

CLS pooling and the dot-product edition

The model’s reference recipe takes the representation of the first token, known as CLS pooling. Its technical table explicitly states that the embeddings are not normalized and lists dot product as the suitable scoring function. Preserve that recipe when setting up the collection and the query side.

L2-normalizing these vectors changes their scoring geometry. The normalized all-mpnet-base-v2 profile uses a different recipe and is a separate checkpoint, despite sharing the MPNet family and a 768-component output. Keep the exact edition and comparison function with the stored collection instead of choosing them from the vector length alone.

Length limit and longer passages

The stated input limit is 512 word pieces; longer text is truncated. The maker also notes that training inputs reached only 250 word pieces and that longer text may work less well. The accepted tokenizer length and the training-length caveat therefore describe different practical boundaries.

For a large help article, prepare short sections that answer specific questions. Retain a heading where it identifies the topic, and keep a rule with the exception needed to read it correctly. A single full-page input may lose later material through truncation, while an overly small fragment can lose the condition that makes it useful.

Use the ranking in retrieval and RAG

The dense vectors support a first retrieval stage. Return the highest-scoring candidate passages with their source references, then inspect whether they address the actual question. In a RAG workflow, supply those passages to the answer-writing stage as source material; retrieval itself does not compose the answer.

Use representative questions with known relevant passages to choose the result count and assess the ranking. Include examples with similar vocabulary but different intent, such as modifying an address versus viewing the current address. Interpret scores relative to candidates under the same model and scoring setup, rather than as calibrated probabilities of correctness.

The useful output of this checkpoint is a question-and-passage embedding space for semantic search. Its unnormalized vectors, CLS pooling, dot-product score and published long-input caveat are the details to preserve when preparing and using results.