All MPNet Base V2
Neural Network
All MPNet Base V2 is an embedding model that encodes sentences and paragraphs into 768-dimensional vectors for semantic search.
Specifications
- Input
- text
- Output
- vectors (embeddings)
- Context size
- 512 tokens
- Vector dimensions
- 768
Prompt cost
(per 1M tokens)
How it works All MPNet Base V2?
All MPNet Base V2 encodes sentences and short paragraphs into a 768-dimensional dense vector space for information retrieval, clustering, and semantic similarity assessment. The fixed dimensionality simplifies storage in Qdrant, pgvector, or other vector databases. In BotHub, the model is available via a unified OpenAI-compatible API: one key for all models, so you can replace the embedder in your pipeline without rewriting the integration. Pay in rubles for tokens used, with no subscription, VPN, or foreign card required; contracts and invoices are available for companies. The model is used for searching corporate knowledge bases, RAG support assistant loops, ticket deduplication, and review clustering.Frequently asked questions about All MPNet Base V2
You can use generated results for commercial purposes. You own all rights to the content you create. The only restriction: make sure your prompt does not include copyrighted third-party material. You are responsible for respecting the rights to any input data.
It is an embedding model: it turns text into a numerical vector used to measure semantic similarity. Suitable for semantic search, RAG on knowledge bases, clustering, deduplication, and similarity-based classification. Context is 512 tokens per chunk.
The output is not text, but a fixed-dimension vector — 768 numbers for each chunk. The response comes as a standard JSON array via an OpenAI-compatible API, so the vector can be written directly to Qdrant, pgvector, or another vector storage without conversion.
Reasoning is not provided by the format itself: the model does not generate text or build a chain of thought, but encodes a chunk into a vector in a single pass. The quality of the result is determined by chunking, normalization, and search configuration, not by the length of reasoning.
Function calling and structured output schemas are not noted in our data — these are tools for generative models. The embedding endpoint response itself already comes in JSON. If you need function calling, connect another model via the same API without rewriting the integration.
Acceptance of anything other than text is not noted in our data, so expect text chunks of up to 512 tokens. For images, audio, or video, choose a suitable model nearby in BotHub — the interface and API key remain the same.
We do not have data on the model's language coverage, so it is more reliable to test it on your own corpus: collect a small set of queries and evaluate the accuracy. There are other embedding models nearby in BotHub — you can compare them via the same API, with payment in rubles.