All MiniLM L12 V2
Neural Network
All MiniLM L12 V2 is a compact embedding model that converts sentences and paragraphs into 384-dimensional vectors for semantic search.
Specifications
- Input
- text
- Output
- vectors (embeddings)
- Context size
- 512 tokens
- Vector dimensions
- 384
Prompt cost
(per 1M tokens)
How it works All MiniLM L12 V2?
All MiniLM L12 V2 maps sentences and short paragraphs into a dense 384-dimensional vector space, outputting compact embeddings for semantic search, clustering, and text similarity comparison. The small dimensionality saves index memory and speeds up nearest neighbor search in Qdrant, pgvector, or other vector stores. In BotHub, the model works via a unified OpenAI-compatible API: one key for all models, access from Russia without VPN or foreign cards, and payment in rubles based on token usage. Connect it to a RAG pipeline over your knowledge base, build search for documentation and support tickets, group reviews by topic, and find duplicates in product listings.Frequently asked questions about All MiniLM L12 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 a compact embedding model from sentence-transformers: it turns text into a numerical vector. Such vectors are needed for semantic search, RAG over document bases, clustering, deduplication, and ticket classification. The context per request is up to 512 tokens, which is convenient for short fragments and chunks.
The output is not text or an image, but an array of floating-point numbers — a vector representation of the fragment. It is written to a vector database like Qdrant or pgvector and compared by cosine similarity. The response format is standard for embeddings via the BotHub OpenAI-compatible API.
Step-by-step reasoning is not noted in our data, and it is not needed for embeddings: the model encodes a fragment into a vector in a single pass. If a chain of thought is required, switch to a reasoning text model in BotHub without changing the integration.
Function calling is not noted in our data — this is not a chat model, but an encoder. However, the response comes structured: a JSON with an array of numbers that is easy to parse in your pipeline. It is more convenient to connect tools and functions at the text model level.
In our data, accepting anything other than text is not noted, so plan your pipeline around text fragments. If you need to index multimodal content, first convert it to text using a suitable BotHub model, and then calculate embeddings with this encoder.
We do not have data on training languages, so you should check the quality on your corpus: collect a small set of queries and reference documents and measure the retrieval accuracy. In BotHub, you can compare several encoders via one API and choose the best one.