Update README.md

README.md

@@ -8,16 +8,25 @@ pipeline_tag: sentence-similarity
 
 # MiniCOIL v1
 
-MiniCOIL - is a sparse contextualized per-token embeddings.
-Read more about it in [the article](https://qdrant.tech/articles/minicoil).
+MiniCOIL is a sparse neural embedding model for textual retrieval. 
+
+It creates 4-dimensional embeddings for each word stem, capturing the word's meaning.  
+These meaning embeddings are combined into a bag-of-words (BoW) representation of the input text.  
+The final sparse representation is calculated by weighting each word using the BM25 scoring formula.
+
+<img src="https://storage.googleapis.com/qdrant-examples/miniCOIL_inference.png" alt="miniCOIL inference" width="600"/>
+
+In the case of a word's absence in the miniCOIL vocabulary, word weight in sparse representation is purely based on the BM25 score.
+
+Read more about miniCOIL in [the article](https://qdrant.tech/articles/minicoil).
 
 
 ## Usage
 
-This model is designed to be used with [FastEmbed](https://github.com/qdrant/fastembed) library.
+This model is designed to be used with the [FastEmbed](https://github.com/qdrant/fastembed) library.
 
 > Note:
-This model is supposed to be used with Qdrant. Vectors have to be configured with [Modifier.IDF](https://qdrant.tech/documentation/concepts/indexing/?q=modifier#idf-modifier).
+This model was designed with Qdrant's specifics in mind; miniCOIL sparse vectors in Qdrant have to be configured with [Modifier.IDF](https://qdrant.tech/documentation/concepts/indexing/?q=modifier#idf-modifier). Otherwise, you'll have to personally calculate & scale the produced sparse representations by the IDF part of the BM25 formula.
 
 ```py
 from fastembed import SparseTextEmbedding