Add Tantivy back to the BM42 article (#1014)

qdrant-landing/content/articles/bm42.md

@@ -335,19 +335,22 @@ After encoding with BM42, the average vector size is only **5.6 elements per doc
 
 With `datatype: uint8` available in Qdrant, the total size of the sparse vector index is about **13Mb** for ~530k documents.
 
-As a referene point, we use the BM25 implementation with the same preprocessing pipeline: tokenization, stop-words removal, and lemmatization. There are other implementations of BM25 which might use different components and produce different results.
+As a reference point, we use:
 
-Those implementations are out of the scope for this article.
+- BM25 with tantivy
+- the [sparse vector BM25 implementation](https://huggingface.co/Qdrant/bm25) with the same preprocessing pipeline like for BM42: tokenization, stop-words removal, and lemmatization
 
-
-|                 | BM25 (Sparse) | BM42     |
-|-------------------|------|----------|
-|~~Precision @ 10~~ * | ~~0.45~~ | ~~0.49~~ |
-| Recall @ 10       | 0.83 | **0.85** |
+|                      | BM25 (tantivy)    | BM25 (Sparse) | BM42     |
+|----------------------|-------------------|---------------|----------|
+| ~~Precision @ 10~~ * | ~~0.45~~          | ~~0.45~~      | ~~0.49~~ |
+| Recall @ 10          | ~~0.71~~ **0.89** | 0.83          | 0.85     |
 
 
  \* - values were corrected after the publication due to a mistake in the evaluation script.
 
+<aside role="status">
+When used properly, BM25 with tantivy achieves the best results. Our initial implementation performed the character escaping that led to understating the value of <code>recall@10</code> for tantivy. 
+</aside>
 
 To make our benchmarks transparent, we have published scripts we used for the evaluation: see [github repo](https://github.com/qdrant/bm42_eval).