PI-2526 Switch to cosine similarity + increase chunk size (#4552)

The Data Science team have been using cosing similarity for their testing, whereas OpenSearch defaults to L2 (Euclidean distance).

Also increasing chunk size as a workaround for the batch size limit of 32 in the mxbai model, which limits the length of the notes we can process in one go to 1792 chars (now 3584).

Author: marcus-bcl

projects/person-search-index-from-delius/container/pipelines/contact/index/index-template-semantic.yml

@@ -799,6 +799,7 @@ template:
           knn:
             type: knn_vector
             dimension: 1024
+            space_type: cosinesimil
       notes:
         copy_to: text
         type: text

projects/person-search-index-from-delius/container/pipelines/contact/index/ingest-pipeline.tpl.json

@@ -25,7 +25,7 @@
       "text_chunking": {
         "algorithm": {
           "fixed_token_length": {
-            "token_limit": 64,
+            "token_limit": 128,
             "overlap_rate": 0.125,
             "tokenizer": "standard"
           }