@@ -9,25 +9,10 @@ highlight:
99 body : |-
1010 New options have been added to the `dense_vector` field type.
1111
12- The first is support for storing vectors in https://en.wikipedia.org/wiki/Bfloat16_floating-point_format[ bfloat16]
13- format. This is a floating-point format that utilises 2-bytes per value rather than 4, halving the storage space
12+ The first is support for storing vectors in bfloat16 format.
13+ This is a floating-point format that utilises 2-bytes per value rather than 4, halving the storage space
1414 required compared to `element_type: float`. This can be specified with `element_type: bfloat16`
15- when creating the index, for all `dense_vector` indexing types:
16-
17- [source,yaml]
18- ----------------------------
19- PUT vectors
20- {
21- "mappings": {
22- "properties": {
23- "vector": {
24- "type": "dense_vector",
25- "element_type": "bfloat16",
26- "index_options": {"type": "bbq_hnsw"}
27- }
28- }
29- }
30- ----------------------------
15+ when creating the index, for all `dense_vector` indexing types.
3116
3217 Float values are automatically truncated to 2-bytes when writing to disk, so this format can be used
3318 with original source vectors at 2- or 4-byte precision. BFloat16 values are zero-expanded back to 4-byte floats
@@ -43,19 +28,5 @@ highlight:
4328 on-disk during rescoring, and to not read it into memory first. This may significantly increase search performance
4429 in such low-memory situations.
4530
46- This is specified as an index option for HNSW vector index types:
47-
48- [source,yaml]
49- ----------------------------
50- PUT vectors
51- {
52- "mappings": {
53- "properties": {
54- "vector": {
55- "type": "dense_vector",
56- "index_options": {"type": "bbq_hnsw", "on_disk_rescore": true}
57- }
58- }
59- }
60- ----------------------------
31+ This is specified using the `on_disk_rescore: true` index option.
6132 notable : true
0 commit comments