@Jasper_Simon actually, I think it might be OK. But I need to add some documentation.
Could you try specifying a unique "name" field for each of your inner_hits objects? Something like:
{
"knn": [
{
"field": "paragraphs.titleEmbedding",
"query_vector": [
0.1, 0.2
],
"k": 20,
"num_candidates": 50,
"inner_hits": {
"name": "title",
"_source": [
"paragraphs.documentTitle"
],
"fields": [
"paragraphs.documentTitle"
]
}
},
{
"field": "paragraphs.textEmbedding",
"query_vector": [
0.1, 0.2
],
"k": 20,
"num_candidates": 50,
"inner_hits": {
"name": "text",
"_source": [
"paragraphs.documentText"
],
"fields": [
"paragraphs.documentText"
]
}
}
]
}
I think the issue is that we attempt to automatically name the inner_hit to appropriately extract the correct field information given a configuration. However, we have no way of distinguishing these inner_hit configurations.