[BUG] KNN Query using FAISS score is incorrect

[NitinAggarwal1]

What is the bug?
We have used the KNN plugin in open search to try doing the vector search .

Details of the index are as follows :

• Index Mapping : "description_dense_vector": {
  "type": "knn_vector",
  "dimension": 768,
  "method": {
  "engine": "faiss",
  "space_type": "innerproduct",
  "name": "hnsw",
  "parameters": {}
  }
  }

• Query we use to get result set :
  GET ml-products/_search
  {
  "track_total_hits": true,
  "profile": "true",
  "_source": {
  "includes": [
  "product_id",
  "description",
  "p_product_catalog_search",
  "Brand_s",
  "multifield_text"
  ]
  },
  "size": 15,
  "query": {
  "neural": {
  "description_dense_vector": {
  "query_text": "mini kobalt tool kt",
  "model_id": "b_wGhJUB8vZIfNtyHHkc",
  "min_score": "0.1"
  }
  }
  }
  }

Issue 1 : For min_score = 0.8 we are getting 60 results , while for min_score = 0.9 we are getting 3356 results .
Issue 2 : Issue 2 : For max_distance = 0.8 we are getting 3 results , while for max_distance = 0.9 we are getting 0 results
It is just opposite the normal behaviour where min_score if reduced the recall should increase.

How can one reproduce the bug?
Steps to reproduce the behavior:

1. Go to '...'
2. Click on '....'
3. Scroll down to '....'
4. See error

What is the expected behavior?
A clear and concise description of what you expected to happen.

What is your host/environment?

• OS: Container-Optimized OS with containerd (cos_containerd)
• Version : 2.19
• Plugins : KNN Native Plugin

Do you have any screenshots?
If applicable, add screenshots to help explain your problem.

Do you have any additional context?
We see the faiss

[jmazanec15]

Thanks @NitinAggarwal1 . That seems backwards. Let me see if I can repro

[jmazanec15]

@NitinAggarwal1 just to confirm, the scores returned in the results are always what you expect, correct? Its more a matter of the number of matches? I imagine this has to do with radial search implementation in faiss. Did you try at all with Lucene engine?

[jmazanec15]

@NitinAggarwal1 are you still seeing a problem around this?