[BUG] Unable to index document containing knn_vector using NLP ingest pipeline

[whittssg]

What is the bug?

Indexing a document with knnvector errors

How can one reproduce the bug?

?

What is the expected behavior?

It works

What is your host/environment?

Windows (just latest stabe release downloaded and setup today)

---

I followed this tutorial and everything works perfectly but when i try and index somthing via this library it errors..
https://opensearch.org/docs/latest/search-plugins/neural-search-tutorial/

This is the code i am using in c# to create the index:

var response = openSearchClient.Indices.Create("df2_kasf", indexes => indexes
        .Settings(s => s
            .Setting("index.knn", true)
            .DefaultPipeline("nlp-ingest-pipeline"))
        .Map<kas>(x => x
            .AutoMap()
            .Properties(p => p
                .KnnVector(kv => kv
                    .Name(n => n.passage_embedding)
                    .Dimension(768)
                    .Method(m => m
                        .Engine("lucene")
                        .SpaceType("l2")
                        .Name("hnsw")))

The index creates without error but indexing gives this error:

{Type: mapper_parsing_exception Reason: "failed to parse field [passage_embedding] of type [knn_vector] in document with id '000017498'. Preview of field's value: 'having a good day'" CausedBy: "Type: illegal_argument_exception Reason: "Vector dimension mismatch. Expected: 768, Given: 0"}

I am filling the passage_embedding with text.

Is there something else i need to specify in the .net library for vector indexing?

Thanks

[Xtansia]

Hi @whittssg,

Could you please provide a more complete code sample including how you've created the ingest pipeline and how you're attempting to index the document?

[whittssg]

@Xtansia I followed the tutorial on https://opensearch.org/docs/latest/search-plugins/neural-search-tutorial/ exactly.... names included.. Manually indexing works via the dashboard dev tools

[Xtansia]

@whittssg If manually indexing via dev tools works then please share the code you're using to attempt to index the document in C#.

[whittssg]

I am just filling the field that is specified as the vector field in the index creation... nothing special:

  newka.passage_embedding = "some cool text to test this search";
  if (counter % 5 == 0 && counter != 0)
  {
      Console.WriteLine("KA - Bulk Inserting: " + counter);

      var result = openSearchClient.Bulk(descriptor);
      descriptor = new BulkDescriptor();


  }

and as i mentioned above this is how i created the index in c#:

` indexes => indexes
.Settings(s => s
.Setting("index.knn", true)

              .DefaultPipeline("nlp-ingest-pipeline"))

                                     .Map<kas>(
                                         x => x.AutoMap()
                                         .Properties(p => p
                                         
                                                 .KnnVector(kv => kv .Name(n => n.passage_embedding).Dimension(768).Method(m => m.Engine("lucene").SpaceType("l2").Name("hnsw")))`

[Xtansia]

In the tutorial the ingest pipeline is configured to take an input field named text and map its embeddings into a field named passage_embeddings: https://opensearch.org/docs/latest/search-plugins/neural-search-tutorial/#step-2a-create-an-ingest-pipeline-for-neural-search
Note the field_map configuration:

PUT /_ingest/pipeline/nlp-ingest-pipeline
{
  "description": "An NLP ingest pipeline",
  "processors": [
    {
      "text_embedding": {
        "model_id": "aVeif4oB5Vm0Tdw8zYO2",
        "field_map": {
          "text": "passage_embedding"
        }
      }
    }
  ]
}

So when you're indexing documents you need to send your string in the text field as is done in the tutorial not directly into the passage_embeddings field: https://opensearch.org/docs/latest/search-plugins/neural-search-tutorial/#step-2c-ingest-documents-into-the-index

PUT /my-nlp-index/_doc/1
{
  "text": "A West Virginia university women 's basketball team , officials , and a small gathering of fans are in a West Virginia arena .",
  "id": "4319130149.jpg"
}

[whittssg]

oh ok, so do i need to change this (i changed the field to text):
.KnnVector(kv => kv .Name(n => n.text).Dimension(768).Method(m => m.Engine("lucene").SpaceType("l2").Name("hnsw")))

I changed it to text as you can see above but now get the same error"

illegal_argument_exception Reason: "Vector dimension mismatch. Expected: 768, Given: 0""}

[whittssg]

So it will index if i change that but querying will give this error:

"error": { "root_cause": [ { "type": "query_shard_exception", "reason": "failed to create query: Field 'text' is not knn_vector type.", "index": "df2_kasf", "index_uuid": "XiphTWaXQpWeq_wST-lNwA" } ], "type

I think i am missing something obvious

[whittssg]

Creating the index should read this right to match the tutorial:

.KnnVector(kv => kv .Name(n => n.text).Dimension(768).Method(m => m.Engine("lucene").SpaceType("l2").Name("hnsw")))
or should it be
.KnnVector(kv => kv .Name(n => n.passage_embedding).Dimension(768) .ModelId("").Method(m => m.Engine("lucene").SpaceType("l2").Name("hnsw")))

[whittssg]

the only way my code indexes is if i rem out this:
KnnVector(kv => kv .Name(n => n.passage_embedding).Dimension(768) .ModelId("").Method(m => m.Engine("lucene").SpaceType("l2").Name("hnsw")))

I can see documents indexing once this is gone.

Maybe i am just missing somethings stupid, is there an example some where for this type of search?

[whittssg]

I think i should have been more clear, i followed everything upto the creation of the index on that tutorial (i went through it all and everything worked perfectly but now i want to do it via c#). So instead of doing the models etc in c# i started at the create index step in c# (since the models etc were created via the puts in the tutorial). Since i am creating the index in c# i need to specify the KnnVector in the index creator and the field that is associated with it? Which i thought should be this:

.KnnVector(kv => kv.Name(n => n.passage_embedding).Dimension(768).ModelId("").Method(m => m.Engine("lucene").SpaceType("l2").Name("hnsw")))

Then i should just fill the test field and it should work but nope.

Thanks for you help by the way.

[Xtansia]

I haven't actually run this code yet but the below should be roughly what's needed. I'm going to work on creating a full working sample.

You document class would look something like:

public class NlpDoc
{
    public NlpDoc()
    {
    }

    public NlpDoc(string id, string text)
    {
        Id = id;
        Text = text;
    }

    public string Id { get; set; }
    public string Text { get; set; }
    [PropertyName("passage_embedding")]
    public float[] PassageEmbedding { get; set; }
}

Creating the index would look something like:

var resp = await client.Indices.CreateAsync(
    indexName,
    i => i
        .Settings(s => s
            .Setting("index.knn", true)
            .DefaultPipeline(pipelineName))
        .Map<NlpDoc>(m => m
            .Properties(p => p
                .Text(t => t.Name(d => d.Id))
                .KnnVector(k => k
                    .Name(d => d.PassageEmbedding)
                    .Dimension(768)
                    .Method(km => km
                        .Engine("lucene")
                        .SpaceType("l2")
                        .Name("hnsw")))
                .Text(t => t.Name(d => d.Text)))));

Indexing the documents would look like:

var docs = new[]
{
    new NlpDoc("4319130149.jpg", "A West Virginia university women 's basketball team , officials , and a small gathering of fans are in a West Virginia arena ."),
    new NlpDoc("1775029934.jpg", "A wild animal races across an uncut field with a minimal amount of trees ."),
    new NlpDoc("2664027527.jpg", "People line the stands which advertise Freemont 's orthopedics , a cowboy rides a light brown bucking bronco ."),
    new NlpDoc("4427058951.jpg", "A man who is riding a wild horse in the rodeo is very near to falling off ."),
    new NlpDoc("2691147709.jpg", "A rodeo cowboy , wearing a cowboy hat , is being thrown off of a wild white horse .")
};
var resp = await client.IndexManyAsync(docs, indexName);

[whittssg]

I will give it a go tomorrow, a sample would be awsome. Thanks again.

[Xtansia]

@whittssg I've created a working sample in #614 and confirmed this isn't a bug in the client. I'm going to close this issue in favour of #372 for adding proper support and typings for the feature.