Failed to create query: the query vector has a different dimension [384] than the index vectors [768]

Elastic Stack Elasticsearch
1

How does one set the dimension in the query to avoid this message?

2

while setting up index mapping, you will have a "dims" field along with type: "dense_vector".

set it to "dims": 768

3

Yes, I have that in the index mappings, but where is the query getting the 384 value -- how would I set that?

    "path": {
      "type": "text",
      "term_vector": "with_positions_offsets",
      "analyzer": "text_keep_stopwords",
      "search_analyzer": "text_drop_stopwords",
      "index_options": "offsets",
      "store": "true"
    },
    "path_embedding.predicted_value": {
      "type": "dense_vector",
      "dims": 768,
      "index": true,
      "similarity": "cosine"
    },
    "passage": {
      "type": "text",
      "analyzer": "text_keep_stopwords",
      "search_analyzer": "text_drop_stopwords",
      "term_vector": "with_positions_offsets",
      "index_options": "offsets",
      "store": "true"
    },
    "passage_embedding.predicted_value": {
      "type": "dense_vector",
      "dims": 768,
      "index": true,
      "similarity": "cosine"
    },
4

Can you share your insert script (only vector key value)

5

I'm using a bulk ingester:

        BulkIngester<ObjectNode> ingester = BulkIngester
            .of(b -> b.client(_client)
                .maxOperations(100).flushInterval(20L, TimeUnit.SECONDS)
                .listener(ESUtils.getBulkListener(_logger)));

With the pipeline defined like:

        client.ingest().putPipeline(pr -> pr.id("rdp_pipeline")
            .processors(p -> p
                .inference(i -> i.modelId(".multilingual-e5-small_linux-x86_64")
                    .fieldMap("passage", JsonData.of("text_field"))
                    .targetField("passage_embedding")))
            .processors(p -> p
                .inference(i -> i.modelId(".multilingual-e5-small_linux-x86_64")
                    .fieldMap("path", JsonData.of("text_field"))
                    .targetField("path_embedding")))

        );

And I build an ObjectNode and send that for ingestion:

        String path = passage.get("path").asText().replace("~", " ~ ");
        int linenum = passage.get("lineNum").intValue();
        String text = passage.get("text").asText();
        ObjectNode content = s_objectMapper.createObjectNode();
        content.put("passage", text);
        content.put("path", path);
        content.put("url", url);
        content.put("title", title);
        content.put("linenum", linenum);
        content.put("filestem", filestem);
        content.put("id", id);
        content.put("language", langFamily);
        ingester
            .add(op -> op.index(idx -> idx.index(indexName).document(content)));
6

I think the model used produces a output embedding of size 384 (intfloat/multilingual-e5-small) whereas you have configured 768 in index mapping

7

You are absolutely correct. I didn't realize this limitation. Thank you.