Unable to perform hybrid search

Elastic Search
1

I am having trouble doing hybrid search. to give you guys bit of background I am using semantic search created inference with openai embeddings.

const addIndex = async () => {
  // try {
  //   await client.indices.delete({ index });
  // } catch (err) {
  //   //
  // }
  await client.indices.create({
    index,
    mappings: {
      properties: {
        phrase: {
          type: "semantic_text",
          inference_id: "openai_embeddings",
        },
        copilot: {
          type: "keyword",
        },
        org: {
          type: "keyword",
        },
        topic: {
          type: "keyword",
        },
        mode: {
          type: "keyword",
        },
      },
    },
  });
  console.log("created successfully");
};

const response = await client.transport.request({
    method: "PUT",
    path: "/_inference/text_embedding/openai_embeddings",
    body: {
      service: "openai",
      service_settings: {
        api_key: "xxxxxxxx-xxxxxxxx",
        model_id: "text-embedding-3-small",
      },
    },
  });

this is my search query

const response = await client.search({
    index,
    min_score: 0.8,
    query: {
      bool: {
        must: {
          semantic: {
            field: "phrase",
            query: "can i talk to a real person",
          },
        },
        filter: [
          { term: { copilot: "67108e2e8c0611df3ad71109" } }, // Filter by copilotId
          { term: { mode: "DRAFT" } }, // Filter by mode
        ],
      },
    },
  });

above query randomly works sometimes and doesnt work other times.
i make page to 1000 then it started giving results. why is pagesize altering the behaviour? Is it performing post filter? so that means it doesnt work if i have 2000 phrases? Please help me here.

2

Hi @harish32, Welcome to the Elastic community.

The page should not impact to the response.

  1. You defining 1000 in size?
  2. Any error you getting when there is no response?
  3. I am also assuming Async is well handled in your script.
3

no its returning empty array
here is the query i have ran

  const response = await client.search({
    index,
    query: {
      bool: {
        must: [
          {
            semantic: {
              field: "phrase",
              query: "talk to a agent",
            },
          },
        ],
        filter: [
          { term: { copilot: "67108e2e8c0611df3ad71109" } }, // Filter by copilotId
          { term: { mode: "DRAFT" } }, // Filter by mode
          { term: { topic: "675a88feb73a5d37e21fc352" } },
        ],
      },
    },
  });
  console.log(response.hits);

i have attached the screenshot where i ran same function 3 times you can see it gave result not completely but returned only one then
3rd time it didnt give anything

Screenshot from 2024-12-26 16-22-16
Screenshot from 2024-12-26 16-22-161275×1031 90.3 KB

4

I would

  1. check the Elasticsearch logs (especially for any errors around inference or search)
  2. log the actual query run to compare it to the successful ones.

If I had to guess, I'd say turning the semantic query into its dense vector representation fails, so you're not running the same Elasticsearch query in the background. Maybe something like a rate limit is kicking in (but that's just a random guess). I'd be surprised if you ran the exact same Elasticsearch query and it would give you different results (assuming the cluster is healthy).

1 Like
5

Another helpful troubleshooting tool could be to run the knn search directly on the semantic_text field a a nested query. See example.

6

I tried knn as well still not getting results

const response = await client.search({
    index,
    query: {
      bool: {
        must: [
          {
            nested: {
              path: "phrase.inference.chunks",
              query: {
                knn: {
                  field: "phrase.inference.chunks.embeddings",
                  query_vector_builder: {
                    text_embedding: {
                      model_id: "openai_embeddings",
                      model_text: "talk to an agent",
                    },
                  },
                },
              },
            },
          },
        ],
        filter: [
          { term: { copilot: "67108e2e8c0611df3ad71109" } }, // Filter by copilotId
          { term: { mode: "DRAFT" } }, // Filter by mode
          { term: { topic: "675a88feb73a5d37e21fc352" } },
        ],
      },
    },

If i remove filter it works fine but i need both to work

7

Thanks for the reply i am new to Elasticsearch. i tried checking logs i cant find anything.

Screenshot from 2025-01-07 15-58-39
Screenshot from 2025-01-07 15-58-393323×1256 228 KB

what i believe is happening it is doing a filter first and then searching on the 10 results it filtered. That is why when i increase the page size it is going to return 100 which will contain the semantic word so it is returning matching 10.

I have followed this

Do you have any reference for hybrid search like mine which would be very helpful.

8

Can you try adding the filters in the knn filter section, to pre filter the results returned by knn instead of post filtering?

9

Thank you so much @Kathleen_DeRusso :pray:. It is working now.
Thanks @xeraa for jumping in here

1 Like