i created a new project and just crawled this website: https://www.tagesschau.de/inland/innenpolitik/spd-miersch-generalsekretaer-100.html

It looks very suspicious to me... The similarity score between ANY image and ANY text is never higher than any text to another text. (commonly known as modality gap).

Might it be that this is a displaying bug or that some sdoc IDs got mixed up? Have you checked the raw output in the SwaggerUI ?

calling the endponit

{
  "proj_id": 1,
  "query": "Kanzler",
  "top_k": 100,
  "threshold": 0,
  "filter": {
    "id": "string",
    "items": [],
    "logic_operator": "or"
  }
}

returns

[
  {
    "sdoc_id": 2,
    "score": 0.9917974472045898,
    "sentence_id": 201
  },
  {
    "sdoc_id": 1,
    "score": 0.9906548261642456,
    "sentence_id": 7
  },
  {
    "sdoc_id": 1,
    "score": 0.9906548261642456,
    "sentence_id": 14
  },
  {
    "sdoc_id": 1,
    "score": 0.9906548261642456,
    "sentence_id": 28
  },
  {
    "sdoc_id": 2,
    "score": 0.9906548261642456,
    "sentence_id": 108
  },
  ....

sdoc_id 2 is an image.

I mean even images have text documents (the automatically generated caption) however, this caption should only be returned once, as it most of the time only consists of one sentence. maybe there a bug related to embedding the caption?