.near_text vector search score is very low

Support
1

I am creating a schema as below

image
image738×566 25.1 KB

I’ve added the data to the collection, but when I query it for the nearest results to the test term “Art Briefs” (which is an exact match in the collection), I’m getting unexpectedly low scores. Here’s an example:
For search of

org_collection = client.collections.get(“Org_Test”)
response = org_collection.query.near_text(
query=“Art Briefs”,
limit=5,
return_metadata=MetadataQuery.full()
)
for o in response.objects:
print(o.properties[‘skill_name’])
cosine_similarity = 2 * o.metadata.certainty - 1
print(f"Cosine similarity: {cosine_similarity}“)
print(f"Certainty: {o.metadata.certainty}”)

I get
{Art Briefs
Cosine similarity: 0.8635821342468262
Certainty: 0.9317910671234131}

I’ve set vectorize_property_name=True for the skill name field only. This low score is causing issues with my API, and I need the score to be much higher. I’ve searched extensively but I’m stuck. Please provide a resolution as soon as possible.

2

hi @Rishi_Prakash !!

Welcome to our community :hugs:

the near_text (similarity/vector search) will not be a literal search. So even if you have a exact match, it may not be placed near to your query.

For that, the bm25/keyword and hybrid search will be more effective.

How far on the ranking the exact match object from the query is?

3

Hi Duda,

Thanks for the response,
Could you update the below code to use bm25/keyword and hybrid search.

org_collection = client.collections.get(“Org_Test”)
response = org_collection.query.near_text(
query=“Art Briefs”,
limit=5,
return_metadata=MetadataQuery.full()
)
for o in response.objects:
print(o.properties[‘skill_name’])
cosine_similarity = 2 * o.metadata.certainty - 1
print(f"Cosine similarity: {cosine_similarity}“)
print(f"Certainty: {o.metadata.certainty}”)

Because I don’t see the V4 code in documentation.

Thanks!

4

Hi!

it is here:

Note that you get distance with a vector search, but for hybrid and bm25/keyword, you get score.

here is the docs for bm25:

Your code will be need to change to:

response = org_collection.query.hybrid(
    query="Art Briefs",
    limit=5,
    return_metadata=MetadataQuery.full()
)