[Bug]: wrong retriever from version 0.6.0

What happened?

Hi, I implemented a prefech retriever.
I am using the model:
SparseTextEmbedding("Qdrant/bm25")
TextEmbedding("sentence-transformers/paraphrase-multilingual-mpnet-base-v2")
LateInteractionTextEmbedding("answerdotai/answerai-colbert-small-v1")

I tried the query: "Errore 0x000006D9"

This is the abnormal behaviour:
with fastembed version 0.5.1 it retrieves the right qdrant docuemnt
From fastembed 0.6.0 to 0.7.1 it retrieves the wrong qdrant document

instead, if the query is: "Erro 0x000006D9" or "Erro 0x000006D9" it works fine.

This is my function;

def retrieve_with_prf(query, limit=5, k_feedback=3):
    try:
        sparse_vec = list(sparse_embedding_model.embed([query]))[0]
        initial_results = client.query_points(
            collection_name=collection_name,
            query=models.SparseVector(**sparse_vec.as_object()),
            using=sparse_vector_name,
            with_payload=True,
            limit=k_feedback
        )

        if not initial_results or not initial_results.points:
            return []

        feedback_texts = [
            p.payload.get('text', '')[:1000]
            for p in initial_results.points
            if p.payload.get('text')
        ]
        if not feedback_texts:
            return []

        feedback_embeddings = list(dense_embedding_model.embed(feedback_texts))
        avg_vec = np.mean(feedback_embeddings, axis=0).tolist()

        return client.query_points(
            collection_name=collection_name,
            query=avg_vec,
            using=dense_vector_name,
            with_payload=True,
            limit=limit
        )
    except Exception:
        return []

What is the expected behaviour?

No response

A minimal reproducible example

No response

What Python version are you on? e.g. python --version

python 3.13

FastEmbed version

v0.7.1

What os are you seeing the problem on?

No response

Relevant stack traces and/or logs

Start with retrieve_with_prf and reproduce the exact "Errore 0x000006D9" query using the listed embedding models, comparing fastembed 0.5.1 with 0.6.0-0.7.1. Confirm which retrieval stage changes and use the working shorter query as a comparison; done means the exact query retrieves the expected Qdrant document again, with a regression test if the repository exposes one.