Multi-vector search

In some cases, a single vector search may not be sufficient to retrieve the most relevant results.

In TopK, you can run multiple vector searches in a single query. Additionally, you can apply a different ranking score to each individual vector search result.

For example, in a research paper database, you might want to:

• Retrieve documents based on both the summary of the entire paper and summaries of individual paragraphs.
• Rank results by combining scores from multiple embeddings such as:
  • 70% weight for entire paper similarity
  • 30% weight for paragraph summary similarity

Let’s define a collection with two semantic indexes:

Python

from topk_sdk.schema import text, vector, semantic_index

client.collections().create(
    "articles",
    schema={
        "title": text().required(),
        "paper_summary": text().index(semantic_index()),  # Embedding of full paper
        "paragraph_summary": text().index(semantic_index()),  # Embedding of a paragraph
    },
)

After, we run a multi-vector search by using the fn.semantic_similarity() function for calculating semantic similarity on the paper_score field as well as the paragraph_score field:

Python

from topk_sdk.query import select, fn

docs = client.collection("articles").query(
    select(
        "title",
        paper_score=fn.semantic_similarity("paper_summary", "deep learning optimization"),
        paragraph_score=fn.semantic_similarity("paragraph_summary", "stochastic gradient descent")
    )
    .topk(field("paper_score") * 0.7 + field("paragraph_score") * 0.3, 10)
)

# Example results:

[
  {
    "_id": "2",
    "title": "On the Importance of Initialization and Momentum in Deep Learning",
    "paper_score": 0.9774298071861267,
    "paragraph_score": 0.9783554673194885,
    # "blend_score": 0.9777075052261353,
  },
  {
    "_id": "1",
    "title": "Understanding the Difficulty of Training Deep Feedforward Neural Networks",
    "paper_score": 0.9773390889167786,
    "paragraph_score": 0.976479709148407,
    # "blend_score": 0.977081298828125,
  }
]

Let’s break down the example above:

1. Calculate semantic similarity between the query and the full paper summary as well as the paragraph summary.
2. Blend the two scores by giving 70% weight to the full paper and 30% weight to the paragraph summary.
3. Retrieve the top 10 results ranked by the blended score.