Querying vectors

Vector Databases for Embeddings with Pinecone

James Chapman

Curriculum Manager, DataCamp

The power of querying

Querying: receive the most semantically similar vectors to an input vector

A user input being embedded and used to query a Pinecone vector database.

The .query() method

index.query(

  vector=[-0.250919762305275, ...],

  top_k=3

)

{'matches': [{'id': '1', 'score': 0.0478537641, 'values': []},
             {'id': '2', 'score': 0.046000585, 'values': []},
             {'id': '3', 'score': 0.0458319113, 'values': []}],
 'namespace': '',
 'usage': {'read_units': 5}}

The .query() method

index.query(
    vector=[-0.250919762305275, ...],
    top_k=3,

    include_values=True

)

{'matches': [{'id': '1', 'score': 0.0478537641, 'values': [-0.0131468913, ...]},
             {'id': '2', 'score': 0.046000585, 'values': [-0.0120476764, ...]},
             {'id': '3', 'score': 0.0458319113, 'values': [0.00285418332, ...]}],
 'namespace': '',
 'usage': {'read_units': 5}}

Read units (RUs) for querying

For querying, RUs is harder to calculate
Dependent on:
- No. of records in the namespace
- Size of records
  - Vector dimensionality
  - Amount of metadata

Table of RUs consumed when querying.

¹ https://docs.pinecone.io/guides/organizations/manage-cost/understanding-cost#query

Distance metrics

The cosine similarity: showing an angle between two vectors.

¹ https://docs.pinecone.io/guides/indexes/understanding-indexes#distance-metrics

Distance metrics

The euclidean distance: showing the straight-line distance between two vectors.

¹ https://docs.pinecone.io/guides/indexes/understanding-indexes#distance-metrics

Distance metrics

The dot product: how the value is calculated from the angle between the vectors.

¹ https://docs.pinecone.io/guides/indexes/understanding-indexes#distance-metrics

Setting the distance metric

pc.create_index(
    name="datacamp-index", 
    dimension=1536,

    metric='dotproduct',

    spec=ServerlessSpec(
        cloud='aws', 
        region='us-east-1'
    )
)

metric → 'cosine', 'euclidean', 'dotproduct'

Your turn to query!

Vector Databases for Embeddings with Pinecone