Python interface for pgvector and pgvectorscale

Working with pgvectorscale and pgvector in python

You use pgai to power production grade AI applications. timescale_vector is the Python interface you use to interact with a pgai on programmatically.

Before you get started with timescale_vector:

• Sign up for pgai on : Get 90 days free to try pgai on .
• Follow the Get Started Tutorial: Learn how to use pgai on for semantic search on a real-world dataset.

If you prefer to use an LLM development or data framework, see pgai’s integrations with LangChain and LlamaIndex.

Prerequisites

timescale_vector depends on the source distribution of psycopg2 and adheres to best practices for psycopg2.

Before you install timescale_vector:

• Follow the psycopg2 build prerequisites.

Install

To interact with pgai on using Python:

1. Install timescale_vector:

   pip install timescale_vector
   

2. Install dotenv:

   pip install python-dotenv
   

   In these examples, you use dotenv to pass secrets and keys.

That is it, you are ready to go.

Basic usage of the timescale_vector library

First, import all the necessary libraries:

from dotenv import load_dotenv, find_dotenv
import os
from timescale_vector import client
import uuid
from datetime import datetime, timedelta

Load up your credentials, the safest way is with a .env file:

_ = load_dotenv(find_dotenv(), override=True)
service_url  = os.environ['TIMESCALE_SERVICE_URL']

Next, create the client. This tutorial, uses the sync client. But the library has an async client as well (with an identical interface that uses async functions).

The client constructor takes three required arguments:

name	description
service_url	URL / connection string
table_name	Name of the table to use for storing the embeddings. Think of this as the collection name
num_dimensions	Number of dimensions in the vector

vec  = client.Sync(service_url, "my_data", 2)

Next, create the tables for the collection:

vec.create_tables()

Next, insert some data. The data record contains:

• A UUID to uniquely identify the embedding
• A JSON blob of metadata about the embedding
• The text the embedding represents
• The embedding itself

Because this data includes UUIDs which become primary keys, upserts should be used for ingest.

vec.upsert([\
    (uuid.uuid1(), {"animal": "fox"}, "the brown fox", [1.0,1.3]),\
    (uuid.uuid1(), {"animal": "fox", "action":"jump"}, "jumped over the", [1.0,10.8]),\
])

You can now create a vector index to speed up similarity search:

vec.create_embedding_index(client.TimescaleVectorIndex())

Then, you can query for similar items:

vec.search([1.0, 9.0])

[[UUID('73d05df0-84c1-11ee-98da-6ee10b77fd08'),
  {'action': 'jump', 'animal': 'fox'},
  'jumped over the',
  array([ 1. , 10.8], dtype=float32),
  0.00016793422934946456],
 [UUID('73d05d6e-84c1-11ee-98da-6ee10b77fd08'),
  {'animal': 'fox'},
  'the brown fox',
  array([1. , 1.3], dtype=float32),
  0.14489260377438218]]

There are many search options which are covered below in the Advanced search section.

A simple search example that returns one item using a similarity search constrained by a metadata filter is shown below:

vec.search([1.0, 9.0], limit=1, filter={"action": "jump"})

[[UUID('73d05df0-84c1-11ee-98da-6ee10b77fd08'),
  {'action': 'jump', 'animal': 'fox'},
  'jumped over the',
  array([ 1. , 10.8], dtype=float32),
  0.00016793422934946456]]

The returned records contain 5 fields:

name	description
id	The UUID of the record
metadata	The JSON metadata associated with the record
contents	the text content that was embedded
embedding	The vector embedding
distance	The distance between the query embedding and the vector

You can access the fields by simply using the record as a dictionary keyed on the field name:

records = vec.search([1.0, 9.0], limit=1, filter={"action": "jump"})
(records[0]["id"],records[0]["metadata"], records[0]["contents"], records[0]["embedding"], records[0]["distance"])

(UUID('73d05df0-84c1-11ee-98da-6ee10b77fd08'),
 {'action': 'jump', 'animal': 'fox'},
 'jumped over the',
 array([ 1. , 10.8], dtype=float32),
 0.00016793422934946456)

You can delete by ID:

vec.delete_by_ids([records[0]["id"]])

Or you can delete by metadata filters:

vec.delete_by_metadata({"action": "jump"})

To delete all records use:

vec.delete_all()

Advanced usage

This section goes into more detail about the Python interface. It covers:

1. Search filter options - how to narrow your search by additional constraints
2. Indexing - how to speed up your similarity queries
3. Time-based partitioning - how to optimize similarity queries that filter on time
4. Setting different distance types to use in distance calculations

Search options

The search function is very versatile and allows you to search for the right vector in a wide variety of ways. This section describes the search option in 3 parts:

1. Basic similarity search.
2. How to filter your search based on the associated metadata.
3. Filtering on time when time-partitioning is enabled.

The following examples are based on this data:

vec.upsert([\
    (uuid.uuid1(), {"animal":"fox", "action": "sit", "times":1}, "the brown fox", [1.0,1.3]),\
    (uuid.uuid1(),  {"animal":"fox", "action": "jump", "times":100}, "jumped over the", [1.0,10.8]),\
])

The basic query looks like this:

vec.search([1.0, 9.0])

[[UUID('7487af96-84c1-11ee-98da-6ee10b77fd08'),
  {'times': 100, 'action': 'jump', 'animal': 'fox'},
  'jumped over the',
  array([ 1. , 10.8], dtype=float32),
  0.00016793422934946456],
 [UUID('7487af14-84c1-11ee-98da-6ee10b77fd08'),
  {'times': 1, 'action': 'sit', 'animal': 'fox'},
  'the brown fox',
  array([1. , 1.3], dtype=float32),
  0.14489260377438218]]

You could provide a limit for the number of items returned:

vec.search([1.0, 9.0], limit=1)

[[UUID('7487af96-84c1-11ee-98da-6ee10b77fd08'),
  {'times': 100, 'action': 'jump', 'animal': 'fox'},
  'jumped over the',
  array([ 1. , 10.8], dtype=float32),
  0.00016793422934946456]]

Narrowing your search by metadata

There are two main ways to filter results by metadata:

• filters for equality matches on metadata.
• predicates for complex conditions on metadata.

Filters are more limited in what they can express, but are also more performant. You should use filters if your use case allows it.

Using filters for equality matches

You could specify a match on the metadata as a dictionary where all keys have to match the provided values (keys not in the filter are unconstrained):

vec.search([1.0, 9.0], limit=1, filter={"action": "sit"})

[[UUID('7487af14-84c1-11ee-98da-6ee10b77fd08'),
  {'times': 1, 'action': 'sit', 'animal': 'fox'},
  'the brown fox',
  array([1. , 1.3], dtype=float32),
  0.14489260377438218]]

You can also specify a list of filter dictionaries, where an item is returned if it matches any dict:

vec.search([1.0, 9.0], limit=2, filter=[{"action": "jump"}, {"animal": "fox"}])

[[UUID('7487af96-84c1-11ee-98da-6ee10b77fd08'),
  {'times': 100, 'action': 'jump', 'animal': 'fox'},
  'jumped over the',
  array([ 1. , 10.8], dtype=float32),
  0.00016793422934946456],
 [UUID('7487af14-84c1-11ee-98da-6ee10b77fd08'),
  {'times': 1, 'action': 'sit', 'animal': 'fox'},
  'the brown fox',
  array([1. , 1.3], dtype=float32),
  0.14489260377438218]]

Using predicates for more advanced filtering on metadata

Predicates allow for more complex search conditions. For example, you could use greater than and less than conditions on numeric values.

vec.search([1.0, 9.0], limit=2, predicates=client.Predicates("times", ">", 1))

[[UUID('7487af96-84c1-11ee-98da-6ee10b77fd08'),
  {'times': 100, 'action': 'jump', 'animal': 'fox'},
  'jumped over the',
  array([ 1. , 10.8], dtype=float32),
  0.00016793422934946456]]

Predicates objects are defined by the name of the metadata key, an operator, and a value.

The supported operators are: ==, !=, <, <=, >, >=

The type of the values determines the type of comparison to perform. For example, passing in "Sam" (a string) performs a string comparison while a 10 (an int) performs an integer comparison, and a 10.0 (float) performs a float comparison. It is important to note that using a value of "10" performs a string comparison as well so it’s important to use the right type. Supported Python types are: str, int, and float.

One more example with a string comparison:

vec.search([1.0, 9.0], limit=2, predicates=client.Predicates("action", "==", "jump"))

[[UUID('7487af96-84c1-11ee-98da-6ee10b77fd08'),
  {'times': 100, 'action': 'jump', 'animal': 'fox'},
  'jumped over the',
  array([ 1. , 10.8], dtype=float32),
  0.00016793422934946456]]

The real power of predicates is that they can also be combined using the & operator (for combining predicates with AND semantics) and |(for combining using OR semantic). So you can do:

vec.search([1.0, 9.0], limit=2, predicates=client.Predicates("action", "==", "jump") & client.Predicates("times", ">", 1))

[[UUID('7487af96-84c1-11ee-98da-6ee10b77fd08'),
  {'times': 100, 'action': 'jump', 'animal': 'fox'},
  'jumped over the',
  array([ 1. , 10.8], dtype=float32),
  0.00016793422934946456]]

Just for sanity, the next example shows a case where no results are returned because of predicates:

vec.search([1.0, 9.0], limit=2, predicates=client.Predicates("action", "==", "jump") & client.Predicates("times", "==", 1))

[]

And one more example where the predicates are defined as a variable and use grouping with parenthesis:

my_predicates = client.Predicates("action", "==", "jump") & (client.Predicates("times", "==", 1) | client.Predicates("times", ">", 1))
vec.search([1.0, 9.0], limit=2, predicates=my_predicates)

[[UUID('7487af96-84c1-11ee-98da-6ee10b77fd08'),
  {'times': 100, 'action': 'jump', 'animal': 'fox'},
  'jumped over the',
  array([ 1. , 10.8], dtype=float32),
  0.00016793422934946456]]

There is also semantic sugar for combining many predicates with AND semantics. You can pass in multiple 3-tuples to Predicates:

vec.search([1.0, 9.0], limit=2, predicates=client.Predicates(("action", "==", "jump"), ("times", ">", 10)))

[[UUID('7487af96-84c1-11ee-98da-6ee10b77fd08'),
  {'times': 100, 'action': 'jump', 'animal': 'fox'},
  'jumped over the',
  array([ 1. , 10.8], dtype=float32),
  0.00016793422934946456]]

Filter your search by time

When using time-partitioning (see below) you can very efficiently filter your search by time. Time-partitioning associates the timestamp embedded in a UUID-based ID with an embedding. First, create a collection with time partitioning and insert some data (one item from January 2018 and another in January 2019):

tpvec = client.Sync(service_url, "time_partitioned_table", 2, time_partition_interval=timedelta(hours=6))
tpvec.create_tables()

specific_datetime = datetime(2018, 1, 1, 12, 0, 0)
tpvec.upsert([\
    (client.uuid_from_time(specific_datetime), {"animal":"fox", "action": "sit", "times":1}, "the brown fox", [1.0,1.3]),\
    (client.uuid_from_time(specific_datetime+timedelta(days=365)),  {"animal":"fox", "action": "jump", "times":100}, "jumped over the", [1.0,10.8]),\
])

Then, you can filter using the timestamps by specifying a uuid_time_filter:

tpvec.search([1.0, 9.0], limit=4, uuid_time_filter=client.UUIDTimeRange(specific_datetime, specific_datetime+timedelta(days=1)))

[[UUID('33c52800-ef15-11e7-be03-4f1f9a1bde5a'),
  {'times': 1, 'action': 'sit', 'animal': 'fox'},
  'the brown fox',
  array([1. , 1.3], dtype=float32),
  0.14489260377438218]]

A UUIDTimeRange can specify a start_date or end_date or both(as in the example above). Specifying only the start_date or end_date leaves the other end unconstrained.

tpvec.search([1.0, 9.0], limit=4, uuid_time_filter=client.UUIDTimeRange(start_date=specific_datetime))

[[UUID('ac8be800-0de6-11e9-889a-5eec84ba8a7b'),
  {'times': 100, 'action': 'jump', 'animal': 'fox'},
  'jumped over the',
  array([ 1. , 10.8], dtype=float32),
  0.00016793422934946456],
 [UUID('33c52800-ef15-11e7-be03-4f1f9a1bde5a'),
  {'times': 1, 'action': 'sit', 'animal': 'fox'},
  'the brown fox',
  array([1. , 1.3], dtype=float32),
  0.14489260377438218]]

You have the option to define whether the start and end dates are inclusive with the start_inclusive and end_inclusive parameters. Setting start_inclusive to true results in comparisons using the >= operator, whereas setting it to false applies the > operator. By default, the start date is inclusive, while the end date is exclusive. One example:

tpvec.search([1.0, 9.0], limit=4, uuid_time_filter=client.UUIDTimeRange(start_date=specific_datetime, start_inclusive=False))

[[UUID('ac8be800-0de6-11e9-889a-5eec84ba8a7b'),
  {'times': 100, 'action': 'jump', 'animal': 'fox'},
  'jumped over the',
  array([ 1. , 10.8], dtype=float32),
  0.00016793422934946456]]

Notice how the results are different when using the start_inclusive=False option because the first row has the exact timestamp specified by start_date.

It is also easy to integrate time filters using the filter and predicates parameters described above using special reserved key names to make it appear that the timestamps are part of your metadata. This is useful when integrating with other systems that just want to specify a set of filters (often these are “auto retriever” type systems). The reserved key names are __start_date and __end_date for filters and __uuid_timestamp for predicates. Some examples below:

tpvec.search([1.0, 9.0], limit=4, filter={ "__start_date": specific_datetime, "__end_date": specific_datetime+timedelta(days=1)})

[[UUID('33c52800-ef15-11e7-be03-4f1f9a1bde5a'),
  {'times': 1, 'action': 'sit', 'animal': 'fox'},
  'the brown fox',
  array([1. , 1.3], dtype=float32),
  0.14489260377438218]]

tpvec.search([1.0, 9.0], limit=4,
             predicates=client.Predicates("__uuid_timestamp", ">", specific_datetime) & client.Predicates("__uuid_timestamp", "<", specific_datetime+timedelta(days=1)))

[[UUID('33c52800-ef15-11e7-be03-4f1f9a1bde5a'),
  {'times': 1, 'action': 'sit', 'animal': 'fox'},
  'the brown fox',
  array([1. , 1.3], dtype=float32),
  0.14489260377438218]]

Indexing

Indexing speeds up queries over your data. By default, the system creates indexes to query your data by the UUID and the metadata.

To speed up similarity search based on the embeddings, you have to create additional indexes.

Note that if performing a query without an index, you always get an exact result, but the query is slow (it has to read all of the data you store for every query). With an index, your queries are order-of-magnitude faster, but the results are approximate (because there are no known indexing techniques that are exact).

Luckily, provides 3 excellent approximate indexing algorithms, StreamingDiskANN, HNSW, and ivfflat.

Below are the trade-offs between these algorithms:

Algorithm	Build speed	Query speed	Need to rebuild after updates
StreamingDiskAnn	Fast	Fastest	No
HNSW	Fast	Faster	No
ivfflat	Fastest	Slowest	Yes

You can see benchmarks on the blog.

You should use the StreamingDiskANN index for most use cases. This can be created with:

vec.create_embedding_index(client.TimescaleVectorIndex())

Indexes are created for a particular distance metric type. So it is important that the same distance metric is set on the client during index creation as it is during queries. See the distance type section below.

Each of these indexes has a set of build-time options for controlling the speed/accuracy trade-off when creating the index and an additional query-time option for controlling accuracy during a particular query. The library uses smart defaults for all of these options. The details for how to adjust these options manually are below.

StreamingDiskANN index

The StreamingDiskANN index is a graph-based algorithm that uses the DiskANN algorithm. You can read more about it in the blog announcing its release.

To create this index, run:

vec.create_embedding_index(client.TimescaleVectorIndex())

The above command creates the index using smart defaults. There are a number of parameters you could tune to adjust the accuracy/speed trade-off.

The parameters you can set at index build time are:

Parameter name	Description	Default value
num_neighbors	Sets the maximum number of neighbors per node. Higher values increase accuracy but make the graph traversal slower.	50
search_list_size	This is the S parameter used in the greedy search algorithm used during construction. Higher values improve graph quality at the cost of slower index builds.	100
max_alpha	Is the alpha parameter in the algorithm. Higher values improve graph quality at the cost of slower index builds.	1.0

To set these parameters, you could run:

vec.create_embedding_index(client.TimescaleVectorIndex(num_neighbors=50, search_list_size=100, max_alpha=1.0))

You can also set a parameter to control the accuracy vs. query speed trade-off at query time. The parameter is set in the search() function using the query_params argument. You can set the search_list_size(default: 100). This is the number of additional candidates considered during the graph search at query time. Higher values improve query accuracy while making the query slower.

You can specify this value during search as follows:

vec.search([1.0, 9.0], limit=4, query_params=TimescaleVectorIndexParams(search_list_size=10))

To drop the index, run:

vec.drop_embedding_index()

pgvector HNSW index

Pgvector provides a graph-based indexing algorithm based on the popular HNSW algorithm.

To create this index, run:

vec.create_embedding_index(client.HNSWIndex())

The above command creates the index using smart defaults. There are a number of parameters you could tune to adjust the accuracy/speed trade-off.

The parameters you can set at index build time are:

Parameter name	Description	Default value
m	Represents the maximum number of connections per layer. Think of these connections as edges created for each node during graph construction. Increasing m increases accuracy but also increases index build time and size.	16
ef_construction	Represents the size of the dynamic candidate list for constructing the graph. It influences the trade-off between index quality and construction speed. Increasing ef_construction enables more accurate search results at the expense of lengthier index build times.	64

To set these parameters, you could run:

vec.create_embedding_index(client.HNSWIndex(m=16, ef_construction=64))

You can also set a parameter to control the accuracy vs. query speed trade-off at query time. The parameter is set in the search() function using the query_params argument. You can set the ef_search(default: 40). This parameter specifies the size of the dynamic candidate list used during search. Higher values improve query accuracy while making the query slower.

You can specify this value during search as follows:

vec.search([1.0, 9.0], limit=4, query_params=HNSWIndexParams(ef_search=10))

To drop the index run:

vec.drop_embedding_index()

pgvector ivfflat index

Pgvector provides a clustering-based indexing algorithm. The blog post describes how it works in detail. It provides the fastest index-build speed but the slowest query speeds of any indexing algorithm.

To create this index, run:

vec.create_embedding_index(client.IvfflatIndex())

Note: ivfflat should never be created on empty tables because it needs to cluster data, and that only happens when an index is first created, not when new rows are inserted or modified. Also, if your table undergoes a lot of modifications, you need to rebuild this index occasionally to maintain good accuracy. See the blog post for details.

Pgvector ivfflat has a lists index parameter that is automatically set with a smart default based on the number of rows in your table. If you know that you’ll have a different table size, you can specify the number of records to use for calculating the lists parameter as follows:

vec.create_embedding_index(client.IvfflatIndex(num_records=1000000))

You can also set the lists parameter directly:

vec.create_embedding_index(client.IvfflatIndex(num_lists=100))

You can also set a parameter to control the accuracy vs. query speed trade-off at query time. The parameter is set in the search() function using the query_params argument. You can set the probes. This parameter specifies the number of clusters searched during a query. It is recommended to set this parameter to sqrt(lists) where lists is the num_list parameter used above during index creation. Higher values improve query accuracy while making the query slower.

You can specify this value during search as follows:

vec.search([1.0, 9.0], limit=4, query_params=IvfflatIndexParams(probes=10))

To drop the index, run:

vec.drop_embedding_index()

Time partitioning

In many use cases where you have many embeddings, time is an important component associated with the embeddings. For example, when embedding news stories, you often search by time as well as similarity (for example, stories related to Bitcoin in the past week or stories about Clinton in November 2016).

Yet, traditionally, searching by two components “similarity” and “time” is challenging for Approximate Nearest Neighbor (ANN) indexes and makes the similarity-search index less effective.

One approach to solving this is partitioning the data by time and creating ANN indexes on each partition individually. Then, during search, you can:

• Step 1: filter partitions that don’t match the time predicate.
• Step 2: perform the similarity search on all matching partitions.
• Step 3: combine all the results from each partition in step 2, re-rank, and filter out results by time.

Step 1 makes the search a lot more efficient by filtering out whole swaths of data in one go.

Timescale-vector supports time partitioning using ‘s hypertables. To use this feature, simply indicate the length of time for each partition when creating the client:

from datetime import timedelta
from datetime import datetime

vec = client.Async(service_url, "my_data_with_time_partition", 2, time_partition_interval=timedelta(hours=6))
await vec.create_tables()

Then, insert data where the IDs use UUIDs v1 and the time component of the UUIDspecifies the time of the embedding. For example, to create an embedding for the current time, simply do:

id = uuid.uuid1()
await vec.upsert([(id, {"key": "val"}, "the brown fox", [1.0, 1.2])])

To insert data for a specific time in the past, create the UUID using the uuid_from_time function

specific_datetime = datetime(2018, 8, 10, 15, 30, 0)
await vec.upsert([(client.uuid_from_time(specific_datetime), {"key": "val"}, "the brown fox", [1.0, 1.2])])

You can then query the data by specifying a uuid_time_filter in the search call:

rec = await vec.search([1.0, 2.0], limit=4, uuid_time_filter=client.UUIDTimeRange(specific_datetime-timedelta(days=7), specific_datetime+timedelta(days=7)))

Distance metrics

Cosine distance is used by default to measure how similarly an embedding is to a given query. In addition to cosine distance, Euclidean/L2 distance is also supported. The distance type is set when creating the client using the distance_type parameter. For example, to use the Euclidean distance metric, you can create the client with:

vec  = client.Sync(service_url, "my_data", 2, distance_type="euclidean")

Valid values for distance_type are cosine and euclidean.

It is important to note that you should use consistent distance types on clients that create indexes and perform queries. That is because an index is only valid for one particular type of distance measure.

Note that the StreamingDiskANN index only supports cosine distance at this time.