Postgres Vector Store#

In this notebook we are going to show how to use Postgresql and pgvector to perform vector searches in LlamaIndex

If you’re opening this Notebook on colab, you will probably need to install LlamaIndex 🦙.

%pip install llama-index-vector-stores-postgres

!pip install llama-index

Running the following cell will install Postgres with PGVector in Colab.

!sudo apt update
!echo | sudo apt install -y postgresql-common
!echo | sudo /usr/share/postgresql-common/pgdg/apt.postgresql.org.sh
!echo | sudo apt install postgresql-15-pgvector
!sudo service postgresql start
!sudo -u postgres psql -c "ALTER USER postgres PASSWORD 'password';"
!sudo -u postgres psql -c "CREATE DATABASE vector_db;"

# import logging
# import sys

# Uncomment to see debug logs
# logging.basicConfig(stream=sys.stdout, level=logging.DEBUG)
# logging.getLogger().addHandler(logging.StreamHandler(stream=sys.stdout))

from llama_index.core import SimpleDirectoryReader, StorageContext
from llama_index.core import VectorStoreIndex
from llama_index.vector_stores.postgres import PGVectorStore
import textwrap
import openai

Setup OpenAI#

The first step is to configure the openai key. It will be used to created embeddings for the documents loaded into the index

import os

os.environ["OPENAI_API_KEY"] = "<your key>"
openai.api_key = os.environ["OPENAI_API_KEY"]

Download Data

!mkdir -p 'data/paul_graham/'
!wget 'https://raw.githubusercontent.com/run-llama/llama_index/main/docs/examples/data/paul_graham/paul_graham_essay.txt' -O 'data/paul_graham/paul_graham_essay.txt'

--2024-03-14 02:56:30--  https://raw.githubusercontent.com/run-llama/llama_index/main/docs/examples/data/paul_graham/paul_graham_essay.txt
Resolving raw.githubusercontent.com (raw.githubusercontent.com)... 185.199.108.133, 185.199.109.133, 185.199.111.133, ...
Connecting to raw.githubusercontent.com (raw.githubusercontent.com)|185.199.108.133|:443... connected.
HTTP request sent, awaiting response... 200 OK
Length: 75042 (73K) [text/plain]
Saving to: ‘data/paul_graham/paul_graham_essay.txt’

data/paul_graham/pa 100%[===================>]  73.28K  --.-KB/s    in 0.001s  

2024-03-14 02:56:30 (72.2 MB/s) - ‘data/paul_graham/paul_graham_essay.txt’ saved [75042/75042]

Loading documents#

Load the documents stored in the data/paul_graham/ using the SimpleDirectoryReader

documents = SimpleDirectoryReader("./data/paul_graham").load_data()
print("Document ID:", documents[0].doc_id)

Document ID: 1306591e-cc2d-430b-a74c-03ae7105ecab

Create the Database#

Using an existing postgres running at localhost, create the database we’ll be using.

import psycopg2

connection_string = "postgresql://postgres:password@localhost:5432"
db_name = "vector_db"
conn = psycopg2.connect(connection_string)
conn.autocommit = True

with conn.cursor() as c:
    c.execute(f"DROP DATABASE IF EXISTS {db_name}")
    c.execute(f"CREATE DATABASE {db_name}")

Create the index#

Here we create an index backed by Postgres using the documents loaded previously. PGVectorStore takes a few arguments.

from sqlalchemy import make_url

url = make_url(connection_string)
vector_store = PGVectorStore.from_params(
    database=db_name,
    host=url.host,
    password=url.password,
    port=url.port,
    user=url.username,
    table_name="paul_graham_essay",
    embed_dim=1536,  # openai embedding dimension
)

storage_context = StorageContext.from_defaults(vector_store=vector_store)
index = VectorStoreIndex.from_documents(
    documents, storage_context=storage_context, show_progress=True
)
query_engine = index.as_query_engine()

Query the index#

We can now ask questions using our index.

response = query_engine.query("What did the author do?")

print(textwrap.fill(str(response), 100))

The author worked on writing and programming before college, initially focusing on writing short
stories and later transitioning to programming on early computers like the IBM 1401 using Fortran.
The author continued programming on microcomputers like the TRS-80, creating simple games and a word
processor. In college, the author initially planned to study philosophy but switched to studying AI
due to a lack of interest in philosophy courses. The author was inspired to work on AI after
encountering works like Heinlein's novel "The Moon is a Harsh Mistress" and seeing Terry Winograd
using SHRDLU in a PBS documentary.

response = query_engine.query("What happened in the mid 1980s?")

print(textwrap.fill(str(response), 100))

AI was in the air in the mid 1980s, with two main influences that sparked interest in working on it:
a novel by Heinlein called The Moon is a Harsh Mistress, featuring an intelligent computer called
Mike, and a PBS documentary showing Terry Winograd using SHRDLU.

Querying existing index#

vector_store = PGVectorStore.from_params(
    database="vector_db",
    host="localhost",
    password="password",
    port=5432,
    user="postgres",
    table_name="paul_graham_essay",
    embed_dim=1536,  # openai embedding dimension
)

index = VectorStoreIndex.from_vector_store(vector_store=vector_store)
query_engine = index.as_query_engine()

response = query_engine.query("What did the author do?")

print(textwrap.fill(str(response), 100))

The author worked on writing short stories and programming before college. Initially, the author
wrote short stories and later started programming on an IBM 1401 using an early version of Fortran.
With the introduction of microcomputers, the author's interest in programming grew, leading to
writing simple games, predictive programs, and a word processor. Despite initially planning to study
philosophy in college, the author switched to studying AI due to a lack of interest in philosophy
courses. The author was inspired to work on AI after encountering a novel featuring an intelligent
computer and a PBS documentary showcasing AI technology.

Hybrid Search#

To enable hybrid search, you need to:

pass in hybrid_search=True when constructing the PGVectorStore (and optionally configure text_search_config with the desired language)
pass in vector_store_query_mode="hybrid" when constructing the query engine (this config is passed to the retriever under the hood). You can also optionally set the sparse_top_k to configure how many results we should obtain from sparse text search (default is using the same value as similarity_top_k).

from sqlalchemy import make_url

url = make_url(connection_string)
hybrid_vector_store = PGVectorStore.from_params(
    database=db_name,
    host=url.host,
    password=url.password,
    port=url.port,
    user=url.username,
    table_name="paul_graham_essay_hybrid_search",
    embed_dim=1536,  # openai embedding dimension
    hybrid_search=True,
    text_search_config="english",
)

storage_context = StorageContext.from_defaults(
    vector_store=hybrid_vector_store
)
hybrid_index = VectorStoreIndex.from_documents(
    documents, storage_context=storage_context
)

hybrid_query_engine = hybrid_index.as_query_engine(
    vector_store_query_mode="hybrid", sparse_top_k=2
)
hybrid_response = hybrid_query_engine.query(
    "Who does Paul Graham think of with the word schtick"
)

/workspaces/llama_index/llama-index-integrations/vector_stores/llama-index-vector-stores-postgres/llama_index/vector_stores/postgres/base.py:571: SAWarning: UserDefinedType REGCONFIG() will not produce a cache key because the ``cache_ok`` attribute is not set to True.  This can have significant performance implications including some performance degradations in comparison to prior SQLAlchemy versions.  Set this attribute to True if this type object's state is safe to use in a cache key, or False to disable this warning. (Background on this warning at: https://sqlalche.me/e/20/cprf)
  res = session.execute(stmt)

print(hybrid_response)

Roy Lichtenstein

Improving hybrid search with QueryFusionRetriever#

Since the scores for text search and vector search are calculated differently, the nodes that were found only by text search will have a much lower score.

You can often improve hybrid search performance by using QueryFusionRetriever, which makes better use of the mutual information to rank the nodes.

from llama_index.core.response_synthesizers import CompactAndRefine
from llama_index.core.retrievers import QueryFusionRetriever
from llama_index.core.query_engine import RetrieverQueryEngine

vector_retriever = hybrid_index.as_retriever(
    vector_store_query_mode="default",
    similarity_top_k=5,
)
text_retriever = hybrid_index.as_retriever(
    vector_store_query_mode="sparse",
    similarity_top_k=5,  # interchangeable with sparse_top_k in this context
)
retriever = QueryFusionRetriever(
    [vector_retriever, text_retriever],
    similarity_top_k=5,
    num_queries=1,  # set this to 1 to disable query generation
    mode="relative_score",
    use_async=False,
)

response_synthesizer = CompactAndRefine()
query_engine = RetrieverQueryEngine(
    retriever=retriever,
    response_synthesizer=response_synthesizer,
)

response = query_engine.query(
    "Who does Paul Graham think of with the word schtick, and why?"
)
print(response)

Paul Graham thinks of Roy Lichtenstein when he uses the word "schtick" because he recognizes paintings resembling a specific type of cartoon style as being created by Roy Lichtenstein.

Metadata filters#

PGVectorStore supports storing metadata in nodes, and filtering based on that metadata during the retrieval step.

Download git commits dataset#

!mkdir -p 'data/git_commits/'
!wget 'https://raw.githubusercontent.com/run-llama/llama_index/main/docs/examples/data/csv/commit_history.csv' -O 'data/git_commits/commit_history.csv'

--2024-03-14 02:56:46--  https://raw.githubusercontent.com/run-llama/llama_index/main/docs/examples/data/csv/commit_history.csv
Resolving raw.githubusercontent.com (raw.githubusercontent.com)... 185.199.111.133, 185.199.108.133, 185.199.109.133, ...
Connecting to raw.githubusercontent.com (raw.githubusercontent.com)|185.199.111.133|:443... connected.
HTTP request sent, awaiting response... 200 OK
Length: 1753902 (1.7M) [text/plain]
Saving to: ‘data/git_commits/commit_history.csv’

data/git_commits/co 100%[===================>]   1.67M  --.-KB/s    in 0.02s   

2024-03-14 02:56:46 (106 MB/s) - ‘data/git_commits/commit_history.csv’ saved [1753902/1753902]

import csv

with open("data/git_commits/commit_history.csv", "r") as f:
    commits = list(csv.DictReader(f))

print(commits[0])
print(len(commits))

{'commit': '44e41c12ab25e36c202f58e068ced262eadc8d16', 'author': 'Lakshmi Narayanan Sreethar<[email protected]>', 'date': 'Tue Sep 5 21:03:21 2023 +0530', 'change summary': 'Fix segfault in set_integer_now_func', 'change details': 'When an invalid function oid is passed to set_integer_now_func, it finds out that the function oid is invalid but before throwing the error, it calls ReleaseSysCache on an invalid tuple causing a segfault. Fixed that by removing the invalid call to ReleaseSysCache.  Fixes #6037 '}
4167

Add nodes with custom metadata#

# Create TextNode for each of the first 100 commits
from llama_index.core.schema import TextNode
from datetime import datetime
import re

nodes = []
dates = set()
authors = set()
for commit in commits[:100]:
    author_email = commit["author"].split("<")[1][:-1]
    commit_date = datetime.strptime(
        commit["date"], "%a %b %d %H:%M:%S %Y %z"
    ).strftime("%Y-%m-%d")
    commit_text = commit["change summary"]
    if commit["change details"]:
        commit_text += "\n\n" + commit["change details"]
    fixes = re.findall(r"#(\d+)", commit_text, re.IGNORECASE)
    nodes.append(
        TextNode(
            text=commit_text,
            metadata={
                "commit_date": commit_date,
                "author": author_email,
                "fixes": fixes,
            },
        )
    )
    dates.add(commit_date)
    authors.add(author_email)

print(nodes[0])
print(min(dates), "to", max(dates))
print(authors)

Node ID: 69513543-dee5-4c65-b4b8-39295f11e669
Text: Fix segfault in set_integer_now_func  When an invalid function
oid is passed to set_integer_now_func, it finds out that the function
oid is invalid but before throwing the error, it calls ReleaseSysCache
on an invalid tuple causing a segfault. Fixed that by removing the
invalid call to ReleaseSysCache.  Fixes #6037
2023-03-22 to 2023-09-05
{'[email protected]', '[email protected]', '[email protected]', '[email protected]', '[email protected]', '[email protected]', '[email protected]', '[email protected]', '[email protected]', '[email protected]', '[email protected]', '[email protected]', '[email protected]', '[email protected]', '[email protected]'}

vector_store = PGVectorStore.from_params(
    database=db_name,
    host=url.host,
    password=url.password,
    port=url.port,
    user=url.username,
    table_name="metadata_filter_demo3",
    embed_dim=1536,  # openai embedding dimension
)

index = VectorStoreIndex.from_vector_store(vector_store=vector_store)
index.insert_nodes(nodes)

print(index.as_query_engine().query("How did Lakshmi fix the segfault?"))

Lakshmi fixed the segfault by removing the invalid call to ReleaseSysCache that was causing the issue.

Apply metadata filters#

Now we can filter by commit author or by date when retrieving nodes.

from llama_index.core.vector_stores.types import (
    MetadataFilter,
    MetadataFilters,
)

filters = MetadataFilters(
    filters=[
        MetadataFilter(key="author", value="[email protected]"),
        MetadataFilter(key="author", value="[email protected]"),
    ],
    condition="or",
)

retriever = index.as_retriever(
    similarity_top_k=10,
    filters=filters,
)

retrieved_nodes = retriever.retrieve("What is this software project about?")

for node in retrieved_nodes:
    print(node.node.metadata)

{'commit_date': '2023-08-07', 'author': '[email protected]', 'fixes': []}
{'commit_date': '2023-08-27', 'author': '[email protected]', 'fixes': []}
{'commit_date': '2023-07-13', 'author': '[email protected]', 'fixes': []}
{'commit_date': '2023-08-07', 'author': '[email protected]', 'fixes': []}
{'commit_date': '2023-08-30', 'author': '[email protected]', 'fixes': []}
{'commit_date': '2023-08-15', 'author': '[email protected]', 'fixes': []}
{'commit_date': '2023-08-23', 'author': '[email protected]', 'fixes': []}
{'commit_date': '2023-08-10', 'author': '[email protected]', 'fixes': []}
{'commit_date': '2023-07-25', 'author': '[email protected]', 'fixes': ['5892']}
{'commit_date': '2023-08-21', 'author': '[email protected]', 'fixes': []}

filters = MetadataFilters(
    filters=[
        MetadataFilter(key="commit_date", value="2023-08-15", operator=">="),
        MetadataFilter(key="commit_date", value="2023-08-25", operator="<="),
    ],
    condition="and",
)

retriever = index.as_retriever(
    similarity_top_k=10,
    filters=filters,
)

retrieved_nodes = retriever.retrieve("What is this software project about?")

for node in retrieved_nodes:
    print(node.node.metadata)

{'commit_date': '2023-08-23', 'author': '[email protected]', 'fixes': []}
{'commit_date': '2023-08-17', 'author': '[email protected]', 'fixes': []}
{'commit_date': '2023-08-15', 'author': '[email protected]', 'fixes': []}
{'commit_date': '2023-08-15', 'author': '[email protected]', 'fixes': []}
{'commit_date': '2023-08-24', 'author': '[email protected]', 'fixes': []}
{'commit_date': '2023-08-15', 'author': '[email protected]', 'fixes': []}
{'commit_date': '2023-08-23', 'author': '[email protected]', 'fixes': []}
{'commit_date': '2023-08-21', 'author': '[email protected]', 'fixes': []}
{'commit_date': '2023-08-20', 'author': '[email protected]', 'fixes': []}
{'commit_date': '2023-08-21', 'author': '[email protected]', 'fixes': []}

Apply nested filters#

In the above examples, we combined multiple filters using AND or OR. We can also combine multiple sets of filters.

e.g. in SQL:

WHERE (commit_date >= '2023-08-01' AND commit_date <= '2023-08-15') AND (author = '[email protected]' OR author = '[email protected]')

filters = MetadataFilters(
    filters=[
        MetadataFilters(
            filters=[
                MetadataFilter(
                    key="commit_date", value="2023-08-01", operator=">="
                ),
                MetadataFilter(
                    key="commit_date", value="2023-08-15", operator="<="
                ),
            ],
            condition="and",
        ),
        MetadataFilters(
            filters=[
                MetadataFilter(key="author", value="[email protected]"),
                MetadataFilter(key="author", value="[email protected]"),
            ],
            condition="or",
        ),
    ],
    condition="and",
)

retriever = index.as_retriever(
    similarity_top_k=10,
    filters=filters,
)

retrieved_nodes = retriever.retrieve("What is this software project about?")

for node in retrieved_nodes:
    print(node.node.metadata)

{'commit_date': '2023-08-07', 'author': '[email protected]', 'fixes': []}
{'commit_date': '2023-08-07', 'author': '[email protected]', 'fixes': []}
{'commit_date': '2023-08-15', 'author': '[email protected]', 'fixes': []}
{'commit_date': '2023-08-10', 'author': '[email protected]', 'fixes': []}

The above can be simplified by using the IN operator. PGVectorStore supports in, nin, and contains for comparing an element with a list.

filters = MetadataFilters(
    filters=[
        MetadataFilter(key="commit_date", value="2023-08-01", operator=">="),
        MetadataFilter(key="commit_date", value="2023-08-15", operator="<="),
        MetadataFilter(
            key="author",
            value=["[email protected]", "[email protected]"],
            operator="in",
        ),
    ],
    condition="and",
)

retriever = index.as_retriever(
    similarity_top_k=10,
    filters=filters,
)

retrieved_nodes = retriever.retrieve("What is this software project about?")

for node in retrieved_nodes:
    print(node.node.metadata)

{'commit_date': '2023-08-07', 'author': '[email protected]', 'fixes': []}
{'commit_date': '2023-08-07', 'author': '[email protected]', 'fixes': []}
{'commit_date': '2023-08-15', 'author': '[email protected]', 'fixes': []}
{'commit_date': '2023-08-10', 'author': '[email protected]', 'fixes': []}

# Same thing, with NOT IN
filters = MetadataFilters(
    filters=[
        MetadataFilter(key="commit_date", value="2023-08-01", operator=">="),
        MetadataFilter(key="commit_date", value="2023-08-15", operator="<="),
        MetadataFilter(
            key="author",
            value=["[email protected]", "[email protected]"],
            operator="nin",
        ),
    ],
    condition="and",
)

retriever = index.as_retriever(
    similarity_top_k=10,
    filters=filters,
)

retrieved_nodes = retriever.retrieve("What is this software project about?")

for node in retrieved_nodes:
    print(node.node.metadata)

{'commit_date': '2023-08-09', 'author': '[email protected]', 'fixes': ['5805']}
{'commit_date': '2023-08-15', 'author': '[email protected]', 'fixes': []}
{'commit_date': '2023-08-15', 'author': '[email protected]', 'fixes': []}
{'commit_date': '2023-08-11', 'author': '[email protected]', 'fixes': []}
{'commit_date': '2023-08-09', 'author': '[email protected]', 'fixes': ['5923', '5680', '5774', '5786', '5906', '5912']}
{'commit_date': '2023-08-03', 'author': '[email protected]', 'fixes': []}
{'commit_date': '2023-08-03', 'author': '[email protected]', 'fixes': ['5908']}
{'commit_date': '2023-08-01', 'author': '[email protected]', 'fixes': []}
{'commit_date': '2023-08-10', 'author': '[email protected]', 'fixes': []}
{'commit_date': '2023-08-10', 'author': '[email protected]', 'fixes': []}

# CONTAINS
filters = MetadataFilters(
    filters=[
        MetadataFilter(key="fixes", value="5680", operator="contains"),
    ]
)

retriever = index.as_retriever(
    similarity_top_k=10,
    filters=filters,
)

retrieved_nodes = retriever.retrieve("How did these commits fix the issue?")
for node in retrieved_nodes:
    print(node.node.metadata)

{'commit_date': '2023-08-09', 'author': '[email protected]', 'fixes': ['5923', '5680', '5774', '5786', '5906', '5912']}

PgVector Query Options#

IVFFlat Probes#

Specify the number of IVFFlat probes (1 by default)

When retrieving from the index, you can specify an appropriate number of IVFFlat probes (higher is better for recall, lower is better for speed)

retriever = index.as_retriever(
    vector_store_query_mode="hybrid",
    similarity_top_k=5,
    vector_store_kwargs={"ivfflat_probes": 10},
)

HNSW EF Search#

Specify the size of the dynamic candidate list for search (40 by default)

retriever = index.as_retriever(
    vector_store_query_mode="hybrid",
    similarity_top_k=5,
    vector_store_kwargs={"hnsw_ef_search": 300},
)