Oracle AI Vector Search Tutorial: Store and Query Embeddings With Python

This tutorial introduces Oracle AI Vector Search, explains its core features, and walks through a practical local project. It also covers setup requirements, semantic search, vector indexing, and common troubleshooting steps.

6 août 2026 · 15 min lire

Explorer avec l’IA

Ouvrir dans ChatGPT Ouvrir dans Claude Ouvrir dans Perplexity

In this tutorial, we’ll build a local semantic search demo with Oracle AI Vector Search, Python, and Oracle Database Free 26ai.

We’ll start with handwritten three-dimensional vectors so the distance calculation is visible, then generate text embeddings and query them from Oracle Database with SQL.

This tutorial is for Python and SQL developers who are new to vector search and comfortable with Docker, environment variables, Python packages, and local database setup.

AI-powered applications need search that understands meaning, not only exact words. A support app, documentation portal, or internal knowledge tool should be able to find “database storage for AI search” even when the best matching document says “store vector embeddings in native columns.”

Vector embeddings make that possible by representing text as numeric vectors. Oracle AI Vector Search lets us store those vectors directly in Oracle Database, query them with SQL, and keep embeddings beside relational application data.

For this local workflow, we do not need a separate vector database.

We’ll run Oracle Database Free 26ai locally, connect from Python with oracledb Thin mode, store manual vectors first, store OpenAI embeddings next, run semantic search, compare semantic retrieval with a simple exact phrase predicate, and create a vector index validated through USER_INDEXES.

By the end, we’ll have a working Python-powered semantic search workflow that stores embeddings in Oracle Database and retrieves the most similar documents locally.

What Is Oracle AI Vector Search?

Oracle AI Vector Search is a set of Oracle Database capabilities for storing, indexing, and querying vector embeddings. A vector embedding is a fixed-size list of numbers that represents the meaning of text, images, or other data in a form that a database can compare mathematically.

An embedding model creates the numeric vectors. It takes input such as a sentence, paragraph, image description, or code snippet and returns a vector with a fixed number of dimensions.

Similar inputs should produce vectors that are close together, while unrelated inputs should land farther apart. We must therefore use the same embedding model for stored documents and incoming queries because distances are meaningful only when vectors share the same coordinate system.

In this tutorial, the manual vectors are deliberately small so we can inspect the math. The real text embeddings are much larger because the model needs enough dimensions to encode more subtle relationships between words, phrases, and topics.

Vector search ranks rows by distance between vectors. For Euclidean and cosine distance queries, smaller distance values mean closer matches. In this tutorial, we’ll start with a small VECTOR(3, FLOAT32) column so we can see distance behavior directly, then move to VECTOR(1536, FLOAT32) embeddings generated from text.

Figure 1. Oracle AI Vector Search flow

Oracle AI Vector Search stores embeddings in a native VECTOR column and ranks results by vector distance.

Oracle Database is useful when embeddings belong beside relational application data such as document IDs, titles, content, users, products, tickets, or metadata.

Some applications use a separate vector database or search engine for specialized retrieval workloads, but for this tutorial’s local semantic search workflow, Oracle Database provides the vector storage and search features we need.

Key Features of Oracle AI Vector Search

We’ll use four Oracle AI Vector Search features in the demo: native vector storage, distance functions, vector indexes, and relational integration.

Native vector data type

Oracle Database supports native VECTOR columns with explicit dimensions and element formats. In the demo, we’ll use VECTOR(3, FLOAT32) for manual vectors and VECTOR(1536, FLOAT32) for the default text-embedding-3-small embedding model.

Distance functions and operators

The runnable examples use VECTOR_DISTANCE() because it keeps the distance metric explicit in SQL.

We’ll use Euclidean distance for the manual vector example and cosine distance for text embeddings.

Oracle Database also includes shorthand vector distance operators and dot-product-related distance behavior, but this tutorial keeps runnable SQL on the explicit VECTOR_DISTANCE() path.

Different distance metrics answer slightly different questions about “closeness.”

For a small tutorial, cosine distance is a good default for text embeddings because it focuses on direction, which often maps well to semantic similarity.

Euclidean distance is easier to visualize with handwritten vectors, and dot-product-style comparisons are useful in some retrieval systems when vector magnitude or normalized vectors are part of the model design.

Metric	What It Compares	Good Fit	Watch For
Cosine	Vector direction	Text embeddings and semantic similarity	Results depend on using the same embedding model and consistent preprocessing
Euclidean	Straight-line distance	Small examples, geometry intuition, some numeric feature vectors	Magnitude affects the distance, so scaling can matter
Dot product	Direction and magnitude together	Workflows designed around normalized vectors or inner-product scoring	Score direction and normalization rules need to be understood before comparing results

Vector indexes

Exact vector search is useful for correctness and small datasets. As datasets grow, vector indexes support approximate nearest-neighbor search workflows.

Oracle AI Vector Search includes in-memory neighbor graph indexes, often associated with HNSW-style search, and neighbor-partitions indexes, often associated with IVF-style partitioning.

In this local demo, we use ORGANIZATION NEIGHBOR PARTITIONS and validate the index with USER_INDEXES.

Approximate indexes trade a small amount of exhaustive-search certainty for faster retrieval at scale.

The right index family depends on the workload, data volume, update pattern, and recall target.

This tutorial uses an IVF-style neighbor-partitions index because it keeps the local setup simple and provides a concrete index-creation step without turning the tutorial into an index-tuning guide.

Index Family	Basic Idea	Strengths	Tradeoffs	When We Might Use It
IVF / neighbor partitions	Partition the vector space, then search likely partitions	Simple mental model, practical for many batch-loaded datasets	Recall and speed depend on partitioning and search settings	We want a straightforward approximate-search path for a growing dataset
HNSW / neighbor graph	Build a graph that links nearby vectors for fast traversal	Strong recall and latency profile for many nearest-neighbor workloads	More memory-oriented, and configuration can matter more	We need low-latency search and can budget memory and tuning effort

Integration with relational queries

A vector column can live beside ordinary relational columns. Our documents table will store title, content, and embedding together, so SQL can return both the similarity score and the original document text.

How to Get Started With Oracle AI Vector Search

We’ll use one local setup path: Oracle Database Free 26ai in Docker, the FREEPDB1 service, and a dedicated vector_demo application user.

The vector-search concepts are approachable, but the setup is intermediate because we use Docker, a database schema, Python packages, environment variables, and an embedding API key.

Prerequisites

We’ll need:

Docker Desktop or Docker Engine.
Access to Oracle Container Registry.
Local ports 1521 and 5500 available.
Python 3.10 or later.
Basic Python and SQL knowledge.
Basic Docker and environment-variable familiarity.
An OpenAI API key for the embedding steps.

This demo requires an Oracle Database version or environment that includes the vector features used here, including native VECTOR columns, VECTOR_DISTANCE(), and CREATE VECTOR INDEX.

The local path uses Oracle Database Free 26ai with the image tag shown below, and the DSN uses the FREEPDB1 pluggable database service.

The manual vector steps do not require an OpenAI API key. The semantic steps send the sample document text and query text to the configured embedding provider, so we’ll use only sample or non-sensitive text.

This tutorial uses oracledb, the current Python driver for Oracle Database. python-oracledb runs in Thin mode by default, so this local demo does not require Oracle Client libraries. Thick mode is useful in other Oracle Database deployments, but we will not call oracledb.init_oracle_client() here.

Start Oracle Database Free 26ai locally

Create a project directory for the scripts we’ll build.

mkdir oracle-ai-vector-search-python
 cd oracle-ai-vector-search-python

Start Oracle Database Free 26ai with the local container image.

docker run --name oracle-free-26ai-vector \
   --detach \
   --publish 1521:1521 \
   --publish 5500:5500 \
   --shm-size=1g \
   --env ORACLE_PWD="replace-with-a-strong-password" \
   container-registry.oracle.com/database/free:23.26.1.0-lite-amd64

Check the container logs and wait until the database reports that startup is complete.

docker logs oracle-free-26ai-vector

FREEPDB1 is the pluggable database service used by this local Oracle Database Free container path. Our Python DSN will be localhost:1521/FREEPDB1.

Create the vector_demo schema in FREEPDB1

Use SYS only for the schema setup step. The password in the sqlplus command is the ORACLE_PWD value from the docker run command; the password in CREATE USER is the application password we will export as DB_PASSWORD.

Run the setup inside the database container.

# Create the application user and grant the privileges needed for the demo.
 docker exec -i oracle-free-26ai-vector \
   sqlplus -s 'sys/"replace-with-oracle-pwd"@FREEPDB1 as sysdba' <<'SQL'
 WHENEVER SQLERROR EXIT SQL.SQLCODE

 CREATE USER vector_demo IDENTIFIED BY "replace-with-vector-demo-password";

 GRANT CREATE SESSION, CREATE TABLE, CREATE SEQUENCE, UNLIMITED TABLESPACE TO vector_demo;

 EXIT
 SQL

This is a local development shortcut for the demo container. CREATE SEQUENCE is needed because the tables use identity columns.

For production, we would review privileges with a database administrator and apply least privilege for the exact application workload.

Install Python dependencies

Create and activate a virtual environment.

python -m venv .venv
 source .venv/bin/activate
 python -m pip install --upgrade pip

Install the Python packages used by the demo.

python -m pip install oracledb openai

Set environment variables

Store database credentials and embedding settings in environment variables instead of hard-coding them in Python files.

export DB_USER="vector_demo"
 export DB_PASSWORD="replace-with-vector-demo-password"
 export DB_DSN="localhost:1521/FREEPDB1"
 export OPENAI_API_KEY="replace-with-your-api-key"
 export EMBEDDING_MODEL="text-embedding-3-small"
 export EMBEDDING_DIM="1536"

OPENAI_API_KEY is required only for the embedding and semantic search steps. EMBEDDING_DIM must match the number of dimensions returned by the selected embedding model.

Oracle AI Vector Search Demo Project

We’ll build the demo in eight checkpoints. Each script adds one concept and produces output we can verify before moving on.

The database-only checkpoint is complete after Steps 1–3. At that point, we will have connected to Oracle Database, stored manual vectors, queried them by distance, and created the semantic table. The OpenAI API key becomes required in Step 4.

Step 1: Connect to the local database

First, we’ll verify that Python can connect to Oracle Database as the vector_demo application user. This script reads DB_USER, DB_PASSWORD, and DB_DSN from the environment and prints the driver mode and database version.

Create the script

Save the following code as 01_check_connection.py.

import os

 import oracledb

 ADMIN_USERS = {"SYS", "SYSTEM", "PDBADMIN"}


 def required_env(name):
 	value = os.getenv(name)
 	if not value:
         raise SystemExit(f"Set {name} before running this script.")
 	return value


 def main():
 	# Load required database settings from environment variables.
 	user = required_env("DB_USER")
     password = required_env("DB_PASSWORD")
 	dsn = required_env("DB_DSN")

 	if user.upper() in ADMIN_USERS:
         raise SystemExit(
             "Use the vector_demo application user, not an admin user."
     	)

 	# Open a Thin-mode connection with the application user.
 	with oracledb.connect(user=user, password=password, dsn=dsn) as connection:
         mode = "thin" if oracledb.is_thin_mode() else "thick"
         print("Connected to Oracle Database.")
         print(f"Driver mode: {mode}")
         print(f"Database version: {connection.version}")


 if __name__ == "__main__":
 	main()

Run the script

Run the connection check.

python 01_check_connection.py

Expected output

The output should confirm that the connection works and that the driver is in Thin mode.

Connected to Oracle Database.
 Driver mode: thin
 Database version: 23.26.1.0.0

We now know that the local database, service name, credentials, and Python driver are working before we create any vector tables.

Step 2: Build intuition with manual vectors

Before we introduce embedding models, let’s store three small vectors we can reason about. The array.array("f", values) call creates a 32-bit floating-point array that python-oracledb can bind into a VECTOR(..., FLOAT32) column.

Figure 2. Manual vectors and Euclidean distance.

The manual example uses Euclidean distance. The apple vector has a distance of 0.0000 because it matches the query vector. The banana vector ranks second because sqrt((1.0 - 0.9)^2 + (0.0 - 0.1)^2 + (0.0 - 0.0)^2) = 0.1414.

Create the script

Save the following code as 02_manual_vectors.py.

import array
 import os

 import oracledb

 ADMIN_USERS = {"SYS", "SYSTEM", "PDBADMIN"}


 def required_env(name):
 	value = os.getenv(name)
 	if not value:
         raise SystemExit(f"Set {name} before running this script.")
 	return value


 def main():
 	user = required_env("DB_USER")
 	if user.upper() in ADMIN_USERS:
         raise SystemExit(
             "Use the vector_demo application user, not an admin user."
     	)

 	# Define three small vectors so the distance calculation is easy to inspect.
 	rows = [
         ("apple", array.array("f", [1.0, 0.0, 0.0])),
         ("banana", array.array("f", [0.9, 0.1, 0.0])),
         ("car", array.array("f", [0.0, 1.0, 0.0])),
 	]

 	with oracledb.connect(
         user=user,
         password=required_env("DB_PASSWORD"),
         dsn=required_env("DB_DSN"),
 	) as connection:
         with connection.cursor() as cursor:
             # Recreate the table so the script can be run repeatedly.
             cursor.execute(
                 """
                 BEGIN
                 	EXECUTE IMMEDIATE 'DROP TABLE manual_vectors PURGE';
                 EXCEPTION
                 	WHEN OTHERS THEN
                     	IF SQLCODE <> -942 THEN
                         	RAISE;
                     	END IF;
                 END;
                 """
             )
             cursor.execute(
                 """
                 CREATE TABLE manual_vectors (
                 	id NUMBER GENERATED BY DEFAULT AS IDENTITY PRIMARY KEY,
                 	label VARCHAR2(100),
                 	embedding VECTOR(3, FLOAT32)
                 )
                 """
             )
             # Bind each vector as a 32-bit floating-point array.
             cursor.executemany(
                 """
                 INSERT INTO manual_vectors (label, embedding)
                 VALUES (:1, :2)
                 """,
                 rows,
             )
             connection.commit()

             # Rank rows by Euclidean distance from the query vector.
             cursor.execute(
                 """
                 SELECT label,
                    	VECTOR_DISTANCE(
                        	embedding,
                        	:query_vector,
                        	EUCLIDEAN
                    	) AS distance
                 FROM manual_vectors
                 ORDER BY distance
                 FETCH EXACT FIRST 2 ROWS ONLY
                 """,
                 query_vector=array.array("f", [1.0, 0.0, 0.0]),
             )

             print("Created manual_vectors.")
             print("Inserted 3 manual vectors.")
             print("\nNearest manual vectors:")
             for rank, (label, distance) in enumerate(cursor, start=1):
                 print(f"{rank}. {label:<6} distance={float(distance):.4f}")


 if __name__ == "__main__":
 	main()

Run the script

Run the manual vector script.

python 02_manual_vectors.py

Expected output

The identical vector appears first with distance 0.0000, and the nearby banana vector appears second.

Created manual_vectors.
 Inserted 3 manual vectors.

 Nearest manual vectors:
 1. apple  distance=0.0000
 2. banana distance=0.1414

This is the core idea behind vector search: rows are ranked by mathematical closeness to a query vector. Embedding models will generate larger vectors, but the database behavior is the same.

Step 3: Create the semantic documents table

Now that we know vectors work, we’ll create a real application-style table. The important detail is that the vector column dimension must match the embedding model output dimension.

The dimension is part of the table contract. If an embedding model returns 1,536 numbers, the column must be declared as VECTOR(1536, FLOAT32).

If we later switch to a model with a different output length, we need to recreate or migrate the vector column so inserts and queries continue to use compatible vectors.

Because DDL cannot bind the vector dimension as a regular SQL bind variable, the script converts EMBEDDING_DIM to an integer before using it in the CREATE TABLE statement.

Create the script

Save the following code as 03_create_documents_table.py.

import os

 import oracledb

 ADMIN_USERS = {"SYS", "SYSTEM", "PDBADMIN"}


 def required_env(name):
 	value = os.getenv(name)
 	if not value:
         raise SystemExit(f"Set {name} before running this script.")
 	return value


 def configured_dimension():
 	try:
         dimension = int(required_env("EMBEDDING_DIM"))
 	except ValueError as exc:
         raise SystemExit("EMBEDDING_DIM must be an integer.") from exc

 	if dimension <= 0:
         raise SystemExit("EMBEDDING_DIM must be a positive integer.")

 	return dimension


 def main():
 	user = required_env("DB_USER")
 	if user.upper() in ADMIN_USERS:
         raise SystemExit(
             "Use the vector_demo application user, not an admin user."
     	)

 	# Validate the configured model name and vector dimension.
     model_name = required_env("EMBEDDING_MODEL")
     dimension = configured_dimension()

 	with oracledb.connect(
         user=user,
         password=required_env("DB_PASSWORD"),
         dsn=required_env("DB_DSN"),
 	) as connection:
         with connection.cursor() as cursor:
             # Recreate the documents table so the script is repeatable.
             cursor.execute(
                 """
                 BEGIN
                 	EXECUTE IMMEDIATE 'DROP TABLE documents PURGE';
                 EXCEPTION
                 	WHEN OTHERS THEN
                     	IF SQLCODE <> -942 THEN
                         	RAISE;
                     	END IF;
                 END;
                 """
             )
             cursor.execute(
                 f"""
                 CREATE TABLE documents (
                 	id NUMBER GENERATED BY DEFAULT AS IDENTITY PRIMARY KEY,
                 	title VARCHAR2(200) NOT NULL,
                 	content VARCHAR2(1000) NOT NULL,
                 	embedding VECTOR({dimension}, FLOAT32)
                 )
                 """
             )

     print(f"Embedding model: {model_name}")
     print(f"Embedding dimension: {dimension}")
     print("Recreated documents table.")


 if __name__ == "__main__":
 	main()

Run the script

Run the table creation script.

python 03_create_documents_table.py

Expected output

The output should show the configured embedding model and vector dimension.

Embedding model: text-embedding-3-small
 Embedding dimension: 1536
 Recreated documents table.

We have completed the database-only portion. Oracle Database can store and compare vectors locally; next, we’ll replace handwritten vectors with embeddings generated from text.

Step 4: Verify the embedding model dimension

Before we insert embeddings, we’ll ask the embedding model for one vector and verify its length. This prevents the most common vector-table error: inserting a vector whose length does not match the VECTOR column dimension.

An embedding model is not just a text-to-number converter; it defines the meaning space for the application. All document and query embeddings in this demo must come from the same model.

Using one model ensures that a small distance represents similar meaning rather than two unrelated vectors that happen to have similar shapes. The model also determines the vector length, which is why we check the returned dimension before loading data.

Create the script

Save the following code as 04_check_embedding_dimension.py.

import os

 from openai import OpenAI


 def required_env(name):
 	value = os.getenv(name)
 	if not value:
         raise SystemExit(f"Set {name} before running this script.")
 	return value


 def configured_dimension():
 	try:
         return int(required_env("EMBEDDING_DIM"))
 	except ValueError as exc:
         raise SystemExit("EMBEDDING_DIM must be an integer.") from exc


 def main():
     required_env("OPENAI_API_KEY")

     model_name = required_env("EMBEDDING_MODEL")
     expected_dimension = configured_dimension()

 	# Request one embedding to confirm the model output dimension.
 	client = OpenAI()
     response = client.embeddings.create(
         model=model_name,
         input="Oracle AI Vector Search dimension check",
 	)

     returned_dimension = len(response.data[0].embedding)

     print(f"Embedding model: {model_name}")
     print(f"Returned embedding length: {returned_dimension}")
     print(f"Configured EMBEDDING_DIM: {expected_dimension}")

 	# Stop before loading data if the table dimension is incompatible.
 	if returned_dimension != expected_dimension:
         raise SystemExit(
             "Dimension check failed. Update EMBEDDING_DIM, rerun "
             "03_create_documents_table.py, and try again."
     	)

     print("Dimension check passed.")


 if __name__ == "__main__":
 	main()

Run the script

Run the dimension check.

python 04_check_embedding_dimension.py

Expected output

For the default model and dimension, the output should look like this.

Embedding model: text-embedding-3-small
 Returned embedding length: 1536
 Configured EMBEDDING_DIM: 1536
 Dimension check passed.

If we choose a different embedding model later, we should update EMBEDDING_MODEL, set EMBEDDING_DIM to the returned vector length, and recreate the documents table.

Step 5: Generate and insert document embeddings

Now we’ll generate embeddings for a small self-contained dataset and insert them into Oracle Database.

Each embedding is converted to array.array("f", embedding) before binding so it matches the FLOAT32 vector column.

Create the script

Save the following code as 05_insert_embeddings.py.

import array
 import os

 import oracledb
 from openai import OpenAI

 DOCUMENTS = [
 	(
         "Oracle AI Vector Search",
         "Oracle Database can store vector embeddings in native VECTOR "
         "columns and query them with SQL similarity search.",
 	),
 	(
         "Semantic Search for Applications",
         "Semantic search ranks results by meaning, helping applications "
         "find related content even when exact words differ.",
 	),
 	(
         "Vector Index Workflow",
         "Vector indexes help approximate nearest-neighbor search scale "
         "when datasets grow beyond small examples.",
 	),
 	(
         "Local Docker Development",
         "Docker containers make it practical to run a local database for "
         "development and repeatable tutorials.",
 	),
 	(
         "Python Database Code",
         "Python applications can connect to Oracle Database with the "
         "oracledb driver and use environment variables for credentials.",
 	),
 	(
         "Acoustic Guitar Practice",
         "Daily guitar practice improves timing, chord transitions, and "
         "confidence when learning new songs.",
 	),
 	(
         "Weeknight Vegetable Soup",
         "A simple soup can combine onions, carrots, beans, herbs, and "
         "broth for an easy weeknight meal.",
 	),
 ]

 ADMIN_USERS = {"SYS", "SYSTEM", "PDBADMIN"}


 def required_env(name):
 	value = os.getenv(name)
 	if not value:
         raise SystemExit(f"Set {name} before running this script.")
 	return value


 def configured_dimension():
 	try:
         return int(required_env("EMBEDDING_DIM"))
 	except ValueError as exc:
         raise SystemExit("EMBEDDING_DIM must be an integer.") from exc


 def main():
 	user = required_env("DB_USER")
 	if user.upper() in ADMIN_USERS:
         raise SystemExit(
             "Use the vector_demo application user, not an admin user."
     	)

     required_env("OPENAI_API_KEY")
     model_name = required_env("EMBEDDING_MODEL")
     expected_dimension = configured_dimension()

 	# Generate embeddings for all sample documents in one API request.
 	client = OpenAI()
     response = client.embeddings.create(
         model=model_name,
         input=[content for _, content in DOCUMENTS],
 	)

 	# Restore the response order using each item's index.
     embeddings = [
         item.embedding
     	for item in sorted(
             response.data,
             key=lambda item: item.index,
     	)
 	]

 	if any(len(embedding) != expected_dimension for embedding in embeddings):
         raise SystemExit(
             "An embedding length did not match EMBEDDING_DIM. "
             "Rerun 04_check_embedding_dimension.py."
     	)

 	# Bind each embedding as a 32-bit floating-point array.
 	rows = [
         (title, content, array.array("f", embedding))
     	for (title, content), embedding in zip(DOCUMENTS, embeddings)
 	]

 	# Replace the sample rows and commit the transaction.
 	with oracledb.connect(
         user=user,
         password=required_env("DB_PASSWORD"),
         dsn=required_env("DB_DSN"),
 	) as connection:
         with connection.cursor() as cursor:
             cursor.execute("DELETE FROM documents")
             cursor.executemany(
                 """
                 INSERT INTO documents (title, content, embedding)
                 VALUES (:1, :2, :3)
                 """,
                 rows,
             )
             cursor.execute("SELECT COUNT(*) FROM documents")
             row_count = cursor.fetchone()[0]
             connection.commit()

     print(f"Embedding model: {model_name}")
     print(f"Generated embeddings for {len(DOCUMENTS)} documents.")
     print(f"Inserted {len(DOCUMENTS)} documents.")
     print(f"documents row count: {row_count}")


 if __name__ == "__main__":
 	main()

Run the script

Run the insert script.

python 05_insert_embeddings.py

Expected output

The row count should match the seven inline documents.

Embedding model: text-embedding-3-small
 Generated embeddings for 7 documents.
 Inserted 7 documents.
 documents row count: 7

We now have text and embeddings stored together in Oracle Database. The unrelated documents about guitar and soup are included, so the search results have obvious near and far matches.

Step 6: Run a semantic search query

Next, we’ll embed a natural-language query and compare it with stored document embeddings. This script uses cosine distance, which is commonly used for text embeddings because it focuses on vector direction rather than raw magnitude.

For semantic search, the query goes through the same embedding model as the documents. We are not asking SQL to understand the English sentence directly.

We are asking the embedding model to turn the sentence into a vector, then asking Oracle Database to rank stored vectors by their distance from that query vector.

Create the script

Save the following code as 06_semantic_search.py.

import array
 import os

 import oracledb
 from openai import OpenAI

 QUERY_TEXT = "Where should an AI app keep meaning-based search data?"
 ADMIN_USERS = {"SYS", "SYSTEM", "PDBADMIN"}


 def required_env(name):
 	value = os.getenv(name)
 	if not value:
         raise SystemExit(f"Set {name} before running this script.")
 	return value


 def main():
 	user = required_env("DB_USER")
 	if user.upper() in ADMIN_USERS:
         raise SystemExit(
             "Use the vector_demo application user, not an admin user."
     	)

     required_env("OPENAI_API_KEY")

 	# Generate an embedding for the natural-language query.
 	client = OpenAI()
     response = client.embeddings.create(
         model=required_env("EMBEDDING_MODEL"),
         input=QUERY_TEXT,
 	)
     query_vector = array.array("f", response.data[0].embedding)

 	with oracledb.connect(
         user=user,
         password=required_env("DB_PASSWORD"),
         dsn=required_env("DB_DSN"),
 	) as connection:
         with connection.cursor() as cursor:
             # Rank stored documents by cosine distance from the query vector.
             cursor.execute(
                 """
                 SELECT title,
                    	content,
                    	VECTOR_DISTANCE(
                        	embedding,
                        	:query_vector,
                        	COSINE
                    	) AS distance
                 FROM documents
                 ORDER BY distance
                 FETCH EXACT FIRST 5 ROWS ONLY
                 """,
                 query_vector=query_vector,
             )
             rows = cursor.fetchall()

 	if not rows:
         raise SystemExit("No documents found. Run 05_insert_embeddings.py first.")

     print(f"Query: {QUERY_TEXT}")
     print("\nTop 5 results:")

 	for rank, (title, content, distance) in enumerate(rows, start=1):
         print(f"{rank}. {title}")
         print(f"   distance: {float(distance):.6f}")
         print(f"   {content[:90]}...")


 if __name__ == "__main__":
 	main()

Run the script

Run the semantic search script.

python 06_semantic_search.py

Expected output

Exact rankings and distances can vary when the embedding model changes. However, documents about Oracle AI Vector Search, semantic search, Python database code, and vector indexing should rank near the top.

Query: Where should an AI app keep meaning-based search data?

 Top 5 results:
 1. Oracle AI Vector Search
    distance: <cosine distance from your run>
    Oracle Database can store vector embeddings in native VECTOR columns and query them with SQL...

 2. Semantic Search for Applications
    distance: <cosine distance from your run>
    Semantic search ranks results by meaning, helping applications find related content even when...

 3. Python Database Code
    distance: <cosine distance from your run>
    Python applications can connect to Oracle Database with the oracledb driver and use environment...

 4. Vector Index Workflow
    distance: <cosine distance from your run>
    Vector indexes help approximate nearest-neighbor search scale when datasets grow beyond small...

 5. Local Docker Development
    distance: <cosine distance from your run>
    Docker containers make it practical to run a local database for development and repeatable...

For the Euclidean and cosine distance queries in this tutorial, smaller values indicate closer matches.

Dot product is another useful vector comparison approach in many vector-search workflows, especially when magnitude or normalized vectors matter, but score direction and normalization need careful handling.

We’ll keep the runnable code on Euclidean and cosine distance so the result interpretation stays simple.

If we normalize embeddings ourselves in a future application, we need to normalize both stored document embeddings and query embeddings consistently; otherwise, distances and rankings can change.

Step 7: Compare semantic search with a phrase match

Semantic search is useful because it does not require the exact same phrase to appear in the document.

Let’s compare the semantic query with a simple exact phrase predicate.

This is not a full-text search engine demo; it is a small SQL comparison that makes the difference between literal matching and meaning-based ranking visible.

Create the script

Save the following code as 07_semantic_vs_keyword.py.

import array
 import os

 import oracledb
 from openai import OpenAI

 SEMANTIC_QUERY = "Where should an AI app keep meaning-based search data?"
 KEYWORD_PHRASE = "meaning-based search data"
 ADMIN_USERS = {"SYS", "SYSTEM", "PDBADMIN"}


 def required_env(name):
 	value = os.getenv(name)
 	if not value:
         raise SystemExit(f"Set {name} before running this script.")
 	return value


 def main():
 	user = required_env("DB_USER")
 	if user.upper() in ADMIN_USERS:
         raise SystemExit(
             "Use the vector_demo application user, not an admin user."
     	)

     required_env("OPENAI_API_KEY")

 	# Generate an embedding for the semantic query.
 	client = OpenAI()
     response = client.embeddings.create(
         model=required_env("EMBEDDING_MODEL"),
         input=SEMANTIC_QUERY,
 	)
     query_vector = array.array("f", response.data[0].embedding)

 	with oracledb.connect(
         user=user,
         password=required_env("DB_PASSWORD"),
         dsn=required_env("DB_DSN"),
 	) as connection:
         with connection.cursor() as cursor:
             # Rank documents by semantic similarity.
             cursor.execute(
                 """
                 SELECT title,
                        VECTOR_DISTANCE(embedding, :query_vector, COSINE)
                 FROM documents
                 ORDER BY VECTOR_DISTANCE(embedding, :query_vector, COSINE)
                 FETCH EXACT FIRST 3 ROWS ONLY
                 """,
                 query_vector=query_vector,
             )
             semantic_rows = cursor.fetchall()

             # Compare semantic results with an exact phrase predicate.
             cursor.execute(
                 """
                 SELECT title
                 FROM documents
                 WHERE LOWER(title || ' ' || content)
                   	LIKE '%' || :phrase || '%'
                 ORDER BY title
                 FETCH FIRST 3 ROWS ONLY
                 """,
                 phrase=KEYWORD_PHRASE.lower(),
             )
             keyword_rows = cursor.fetchall()

     print(f"Semantic query: {SEMANTIC_QUERY}")
     print(f"Keyword phrase: {KEYWORD_PHRASE}")

     print("\nSemantic results:")
 	for rank, (title, distance) in enumerate(semantic_rows, start=1):
         print(f"{rank}. {title}")
         print(f"   distance: {float(distance):.6f}")

     print("\nKeyword results:")
 	if keyword_rows:
     	for rank, (title,) in enumerate(keyword_rows, start=1):
             print(f"{rank}. {title}")
 	else:
         print("0 exact phrase matches")


 if __name__ == "__main__":
 	main()

Run the script

Run the comparison script.

python 07_semantic_vs_keyword.py

Expected output

The semantic search should return relevant documents even though the exact keyword phrase is not present.

Semantic query: Where should an AI app keep meaning-based search data?
 Keyword phrase: meaning-based search data

 Semantic results:
 1. Oracle AI Vector Search
    distance: <cosine distance from your run>
 2. Semantic Search for Applications
    distance: <cosine distance from your run>
 3. Python Database Code
    distance: <cosine distance from your run>

 Keyword results:
 0 exact phrase matches

Traditional SQL predicates are still important for filters, joins, permissions, and exact matching. Vector search adds a meaning-based ranking signal that we can combine with relational data when an application needs semantic retrieval.

Step 8: Add and validate a vector index

Finally, we’ll create a vector index and validate that Oracle Database reports it as a valid index owned by the demo schema.

This step demonstrates the index creation workflow; our tiny dataset is not large enough for meaningful performance conclusions.

Without an index, Oracle Database can compare the query vector with each stored vector exactly. That approach is suitable for tiny datasets and useful while learning.

With larger datasets, approximate vector indexes reduce the search space so queries can return quickly while still finding close neighbors.

Choose the index family after considering data size, latency goals, recall requirements, memory budget, and how often vectors change.

The query at the end confirms that semantic search still returns results after index creation.

We will not use EXPLAIN PLAN or DBMS_XPLAN here because this tutorial validates index creation with schema metadata, not query-plan behavior.

Create the script

Save the following code as 08_create_vector_index.py.

import array
 import os

 import oracledb
 from openai import OpenAI

 INDEX_NAME = "DOCUMENT_EMBEDDING_IDX"
 QUERY_TEXT = "Where should an AI app keep meaning-based search data?"
 ADMIN_USERS = {"SYS", "SYSTEM", "PDBADMIN"}


 def required_env(name):
 	value = os.getenv(name)
 	if not value:
         raise SystemExit(f"Set {name} before running this script.")
 	return value


 def main():
 	user = required_env("DB_USER")
 	if user.upper() in ADMIN_USERS:
         raise SystemExit(
             "Use the vector_demo application user, not an admin user."
     	)

     required_env("OPENAI_API_KEY")

 	# Generate the query vector used for the validation search.
 	client = OpenAI()
     response = client.embeddings.create(
         model=required_env("EMBEDDING_MODEL"),
         input=QUERY_TEXT,
 	)
     query_vector = array.array("f", response.data[0].embedding)

 	with oracledb.connect(
         user=user,
         password=required_env("DB_PASSWORD"),
         dsn=required_env("DB_DSN"),
 	) as connection:
         with connection.cursor() as cursor:
             # Remove a prior index so the script can be rerun safely.
             try:
                 cursor.execute(f"DROP INDEX {INDEX_NAME}")
             except oracledb.DatabaseError as exc:
                 error = exc.args[0]
                 if error.code != 1418:
                 	raise

             # Create an approximate neighbor-partitions vector index.
             cursor.execute(
                 """
                 CREATE VECTOR INDEX document_embedding_idx
                 ON documents (embedding)
                 ORGANIZATION NEIGHBOR PARTITIONS
                 DISTANCE COSINE
                 WITH TARGET ACCURACY 95
                 """
             )

             # Confirm that Oracle reports the vector index as valid.
             cursor.execute(
                 """
                 SELECT index_name, index_type, status
                 FROM user_indexes
                 WHERE index_name = :index_name
                 """,
                 index_name=INDEX_NAME,
             )
             index_row = cursor.fetchone()

             # Confirm that semantic search still returns rows after index creation.
             cursor.execute(
                 """
                 SELECT title
                 FROM documents
                 ORDER BY VECTOR_DISTANCE(embedding, :query_vector, COSINE)
                 FETCH EXACT FIRST 3 ROWS ONLY
                 """,
                 query_vector=query_vector,
             )
             rows = cursor.fetchall()

 	if not index_row:
         raise SystemExit("Vector index was not found in USER_INDEXES.")

     print("Created vector index DOCUMENT_EMBEDDING_IDX.")
     print("\nIndex validation:")
     print(f"Index name: {index_row[0]}")
     print(f"Index type: {index_row[1]}")
     print(f"Status: {index_row[2]}")
     result_count = len(rows)

 	print(
         "\nSemantic search still returns "
         f"{result_count} results after index creation."
 	)


 if __name__ == "__main__":
 	main()

Run the script

Run the index script.

python 08_create_vector_index.py

Expected output

The output should include a row for DOCUMENT_EMBEDDING_IDX, with INDEX_TYPE set to VECTOR and STATUS set to VALID.

Created vector index DOCUMENT_EMBEDDING_IDX.

 Index validation:
 Index name: DOCUMENT_EMBEDDING_IDX
 Index type: VECTOR
 Status: VALID

 Semantic search still returns 3 results after index creation.

USER_INDEXES confirms that the index exists and is valid. It does not prove that a specific query used the index, and this tutorial does not use execution-plan output as a validation method.

For production workloads, we would test with realistic data volume, query patterns, and performance goals.

Oracle AI Vector Search Tips and Troubleshooting

The database container is still starting

Wait until startup completes, then run docker logs oracle-free-26ai-vector again. The database must be ready before Python can connect to localhost:1521/FREEPDB1.

Port 1521 is already in use

Stop the conflicting local service or change the Docker port mapping. Update DB_DSN if the host port changes.

The database connection fails

Verify DB_USER, DB_PASSWORD, and DB_DSN. Use FREEPDB1 in the DSN, and connect application scripts as vector_demo, not SYS, SYSTEM, or PDBADMIN.

The embedding API key is missing

Steps 1-3 are database-only. Steps 4-8 require OPENAI_API_KEY and may incur API costs.

A vector dimension does not match

Run 04_check_embedding_dimension.py, update EMBEDDING_DIM, rerun 03_create_documents_table.py, and reload the embeddings with 05_insert_embeddings.py.

A vector bind fails

Ensure every FLOAT32 vector bind uses array.array("f", values). The tutorial uses this format for manual vectors, stored document embeddings, and query embeddings.

Semantic rankings differ from the example

This is expected. Embedding providers can update models, and floating-point distances can vary between runs.

Vector index creation fails

Confirm that the documents table exists, the vector_demo schema owns it, and DOCUMENT_EMBEDDING_IDX is not left over from a partially completed run.

Preparing for production

Use oracledb.create_pool() for connection pooling, benchmark realistic data volumes, review security and secret management, and tune indexes against real query patterns.

Stop and remove the container

When the tutorial is complete, stop and remove the local container.

docker stop oracle-free-26ai-vector
docker rm oracle-free-26ai-vector

Conclusion

We built a local semantic search workflow with Oracle AI Vector Search, Oracle Database Free 26ai, and Python.

We started with handwritten VECTOR(3, FLOAT32) values, inserted model-generated embeddings with array.array("f", values), queried similar documents with VECTOR_DISTANCE(), compared semantic retrieval with a simple exact phrase predicate, and validated a vector index through USER_INDEXES.

This approach fits well when embeddings belong beside relational data and we want SQL-visible storage and search without adding a separate vector database for the local workflow.

The demo is intentionally small: it teaches the primitives, not production scale. Larger applications should benchmark with realistic data, review security and privileges, use connection pooling, and validate indexing choices under real query patterns.

Next, we can deepen the implementation with these resources:

Oracle AI Vector Search User’s Guide is the main reference for vector storage, search, indexing, and related SQL capabilities.
Oracle Vector Data Type Documentation explains vector dimensions, element formats, and how VECTOR columns are defined.
CREATE VECTOR INDEX SQL Reference is useful when you are ready to explore vector-index syntax beyond this local path.
python-oracledb Vector Data Type Guide covers additional Python binding patterns for Oracle vector columns.
Oracle AI Vector Search LiveLabs provides hands-on workshops for vector embeddings, exact and approximate search, image search, and retrieval-augmented generation.

What is Oracle AI Vector Search?

Do I need an OpenAI API key to complete the tutorial?

Why must EMBEDDING_DIM match the embedding model?

Which distance metrics does the tutorial use?

Does a valid vector index prove that a query used the index?

Author

Mark Nelson

Mark Nelson is an architect and developer evangelist at Oracle. He works at the convergence of AI, microservices and database technologies. He is an active blogger, published author, technical reviewer at Manning Publications, Section Leader at Stanford Code in Place, and mentor at DeepLearning.ai. He regularly presents at Java and Oracle User Groups, AI meetups and large conferences. He has a passion for learning and teaching. He has over thirty years industry experience, at IBM and Oracle.

Sujets

Artificial Intelligence

Python

Top DataCamp Courses

Cursus

Associate AI Engineer pour développeurs

26 h

Apprenez à intégrer l'IA dans des applications logicielles en utilisant des API et des bibliothèques open source. Commencez dès aujourd'hui votre parcours pour devenir AI Engineer !

Afficher les détails

Commencer Le Cours

Cours

Introduction à Oracle SQL

4 h

18K

Renforcez vos compétences avec Oracle SQL, notamment les bases du SQL, l'agrégation, la combinaison et la personnalisation des données.

Afficher les détails

Commencer Le Cours

Cours

Créer des applications d'IA avec Pinecone

3 h

10.5K

Découvrez comment la base de données vectorielle Pinecone révolutionne le développement d'applications d'IA.

Afficher les détails

Commencer Le Cours

Contenus associés

blog

7 Best Vector Databases for AI in 2026: A Complete Guide

Compare the 7 best vector databases in 2026. Learn about embeddings, similarity search, and how to choose the right vector database for your AI applications.

Moez Ali

14 min

Tutoriel

How to Build a Vector Search Application with MongoDB Atlas and Python

Learn how to run your first MongoDB vector search. This tutorial walks you through finding similar items with embeddings and step-by-step examples.

Nilesh Soni

Tutoriel

How to Implement Semantic Search in MongoDB

Learn how to implement semantic search in MongoDB with Python. Generate vector embeddings, create a Vector Search index, and run $vectorSearch queries.

Damilola Oladele

Tutoriel

How to Store and Query Embeddings in MongoDB

Learn how to store, index, and query embeddings in MongoDB using Atlas Vector Search.

Nilesh Soni

Tutoriel

pgvector Tutorial: Integrate Vector Search into PostgreSQL

Discover how to enhance PostgreSQL with vector search capabilities using pgvector. This tutorial guides you through installation, basic operations, and integration with AI tools.

Moez Ali

code-along

Semantic Search with Pinecone

Learn the fundamentals of text embedding and vector databases with Pinecone to build a simple search engine.

James Briggs

Voir Plus Voir Plus

What Is Oracle AI Vector Search?

Key Features of Oracle AI Vector Search

Native vector data type

Distance functions and operators

Vector indexes

Integration with relational queries

How to Get Started With Oracle AI Vector Search

Prerequisites

Start Oracle Database Free 26ai locally

Create the vector_demo schema in FREEPDB1

Install Python dependencies

Set environment variables

Oracle AI Vector Search Demo Project

Step 1: Connect to the local database

Create the script

Run the script

Expected output

Step 2: Build intuition with manual vectors

Create the script

Run the script

Expected output

Step 3: Create the semantic documents table

Create the script

Run the script

Expected output

Step 4: Verify the embedding model dimension

Create the script

Run the script

Expected output

Step 5: Generate and insert document embeddings

Create the script

Run the script

Expected output

Step 6: Run a semantic search query

Create the script

Run the script

Expected output

Step 7: Compare semantic search with a phrase match

Create the script

Run the script

Expected output

Step 8: Add and validate a vector index

Create the script

Run the script

Expected output

Oracle AI Vector Search Tips and Troubleshooting

The database container is still starting

Port 1521 is already in use

The database connection fails

The embedding API key is missing

A vector dimension does not match

A vector bind fails

Semantic rankings differ from the example

Vector index creation fails

Preparing for production

Stop and remove the container

Conclusion

FAQs

Why must EMBEDDING_DIM match the embedding model?

Which distance metrics does the tutorial use?

Does a valid vector index prove that a query used the index?

7 Best Vector Databases for AI in 2026: A Complete Guide

How to Build a Vector Search Application with MongoDB Atlas and Python

How to Implement Semantic Search in MongoDB

How to Store and Query Embeddings in MongoDB

pgvector Tutorial: Integrate Vector Search into PostgreSQL

Semantic Search with Pinecone

.css-1531qan{-webkit-text-decoration:none;text-decoration:none;color:inherit;}Associate AI Engineer pour développeurs

Introduction à Oracle SQL

Créer des applications d'IA avec Pinecone

7 Best Vector Databases for AI in 2026: A Complete Guide

How to Build a Vector Search Application with MongoDB Atlas and Python

How to Implement Semantic Search in MongoDB

How to Store and Query Embeddings in MongoDB

pgvector Tutorial: Integrate Vector Search into PostgreSQL

Semantic Search with Pinecone

Associate AI Engineer pour développeurs