WMS 12041: Build multimodal AI Vector Search using Oracle Private AI Service Container

WMS 12041: Build multimodal AI Vector Search using Oracle Private AI Service Container

Author: klazarz

private-ai-services-container/introduction/images/arch.png

@@ -0,0 +1,59 @@
+# Introduction
+
+## About this Workshop
+
+Welcome to **Build multimodal AI Vector Search using Oracle Private AI Service Container**.
+
+Oracle Private AI Services Container exists to give you modern model inference inside your own environment. You get a local model endpoint without sending your text or images to a public AI service.
+
+Its core value is control with flexibility:
+- Keep **data inside your network** and security boundary
+- Update model services **without changing core database** deployment
+- **Reuse one model endpoint** across notebooks, SQL flows, and apps
+- **Support multimodal patterns**, such as image and text embeddings in one solution
+
+Use each path for a different job:
+- **In-database embeddings (`provider=database`)** fit SQL-first workflows with minimal moving parts.
+- **Private AI Services Container (`provider=privateai`)** fits teams that need model agility, multimodal use cases, or shared model serving across tools.
+
+Compared with public embedding APIs, a private container is often the stronger enterprise choice:
+- Sensitive data does not leave your environment
+- Latency and cost are more predictable on local network paths
+- Development is less exposed to external quotas, endpoint drift, and service outages
+
+In the following labs you will work not only with in-database embedding but **specifically** with the Oracle Private AI Services Container to:
+- discover available models in the Oracle Private AI Services Container
+- generate embeddings using ONXX models stored in the Oracle AI Database and via the API endpoint provided by the Oracle Private AI Services Container.
+- store vectors in Oracle AI Database
+- run cosine similarity search
+- build a simple image search app that used multimodal embedding models
+
+Estimated Workshop Time: 90 minutes
+
+### Architecture at a Glance
+
+- `jupyterlab` runs Python notebooks.
+- `privateai` serves embedding models at `http://privateai:8080` on the container network.
+- `aidbfree` stores documents and vectors.
+
+![architecture](./images/arch.png)
+
+
+### Objectives
+
+In this workshop, you will:
+- Validate the runtime services required for the lab
+- Generate embeddings with both database-stored ONNX models and Oracle Private AI Services Container
+- Perform semantic similarity search in Oracle AI Database 26ai
+- Build a simple image app that uses multimodal embeddings for similarity search
+
+
+## Learn More
+
+- [Oracle Private AI Services Container User Guide](https://docs.oracle.com/en/database/oracle/oracle-database/26/prvai/oracle-private-ai-services-container.html)
+- [Private AI Services Container API Reference](https://docs.oracle.com/en/database/oracle/oracle-database/26/prvai/private-ai-services-container-api-reference.html)
+- [DBMS_VECTOR UTL_TO_EMBEDDING](https://docs.oracle.com/en/database/oracle/oracle-database/26/vecse/utl_to_embedding-and-utl_to_embeddings-dbms_vector.html)
+
+## Acknowledgements
+- **Author** - Oracle LiveLabs Team
+- **Last Updated By/Date** - Oracle LiveLabs Team, March 2026

private-ai-services-container/introduction/introduction.md

@@ -0,0 +1,57 @@
+# Lab 1: Verify the Runtime Environment
+
+## Introduction
+
+In this lab you verify that all the Oracle Private AI Services API endpoint is reachable on the netwokk.
+You will also learn how to list all the models available in the container.
+All checks are executed from a **JupyterLab Terminal**.
+
+Estimated Time: 10 minutes
+
+### Objectives
+
+In this lab, you will:
+- Verify internal container DNS resolution from JupyterLab
+- Validate Private AI health and model list
+- Confirm Oracle AI Database and ORDS reachability from JupyterLab
+- Confirm `/home/.env` is available to the notebook
+
+### Prerequisites
+
+This lab assumes:
+- You can open a terminal in JupyterLab (`File` -> `New` -> `Terminal`)
+
+## Task 1: Verify Internal Hostname Resolution
+
+1. In JupyterLab, open a new terminal.
+
+2. Verify that runtime service names resolve:
+
+    ```bash
+    <copy>getent hosts privateai aidbfree ords</copy>
+    ```
+
+    Expected: one IP entry for each service name.
+
+## Task 2: Validate Private AI REST Endpoints and list available models
+
+1. Health endpoint:
+
+    ```bash
+    <copy>curl -sS -i http://privateai:8080/health</copy>
+    ```
+
+    Expected: HTTP `200`.
+
+2. List deployed models:
+
+    ```bash
+    <copy>curl -sS http://privateai:8080/v1/models | jq .</copy>
+    ```
+
+    Expected: JSON payload with a `data` array of model IDs.
+
+
+## Acknowledgements
+- **Author** - Oracle LiveLabs Team
+- **Last Updated By/Date** - Oracle LiveLabs Team, March 2026

private-ai-services-container/lab1-verify-environment/lab1-verify-environment.md

@@ -0,0 +1,259 @@
+{
+  "cells": [
+    {
+      "cell_type": "markdown",
+      "metadata": {},
+      "source": [
+        "# Notebook B: Database-Stored Model Embeddings\n",
+        "\n",
+        "This notebook demonstrates vector search using embedding models already stored in Oracle AI Database (for example `ALL_MINILM_L12_V2`).\n",
+        "\n",
+        "Flow:\n",
+        "- discover available database models\n",
+        "- generate embeddings with `provider=database`\n",
+        "- store vectors and run similarity search\n"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {},
+      "outputs": [],
+      "source": [
+        "import os\n",
+        "import json\n",
+        "import re\n",
+        "import oracledb\n",
+        "from dotenv import dotenv_values\n"
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {},
+      "source": [
+        "## 1) Load DB configuration from `/home/.env`"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {},
+      "outputs": [],
+      "source": [
+        "ENV_PATH = os.getenv('LAB_ENV_FILE', '/home/.env')\n",
+        "env = dotenv_values(ENV_PATH) if os.path.exists(ENV_PATH) else {}\n",
+        "\n",
+        "DB_USER = os.getenv('DB_USER') or env.get('USERNAME') or 'ADMIN'\n",
+        "DB_PASSWORD = os.getenv('DB_PASSWORD') or env.get('DBPASSWORD')\n",
+        "DB_HOST = os.getenv('DB_HOST', 'aidbfree')\n",
+        "DB_PORT = os.getenv('DB_PORT', '1521')\n",
+        "DB_SERVICE = os.getenv('DB_SERVICE', 'FREEPDB1')\n",
+        "DB_DSN = os.getenv('DB_DSN', f'{DB_HOST}:{DB_PORT}/{DB_SERVICE}')\n",
+        "PREFERRED_DB_MODEL = os.getenv('DB_EMBED_MODEL', 'ALL_MINILM_L12_V2').upper()\n",
+        "\n",
+        "print('ENV file:', ENV_PATH if os.path.exists(ENV_PATH) else 'not found')\n",
+        "print('DB user:', DB_USER)\n",
+        "print('DB dsn :', DB_DSN)\n",
+        "print('Preferred DB model:', PREFERRED_DB_MODEL)\n",
+        "\n",
+        "if not DB_PASSWORD:\n",
+        "    raise ValueError('DB password not found. Set DB_PASSWORD or DBPASSWORD in /home/.env')\n"
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {},
+      "source": [
+        "## 2) Connect and discover stored models"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {},
+      "outputs": [],
+      "source": [
+        "conn = oracledb.connect(user=DB_USER, password=DB_PASSWORD, dsn=DB_DSN)\n",
+        "cur = conn.cursor()\n",
+        "\n",
+        "cur.execute('select user from dual')\n",
+        "print('Connected as:', cur.fetchone()[0])\n",
+        "\n",
+        "cur.execute('''\n",
+        "    SELECT model_name\n",
+        "    FROM user_mining_models\n",
+        "    ORDER BY model_name\n",
+        "''')\n",
+        "models = [row[0] for row in cur.fetchall()]\n",
+        "print('Models in USER_MINING_MODELS:', models)\n",
+        "\n",
+        "if not models:\n",
+        "    raise RuntimeError('No models found in USER_MINING_MODELS. Provision an ONNX model first.')\n",
+        "\n",
+        "MODEL_NAME = PREFERRED_DB_MODEL if PREFERRED_DB_MODEL in models else models[0]\n",
+        "print('Selected DB model:', MODEL_NAME)\n",
+        "\n",
+        "if not re.match(r'^[A-Z][A-Z0-9_$#]*$', MODEL_NAME):\n",
+        "    raise ValueError(f'Unsafe model identifier: {MODEL_NAME}')\n"
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {},
+      "source": [
+        "## 3) Determine embedding dimension from the DB model"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {},
+      "outputs": [],
+      "source": [
+        "db_params = json.dumps({\n",
+        "    'provider': 'database',\n",
+        "    'model': MODEL_NAME,\n",
+        "})\n",
+        "\n",
+        "cur.execute('''\n",
+        "    SELECT VECTOR_DIMENSION_COUNT(\n",
+        "        DBMS_VECTOR.UTL_TO_EMBEDDING(:txt, JSON(:params))\n",
+        "    )\n",
+        "    FROM dual\n",
+        "''', {'txt': 'dimension probe', 'params': db_params})\n",
+        "\n",
+        "EMBEDDING_DIM = int(cur.fetchone()[0])\n",
+        "print('Embedding dimension:', EMBEDDING_DIM)\n"
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {},
+      "source": [
+        "## 4) Create table and store in-database embeddings"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {},
+      "outputs": [],
+      "source": [
+        "TABLE_NAME = 'PRIVATEAI_DOCS_DBMODEL'\n",
+        "\n",
+        "try:\n",
+        "    cur.execute(f'DROP TABLE {TABLE_NAME} PURGE')\n",
+        "except oracledb.DatabaseError:\n",
+        "    pass\n",
+        "\n",
+        "cur.execute(f'''\n",
+        "    CREATE TABLE {TABLE_NAME} (\n",
+        "        doc_id NUMBER GENERATED ALWAYS AS IDENTITY PRIMARY KEY,\n",
+        "        title VARCHAR2(200) NOT NULL,\n",
+        "        content CLOB NOT NULL,\n",
+        "        embedding VECTOR({EMBEDDING_DIM}, FLOAT32)\n",
+        "    )\n",
+        "''')\n",
+        "\n",
+        "docs = [\n",
+        "    ('Database Model Path', 'Embeddings can be generated directly in Oracle AI Database with a stored ONNX model.'),\n",
+        "    ('Operational Simplicity', 'Keeping model inference in-database can simplify architecture and governance.'),\n",
+        "    ('Vector Search SQL', 'Use VECTOR_DISTANCE with COSINE to rank semantic similarity results.'),\n",
+        "    ('Model Governance', 'Database-stored models can be versioned and controlled with DB privileges.'),\n",
+        "]\n",
+        "\n",
+        "inserted = 0\n",
+        "for title, content in docs:\n",
+        "    cur.execute(f'''\n",
+        "        INSERT INTO {TABLE_NAME} (title, content, embedding)\n",
+        "        VALUES (\n",
+        "            :title,\n",
+        "            :content,\n",
+        "            DBMS_VECTOR.UTL_TO_EMBEDDING(:content, JSON(:params))\n",
+        "        )\n",
+        "    ''', {'title': title, 'content': content, 'params': db_params})\n",
+        "    inserted += 1\n",
+        "\n",
+        "conn.commit()\n",
+        "print('Inserted rows:', inserted)\n"
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {},
+      "source": [
+        "## 5) Similarity search with the database model"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {},
+      "outputs": [],
+      "source": [
+        "query_text = 'Which approach keeps embedding generation inside Oracle Database?'\n",
+        "\n",
+        "sql = f'''\n",
+        "SELECT\n",
+        "    title,\n",
+        "    ROUND(1 - VECTOR_DISTANCE(\n",
+        "        embedding,\n",
+        "        DBMS_VECTOR.UTL_TO_EMBEDDING(:query_text, JSON(:params)),\n",
+        "        COSINE\n",
+        "    ), 4) AS similarity,\n",
+        "    SUBSTR(content, 1, 120) AS preview\n",
+        "FROM {TABLE_NAME}\n",
+        "ORDER BY VECTOR_DISTANCE(\n",
+        "    embedding,\n",
+        "    DBMS_VECTOR.UTL_TO_EMBEDDING(:query_text, JSON(:params)),\n",
+        "    COSINE\n",
+        ")\n",
+        "FETCH FIRST 3 ROWS ONLY\n",
+        "'''\n",
+        "\n",
+        "cur.execute(sql, {'query_text': query_text, 'params': db_params})\n",
+        "rows = cur.fetchall()\n",
+        "\n",
+        "print('Query:', query_text)\n",
+        "for idx, (title, score, preview) in enumerate(rows, 1):\n",
+        "    print(f'{idx}. {title} | score={score}')\n",
+        "    print(f'   {preview}')\n"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {},
+      "outputs": [],
+      "source": [
+        "# Optional cleanup\n",
+        "# cur.execute(f'DROP TABLE {TABLE_NAME} PURGE')\n",
+        "# conn.commit()\n"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {},
+      "outputs": [],
+      "source": [
+        "cur.close()\n",
+        "conn.close()\n",
+        "print('Connection closed.')\n"
+      ]
+    }
+  ],
+  "metadata": {
+    "kernelspec": {
+      "display_name": "Python 3",
+      "language": "python",
+      "name": "python3"
+    },
+    "language_info": {
+      "name": "python",
+      "version": "3.11"
+    }
+  },
+  "nbformat": 4,
+  "nbformat_minor": 5
+}