Skip to main content

Walkthrough

This walkthrough will guide you through a basic installation of the Oracle AI Optimizer and Toolkit (the AI Optimizer). It will allow you to experiment with GenAI, using Retrieval-Augmented Generation (RAG) and Natural Language to SQL (NL2SQL) with the Oracle AI Database at the core.

By the end of the walkthrough you will be familiar with:

  • Configuring a Language Model
  • Configuring an Embedding Model
  • Configuring the Vector Storage
  • Splitting, Embedding, and Storing vectors
  • Experimenting with the AI Optimizer

What you'll need for the walkthrough:

  • Internet Access (docker.io and container-registry.oracle.com)
  • Access to an environment where you can run container images (Podman or Docker).
  • 100G of free disk space.
  • 12G of usable memory.
  • Sufficient GPU/CPU resources to run the language model, embedding model, and database (see below).
Performance: A Word of Caution

The performance will vary depending on the infrastructure.

Language and Embedding Models are designed to use GPUs, but this walkthrough can work on machines with just CPUs; albeit much slower! When testing the Language Model, if you don't get a response in a couple of minutes; your hardware is not sufficient to continue with the walkthrough.

Installation​

You will run four container images to establish the "Infrastructure":

  • On-Premises Language Model - granite4.1:8b
  • On-Premises Embedding Model - mxbai-embed-large
  • Vector Storage - Oracle AI Database Free
  • The AI Optimizer

Language Model - granite4.1:8b​

To enable the ChatBot functionality, access to a Language Model is required. The walkthrough will use Ollama to run the granite4.1:8b model.

Same... but different!

The commands in this section work in a macOS/Linux shell or PowerShell 7. The documentation will reference podman commands. If applicable to your environment, podman can be substituted with docker.

If you are using docker, consider aliasing the podman command for convenience:

alias podman=docker

  1. Start the Ollama container:

    The Container Runtime is native. The command below makes all configured GPUs available; omit --gpus=all when running with CPUs only.

    podman run -d --gpus=all -v ollama:$HOME/.ollama -p 11434:11434 --name ollama docker.io/ollama/ollama

  2. Pull the Language Model into the container:

    podman exec -it ollama ollama pull granite4.1:8b

  3. Test the Language Model:

    Performance: Fail Fast...

    Unfortunately, if the below curl does not respond within 5-10 minutes, the rest of the walkthrough will be unbearable. If this is the case, please consider using different hardware.

    curl http://127.0.0.1:11434/api/generate -d '{
    "model": "granite4.1:8b",
    "prompt": "Why is the sky blue?",
    "stream": false
    }'

Embedding - mxbai-embed-large​

To enable the RAG functionality, access to an embedding model is required. The walkthrough will use Ollama to run the mxbai-embed-large embedding model.

  1. Pull the embedding model into the container:

    podman exec -it ollama ollama pull mxbai-embed-large

The AI Optimizer​

The AI Optimizer provides an easy-to-use front-end for experimenting with Language Model parameters and RAG.

  1. Download and extract the latest release into a new directory:

    Latest Release: TAR.GZ archive

    For example:

    curl -LO https://github.com/oracle/ai-optimizer/releases/latest/download/ai-optimizer-src.tar.gz
    mkdir ai-optimizer
    tar zxf ai-optimizer-src.tar.gz -C ai-optimizer

    cd ai-optimizer

  2. Build the container image.

    note

    macOS Silicon users may need to specify --arch amd64

    podman build -f src/Dockerfile -t localhost/ai-optimizer-aio:latest .

  1. Start the AI Optimizer:

    podman run -d --name ai-optimizer-aio --network=host localhost/ai-optimizer-aio:latest

Vector Storage - Oracle AI Database Free​

AI Vector Search in Oracle AI Database provides the ability to store and query private business data using a natural language interface. The AI Optimizer uses these capabilities to provide more accurate and relevant Language Model responses via Retrieval-Augmented Generation (RAG). Oracle AI Database Free provides an ideal, no-cost vector store for this walkthrough.

  1. Start the database container:

    podman run -d --name ai-optimizer-db -p 1521:1521 container-registry.oracle.com/database/free:latest-lite

  2. Wait until the database is ready:

    until podman logs ai-optimizer-db 2>&1 | grep -q 'DATABASE IS READY TO USE!'; do sleep 5; done

  3. Set vector_memory_size:

    podman exec -it ai-optimizer-db sqlplus '/ as sysdba'
    ALTER SYSTEM SET vector_memory_size=512M SCOPE=spfile;
    EXIT;

  4. Restart the database and wait for a new ready message:

    ready_lines=$(podman logs ai-optimizer-db 2>&1 | grep -c 'DATABASE IS READY TO USE!' || true)
    podman container restart ai-optimizer-db
    until [ "$(podman logs ai-optimizer-db 2>&1 | grep -c 'DATABASE IS READY TO USE!')" -gt "$ready_lines" ]; do sleep 5; done

  1. Create a database user for the walkthrough:

    podman exec -it ai-optimizer-db sqlplus '/ as sysdba'
    alter session set container=FREEPDB1;

    CREATE TABLESPACE IF NOT EXISTS USERS DATAFILE '/opt/oracle/oradata/FREE/FREEPDB1/users_01.dbf' SIZE 100M;
    CREATE USER "WALKTHROUGH" IDENTIFIED BY OrA_41_OpTIMIZER
       DEFAULT TABLESPACE "USERS"
       TEMPORARY TABLESPACE "TEMP";
    GRANT "DB_DEVELOPER_ROLE" TO "WALKTHROUGH";
    ALTER USER "WALKTHROUGH" DEFAULT ROLE ALL;
    ALTER USER "WALKTHROUGH" QUOTA UNLIMITED ON USERS;
    -- For Deep Data Security
    CREATE ROLE AIO_DDS_ROLE;
    GRANT CREATE SESSION TO AIO_DDS_ROLE;
    GRANT AIO_DDS_ROLE TO "WALKTHROUGH" WITH ADMIN OPTION;
    EXIT;

Configuration​

Operating System specific instructions:

If you are running on a remote host, you may need to allow access to the 8501 port.

For example, in Oracle Linux 8/9 with firewalld:

firewall-cmd --zone=public --add-port=8501/tcp

With the "Infrastructure" in-place, you're ready to configure the AI Optimizer.

In a web browser, navigate to http://localhost:8501: Chatbot

Notice that there are no language models configured to use. Let's start the configuration.

Configure the Language Model​

To configure the On-Premises Language Model, navigate to Configuration > Models:

  1. Enable the granite4.1:8b model that you pulled earlier by clicking the Edit button Configure Language Model
  2. Tick the Enabled checkbox, leave all other settings as-is, and Save Enable Language Model ⭐️ More information about configuring Language Models can be found in the Model Configuration documentation.

Say "Hello?"​

Navigate to the ChatBot screen:

Say Hello?

The error about language models will have disappeared, but there is a new warning about the database. You'll take care of that in the next steps.

The Chat model: will have been pre-set to the only enabled Language Model and a dialog box to interact with the Language Model will be ready for input.

Feel free to play around with the different Language Model Parameters, hovering over the ︖ icons to get more information on what they do.

You'll come back to the ChatBot later to experiment further.

Configure the Embedding Model​

To configure the On-Premises Embedding Model, navigate back to the Configuration screen and Models tab:

  1. Enable the mxbai-embed-large Embedding Model following the same process as you did for the Language Model. Configure Embedding Model

⭐️ More information about configuring embedding models can be found in the Model Configuration documentation.

Configure the Database​

To configure Oracle AI Database Free, navigate to the Configuration screen and Databases tab:

  1. Enter the Database Username: WALKTHROUGH
  2. Enter the Database Password for the database user: OrA_41_OpTIMIZER
  3. Enter the Database Connection String: //localhost:1521/FREEPDB1
  4. Save Database

Configure Database

⭐️ More information about configuring the database can be found in the Database Configuration documentation.

Split and Embed​

With the embedding model and database configured, you can now split and embed documents for use in Vector Search.

Navigate to Tools > Split/Embed:

  1. Change the Knowledge Base Source to Web Split and Embed

  2. Enter the URL and press Enter: Split and Embed Web Pages

    https://docs.oracle.com/en/database/oracle/oracle-database/26/xeinl/oracle-ai-database-free-installation-guide-linux.pdf

  3. Give the Vector Store an Alias: WALKTHROUGH and press Enter Split and Embed Populate

  4. Click Populate Vector Store

  5. Please be patient...

Performance: Grab a beverage of your choosing...

Depending on the infrastructure, the embedding process can take a few minutes. As long as the "Populating Vector Store..." timer is running... it's working.

You can watch the progress of the embedding by streaming the server logs:

podman exec -it ai-optimizer-aio tail -f /app/src/apiserver_8000.log

Query the Vector Store​

After the splitting and embedding process completes, you can query the Vector Store to see the chunked and embedded document:

From the command line:

  1. Connect to the Oracle AI Database:

    podman exec -it ai-optimizer-db sqlplus 'WALKTHROUGH/OrA_41_OpTIMIZER@FREEPDB1'

  2. Query the Vector Store:

    SELECT * FROM
     WALKTHROUGH_OLLAMA_MXBAI_EMBED_LARGE_512_103_COSINE_HNSW;

With the AI Optimizer configured, you're ready for some experimentation.

Navigate back to the ChatBot.

For this guided experiment, perform the following:

  1. Ask the ChatBot: 
    What are the required packages for a successful
    installation of an Oracle AI Database?

Responses may vary, but generally the ChatBot's response will be inaccurate, including:

  • Not understanding that there is an Oracle AI Database release. This is known as knowledge-cutoff.
  • Suggestions of requiring unrelated software. These are hallucinations.

Now select "Vector Search" in the Toolkit options, ensure "Store Discovery" and "Document Grading" are disabled, while "Prompt Rephrase" is enabled.

Enable RAG

and simply ask: Are you sure?

Performance: Host Overload...

With RAG enabled, all the services (Language/Embedding Models and Database) are being utilized simultaneously:

  • The Language Model is rephrasing "Are you sure?" into a query that takes into account the conversation history and context
  • The embedding model is being used to convert the rephrased query into vectors for a similarity search
  • The database is being queried for documentation chunks similar to the rephrased query (AI Vector Search)
  • The Language Model is completing its response using the documents from the database (if the documents are relevant)

Depending on your hardware, this may cause the response to be significantly delayed.

By asking Are you sure?, you are taking advantage of the AI Optimizer's history and context functionality. The response should be different and include a list of Operating System packages and maybe even an apology!

Under "Vector Search Details" you should see the PDF source, the vector store tables searched, and the rephrased query.

What's Next?​

You should now have a solid foundation using the AI Optimizer.

Try a full Use Case

The Use Cases walk you through using your new AI Optimizer installation through an end-to-end demo that shows the progressive value of NL2SQL, Vector Search, and combined-mode grounding a synthetic dataset.

To take your experiments further, consider exploring:

  • Turn On/Off/Clear history
  • Experiment with different Language Models and Embedding Models
  • Tweak Language Model parameters, including Temperature and Penalties, to fine-tune model performance
  • Investigate various strategies for splitting and embedding text data, such as adjusting chunk-sizes, overlaps, and distance metrics

Clean Up​

To cleanup the walkthrough "Infrastructure", stop and remove the containers.

podman container rm ai-optimizer-db --force
podman container rm ai-optimizer-aio --force
podman container rm ollama --force