Summary
How can you explore five ways to query tables, documents, audio and images from one application?
Oracle AI Accelerator combines Oracle Autonomous Database, files in OCI Object Storage and OCI Generative AI models. The walkthrough prepares the environment, connection, credentials and users with their modules before starting the tests.
The five demos cover Select AI over a Titanic CSV file; Select AI with RAG over HTML and TXT files; OCR and PDF questions with OCI Document Understanding; audio transcription to SRT with OCI Speech; and JPG document processing and image queries with a multimodal model.
The transcribed audio is then analyzed with a generative model to summarize conversations, answer questions and explain sentiment analysis. The walkthrough also reviews generated SQL, column comments, response sources, metadata and embeddings stored in Autonomous, along with application history and metrics.
Start
The walkthrough introduces environment preparation before running five Oracle AI Accelerator demos. The application brings together table queries, RAG, document processing, audio transcription and multimodal interaction.
Due and connection to Autonomous
A virtual machine, an OCI Object Storage bucket and an Autonomous database are prepared. The wallet is downloaded from the connection options, with a password to protect it. The file is placed in the application folder, and connection parameters are completed with the service values used by the example.
OCI Credentials and library installation
An API Key is generated to access the tenancy, and the key file and corresponding configuration are downloaded. These files are placed in the OCI configuration folder of the user profile. The libraries needed to communicate with OCI services and Oracle Database are then installed, allowing environment preparation to finish.
Database and implementation initiation
The initialization script checks the Conda installation and prepares the example components. It creates tables, inserts information and sets up stored procedures that the application uses to process data. The walkthrough follows the tasks running in the application and database while waiting; the interface opens when initialization finishes.
Access to and review of implementation
The demonstration administrator signs in to the application and opens the development schema in Autonomous. Table creation is checked, and the user table is reviewed for access and module settings. This inspection compares initialization with the available records before another user is created.
User creation and module empowerment
A demonstration user is created with the Select AI, RAG, OCI Document Understanding, OCI Speech and multimodal AI modules enabled. Provisioning creates the required settings and profiles. The new record and database user are then checked, before any tables have been loaded for that user. The administrator signs out, and the demonstration user signs in.
1 Demo: Select AI and CSV Titanic Loading
The first demo loads a CSV file containing Titanic passengers. Column comments give the model used by Select AI context about the fields. The walkthrough shows the flow between the application, Autonomous and the bucket, runs the load and reviews processing time and character counts. It also checks the CSV stored in Object Storage as a backup for the example.
Validation of the table and its comments
The schema view is refreshed in Autonomous to check that the Titanic table was created. Its contents, structure, data types and column comments entered through the application are inspected. This validation connects the uploaded file with the actual table used for Select AI questions.
Natural-language consultations with Select AI
The Select AI chat opens with a JSON view of the table structure and contents. The test asks how many survivors appear in the Titanic dataset; this particular load returns 342 people and shows the generated query. The example follows a natural language question through to its result over the imported data.
SQL and chat export
The walkthrough reviews response narration, SQL display and query explanation. A second question asks how many passengers traveled in first class; the example dataset returns 216. Conversation export and the application time and character metrics are then shown for reviewing the queries.
2 Demo: Select AI with RAG about HTML
The second demo saves a web page as an HTML file and loads it into the Select AI with RAG module. Processing results and the file record are checked alongside the earlier CSV. The chat then asks about the benefits of Select AI with RAG, using the newly ingested page as reference content.
Loading of TXT and responses with sources
A TXT file containing Oracle information is added to the content already available for RAG. After loading finishes, a general Oracle question is asked. The response includes retrieved information and its source file, along with processing time. The walkthrough shows how to extend the assistant sources while keeping the same chat interaction.
Summaries with RAG and Vector storage
A summary of Oracle and Select AI is requested using the loaded files. The presenter connects the bucket files with their vector storage in Autonomous, where content is retrieved to answer questions. The conversation can also be saved for later analysis. The example focuses on querying semistructured information through RAG.
3 Demo: OCR with OCI Document Understanding
The initially empty document table and its embeddings field are inspected before a PDF is loaded. The application sends the file to the bucket and creates an OCI Document Understanding job to extract text using OCR. Processing time and extracted character counts are then reviewed, preparing content for the next query step.
Embeddings and consultations about the document
The document table is checked for the extracted text, metadata and embeddings stored by processing. A generative model is then selected to ask a question about the PDF. The demo uses a Meta Llama model and keeps the question and answer history for continuing the document conversation.
Demo 4: Audio transcription with OCI Speech
An MP3 file is uploaded for transcription through OCI Speech, with SRT output. The application records the file and prepares its content for later queries. Once processing finishes, the transcript is inspected, and the recording and agent are selected for analyzing the text in the chat.
The analysis of conversation and feeling
The prompt of an agent configured to analyze SRT transcripts is reviewed. A generative model summarizes the conversation, answers questions and changes the presentation of the operator and customer dialogue. Sentiment analysis and its explanation are then requested using the transcript words. The application also shows token metrics for the exchange.
Transcripts and embeddings in Autonomous
The file table is queried to locate the processed MP3 and stored conversation. The presenter connects embeddings generation with DBMS_VECTOR_CHAIN inside the database. The document table is then checked for the transcript text and its vector representation, which prepare searches and questions over this content.
Audio recording from the application
Alongside file uploads, the application recording control captures a voice test. When recording ends, an OCI Speech job produces an SRT transcript and stores the result in Autonomous. The latest record is located, and the recording is selected in the chat for summaries or questions using a generative model.
5 Demo: OCR and Multimodal AI consultations
The multimodal demo ingests a scanned JPG document. A generative model acts as an intermediary for extraction, using an agent configured for OCR. The content is processed and stored in the Autonomous vector store. Once loading finishes, the multimodal module opens and the document is selected for queries.
Questions about documents and images
A question asks what the document is about and returns an excerpt of its contents. An image is then uploaded for visual information queries and a summary. Another test uses an apple image and asks about its colors; the response identifies red, green and yellow. The walkthrough shows multimodal interaction with text and visual information.
Prompts agents and closure
The walkthrough reviews prompts configured for different inputs: SRT transcripts, multimodal files and text extraction with Markdown output. The presenter describes these settings as a way to adapt interaction with the information. The closing connects Object Storage files with data stored in Autonomous and concludes the five demos.
Resources related
- Oracle AI Accelerator - code and installation guides
- Oracle Select AI - Consultation in Natural Language
- Select AI with RAG - Search about your documents
- OCI Document Understanding - Extraction and OCR
- OCI Speech - Audio Transcription
- OCI Generative AI - Models and capabilities
- DBMS_VECTOR_CHAIN - Embeddings and vector processing