Tutorials

Build shared upload rules, document workflows and AI features with ADM. Each lesson adds to the same working example. For individual API topics, use the developer guides.

The SQL scripts are in the public ADM repository and in the examples folder of the delivery archive. Recorded output comes from real runs.

At a glance

Build workflows on hooks →Register shared upload rules, add category and duplicate tags, and create standard project folders. Six lessons with runnable SQL examples.

Classify uploads with an LLM →Let a language model sort new PDFs into document types, measure it against labelled files, review what it holds back, and classify the documents that already exist. Five lessons.

Ask questions across your documents →Index ADM documents in RAG collections, build an agent that searches them, put it in an APEX chat page, check that each user gets only what ADM lets them see, measure what the search settings cost, and search the collections from the agent of another schema. Seven lessons.

Extract data from PDF forms →Configure the form extraction of the AI Pack for invoices, delivery notes and images, run it with a language model of your choice, check every result, and send only the failures to a person. Six lessons.

The courses in detail

Build workflows on hooks

An ADM hook applies the same rule across upload paths: the ADM APEX app, REST API, and applications using ADM’s document API. A quick validation can refuse an upload immediately. For shared analysis that can finish later, the hook queues work for a background worker. If one application needs analysis before its next step, it can call that analysis directly in its own upload flow.

You build three useful results without changing ADM’s packages:

• Refuse blocked file types and empty files, and keep a record of the refusal.
• Add category tags and flag files with matching checksums as duplicate-of. The sample uses a 12-byte text file to demonstrate the match.
• Give every new project the same starting folders, regardless of which application creates it.

The queue supports these workflows: it lets uploads finish before processing, keeps pending work, and allows failed processing to be retried. You choose when to use it before building it. The course also shows how to start the worker automatically.

Use a test schema with ADM installed. The recordings use Oracle Database 26ai Free; the scripts target 19c or later, but execution on 19c has not been verified.

The course needs two privileges and a free hook slot

A DBA must grant your schema execute on sys.dbms_aq and sys.dbms_aqadm. ADM allows one snippet for each hook event, and the course takes the slots, so it cannot run next to a hook you already registered. 99_teardown.sql frees them. Lesson 1 makes sure that both are in place before it changes anything.

Start: Refuse an upload from a hookPrepare the schema, register a hook that refuses blocked file types and empty files, and show the user the reason it gives.

Classify uploads with an LLM

This course adds a workflow to the hooks demo. A language model reads each new PDF or image and decides its document type. The workflow writes the type into ADM as a tag and as annotations, and it keeps the model’s reason in a table.

You build these results:

• Classify a new upload from a hook, without making the upload wait for the model.
• Measure the classifier against labelled files, find a wrong answer that took effect on its own, and correct it by changing a type description.
• Keep what a person decided when a document is classified again, classify a new version, and record a file that the model cannot read.
• Review the verdicts that wait for a person, accept them with the model’s type or another type, and take a release back.
• Estimate the tokens of a backfill and classify the documents that were in ADM before the classifier.

The course covers the ADM side. It links to the UC AI tutorials for how structured output works. Five lessons, about 115 minutes.

The course needs the hooks demo, the AI Pack and a provider key

The course installs the hooks demo itself, so you do not need to complete the hooks course first. It takes the same hook slots as the hooks course. It needs the AI Pack and an APEX web credential named OPENAI that holds an OpenAI key. Lesson 1 lists the setup, and keeps the steps that only some installations need in boxes: the DBA grants, the credential, and network access. After the install, every upload in the schema goes to your provider, so use a test schema. An optional section of lesson 4 needs Oracle 23ai for the native vector store.

Start: Classify a new upload with a language modelCheck the prerequisites, install the classifier, upload an invoice, and read the tag and annotations the model produced.

Ask questions across your documents

This course builds a chat page that answers from the contracts and policies of a sample company. The answers come from RAG collections of ADM documents, and every passage passes the ADM access check of the signed-in user.

You build these results:

• Index two groups of documents in RAG collections and ask one of them a question.
• Register a search tool for each collection, and run an agent that picks the collection and answers follow-up questions.
• Put the agent in an APEX page with the UC AI Chat plug-in, and hand the signed-in user to ADM.
• Show with two users that a user gets only the passages of documents that ADM lets them see, why a restricted user sees “nothing found”, and what the trace of a run shows.
• Measure retrieval and answers with a question list, store the score threshold that the measurement supports, and make sure that the stored search settings reach the agent.
• Let an agent in the schema of another application search the collections through one granted package that searches as a fixed ADM user.

The course covers the ADM side: collections, access control and identity. It links to the UC AI tutorials for how agents, tools and the chat region work. Seven lessons, about 3 hours and 30 minutes.

The course needs Oracle 23ai and the AI Pack

The collections use the native vector store of Oracle Database 23ai. You need the AI Pack and an APEX web credential named OPENAI that holds an OpenAI key. The chat page in lesson 4 needs APEX 26.1 or later to import the application. The scripts create two ADM users and two APEX accounts with a demo password: use a test schema. Lesson 7 needs a second schema with UC AI and an APEX application, in the APEX workspace of the ADM schema. If you have none, a DBA creates the demo schema with a script of the lesson. Lesson 1 lists the setup.

Start: Answer a question from a RAG collectionInstall the sample company, index its documents, and ask a collection a question.

Extract data from PDF forms

This course configures the form extraction of the AI Pack. ADM reads the values of PDF forms and images into ADM annotations and into a table of your application. Each form is configuration, not code: markers that find it, rules for the prompt, field patterns, and check expressions on the values. You can configure it in the form extraction manager or with SQL. The course uses OpenAI, and the model is a setting: you can use another provider, including an Ollama server that you operate.

You build these results:

• Configure one form and extract its fields and its line table, with a line pattern for long tables.
• Add check expressions, measure a form against a test set, and count the wrong results that pass your checks.
• Add a form and a document type, and make sure that no two forms match one document.
• Read images with a vision model, choose the models with a measurement, and size the context for an Ollama server that you operate.
• Let ADM extract the new documents of an inbox folder in the background, copy the accepted invoices into your own table with a hook, approve the documents that need a person, and watch the review rate.
• Read a summary of a test with five other document types: what carries over from invoices, and which layouts need the file setting.

Six lessons, about two and a half hours. Lesson 6 is a short read with no steps to run.

The course needs the AI Pack and a model

You need the AI Pack and a language model that the database can call, for example OpenAI with an APEX web credential. Lesson 1 lists the setup.

Start: Extract the data of one formLoad the sample documents, configure the first form, and extract one invoice with a language model.