Installing the AI Pack

The AI Pack is an optional, separately licensed add-on. It adds retrieval augmented generation over the documents already in ADM: text extraction, chunking, embeddings, semantic search, generated answers with sources, and read-only tools your own AI agents can call.

It is installed on top of a working base installation and depends on it.

What it adds

RAG collections	A subset of your documents — by folder or by annotation — kept in sync, chunked and embedded automatically.
Search and answers	search_chunks, prepare_sources and generate_answer, with optional query rewriting and hybrid keyword + vector search.
Agent tools	Six read-only tools, pre-registered with uc_ai, that let an LLM agent browse and read ADM under a user’s own permissions.
LLM text extraction	A vision model as a fallback for documents Oracle’s text filter reads badly, such as scanned PDFs.
An hourly job	ADM_AI_HOURLY_JOB, which syncs every collection.
Admin pages	RAG Collections, collection file details, and an answer-debug view.

Prerequisites

• The base product, installed and working. The AI Pack extends it and will not install on its own.
• uc_ai, the United Codes AI library, installed in the same schema. It is what talks to the LLM providers.
• A vector store, one of:
  • Oracle AI Vector Search — native, on Oracle 23ai or later. Nothing external to run.
  • Qdrant — external, free and open-source, run from its Docker image. The option for databases without native vector support.
• An embedding provider, and — if you want generated answers, query rewriting or LLM text extraction — a chat model provider. Both are reached through APEX Web Credentials.
• Outbound network access from the database to the vector store and the provider, which on most installations means a network ACL and a proxy.

Installing

1. Run the AI install script as the ADM schema owner, after install.sql:

   sqlplus adm/password@db @install_ai.sql

   It logs to install_adm_ai.log. As with the base install, the script continues past errors, so read the log:

   grep -iE "ORA-|PLS-" install_adm_ai.log

2. Recompile adm_oci with the DBMS_CLOUD flag if you use object storage. The AI Pack reads file content, and on an object-storage installation that goes through DBMS_CLOUD:

   alter package adm_oci compile plsql_ccflags = 'DBMS_CLOUD:TRUE';

   Note

   On a database where DBMS_CLOUD is not available, leave adm_oci as it is. It stays invalid, which is expected; only object storage needs it.

3. Create the APEX Web Credentials for your embedding and chat providers, and for Qdrant if you use it.

4. Point the settings at them. The AI settings are documented in the settings reference; the ones every installation sets are AI_QDRANT_URL and AI_QDRANT_WEB_CREDENTIAL_ID (Qdrant only) and AI_CHUNK_FILE_LIMIT. If you will use the bundled agents, set AI_AGENT_PROVIDER, AI_AGENT_MODEL and AI_AGENT_WEB_CREDENTIAL_ID too — they ship empty, and filling them in points both shipped prompt profiles at your provider instead of the OpenAI model they were created with. See Choosing the provider and model.

5. Check the hourly job exists and is enabled:

   select job_name, enabled, state, next_run_date
     from user_scheduler_jobs
    where job_name = 'ADM_AI_HOURLY_JOB';

   Without it, collections never sync. See Jobs and maintenance.

6. Create your first collection and watch it sync: RAG configurations.

Text extraction has to be working first

RAG reads the indexed text of your documents, not the raw files. If the base product’s indexing queue is not being worked, or OBJECT_STORAGE_INDEX is off for object-storage files, collections will sync and produce nothing. Confirm content search works before debugging embeddings — see how content becomes searchable.

Upgrading

Run the new release’s install_ai.sql after its install.sql, never before. See Upgrading.

Also refer to

• RAG configurations
• AI agent tools
• Settings reference
• Installation — the base product