Oracle AI Database Private Agent Factory: Enterprise FAQ

Which model providers are supported?

Private Agent Factory supports OCI Generative AI, OpenAI, Gemini, Ollama, and vLLM/self-hosted endpoints, depending on model type and configuration. New model providers are added in every release.

Which OCI Generative AI models are recommended?

Choice of model depends on the use case, cost, and response time requirements. We recommend testing with different models to identify the best fit.  

Does it support OpenAI models?

Yes. Custom base URL will be supported in future releases.  

Does it support Gemini?

Yes. Release 25.3.0.0.8 added Gemini LLM and embedding model support. Gemini LLMs can be connected using service account or API keys. 

Can it use locally hosted models?

Yes, Private Agent Factory comes bundled with local embedding model, multilingual-e5-base. Ollama is also supported for local hosting.

Can it use self-hosted vLLM endpoints?

Yes, vLLM self-hosted model endpoint can be configured using a model ID, URL, and port.  

What embedding models are supported?

Oracle lists local multilingual-e5-base (bundled), OCI Cohere embeddings such as cohere.embed-v4.0, vLLM/Ollama endpoints, and Gemini embedding models.  

Can I use smaller language models instead of frontier LLMs?

Yes. Smaller language models can be used through supported serving options such as vLLM, Ollama, OCI Generative AI, and OpenAI.  

What data can Knowledge Agents use?

Knowledge Agents can use configured and approved content from SharePoint, Google Drive, internal sites, uploaded files, file systems, and public web sources.  

How does Knowledge Agent ingestion work?

The ingestion path is crawling, parsing, storing, chunking, embedding, and vector database ingestion.  

What file types are supported?

Oracle’s file source page says .pdf, .txt, and .rtf files up to 1 GB each are supported.  

Can Knowledge Agent crawl websites?

Yes. Web sources can crawl public unauthenticated pages, with URL filters, crawl depth, crawl frequency, and proxy configuration.  

Can pre-built Data Analysis analyze structured data in Oracle Database?

Yes. Data Analysis Agents connect to Oracle Database 19c and later, understand schemas, generate SQL, and return explanations, tables, charts, and queries.  

Does it support MCP?

Yes. Private Agent Factory supports MCP tools, can attach multiple MCP servers to one agent, and supports Streamable HTTP transport with multiple authentication methods such as direct, bearer token, oAuth, Auth Request

Does it support REST and OpenAPI tools?

Yes. you can upload OpenAPI 2.0/3.0 JSON to auto-create tools and orchestrate SQL and REST in one flow. These are useful when connecting with Fusion Apps or other SaaS apps.

How does it protect enterprise data?

It protects data by running agents close to the database, supporting private/local model paths, and using Oracle Database security controls rather than relying only on application-layer filtering.  

Does data have to leave the enterprise boundary?

No. Oracle says Agent Factory can run in public clouds or on-premises without requiring customers to share data with third parties; model endpoints can also be private or local depending on architecture.  

Does it support SSO?

Yes. Oracle documents SSO support for Oracle IDCS, Google, Okta, Auth0, Microsoft Azure AD, and Amazon Cognito.  

Does it support role-based access?

Yes. Private Agent Factory provides three user roles: Chat-only Users, Editors, and Administrators.

Are answers grounded and traceable?

Yes for Knowledge Agents responses are grounded and traceable to enterprise-approved sources, with source links.  

How does Oracle position database-level security?

row-, column-, and cell-level controls, dynamic masking, SQL Firewall, and Deep Data Security so users and agents only see authorized data.  

Can it run in air-gapped environments?

Yes. it can operate in secure isolated environments.  

What should teams do for MCP security?

Use least-privilege tool design, authenticated MCP endpoints, scoped database users, and strong tool descriptions so the model selects only appropriate tools.  

Is there an additional Agent Factory license cost?

Private Agent Factory is included as a no-cost add-on to Oracle AI Database 26ai and is available at no additional cost to Oracle AI Database customers.  

What costs still apply?

Expect costs for database infrastructure, compute, storage, network, GPUs if used, OCI services, and external model/API usage.  

Does model choice affect cost?

Yes. OCI Generative AI, OpenAI, Gemini, local Ollama, and vLLM/self-hosted endpoints have different cost and operations profiles.  

Does Quickstart cost less than Production?

Not automatically. Quickstart lowers setup friction by using local containers, but production cost depends on your database, compute, model, storage, and network architecture.  

How should teams budget for it?

Budget by agent workload: number of users, prompts, tools, retrieval volume, embedding refresh frequency, model endpoint choice, and availability requirements.  

How do agents get published?

After testing, teams publish an agent and copy an Agent API Endpoint URL or SDK for external invocation. Agents can be invoked programmatically, business event triggers, embedded in customer apps such as APEX, Streamlit, Visual Builder, etc.

Can applications call agents through APIs?

Yes. Published Knowledge Agents, Data Analysis Agents, and Agent Builder flows can be called via REST-style POST requests.  

Does Agent Builder have memory?

Yes. Custom flows built in Agent Builder support conversational memory, and the Agent Builder page states the previous 10 queries or messages are retained as conversational history. Private Agent Factory will provide first-class support for Oracle Agent Memory in upcoming releases.  

Are Knowledge Agents conversational?

Not currently in the same way. Knowledge Agents are stateless and respond to one question at a time.   

Is observability and monitoring built in?

Private Agent Factory uses Oracle AI Database to store all configuration and agent conversation history and will support exporting Wayflow traces to OpenTelemetry compatible endpoints in next release.  

How do teams start hands-on?

Start with Oracle’s LiveLabs workshop, which covers installation, first agent creation, template bootstrapping, and Agent Builder customization.  

What should a first pilot prove?

A pilot should prove secure data access, answer traceability, model quality, cost per workflow, operational ownership, and whether the agent can be embedded into a real application or process.