Security and governance in AI Factory
=====================================

AI Factory helps you build Gen AI and data‑driven applications on your
infrastructure with clear guardrails. This page explains who is
responsible for what, where governance is enforced, and how to apply
least‑privilege patterns across components.

Why security and governance matter
----------------------------------

Gen AI systems touch sensitive data, call powerful models, and may
execute actions. Proper controls reduce risk while keeping teams
productive:

- Protect sensitive data at rest and in flight

- Limit who can deploy models, change retrieval rules, or publish
  assistants

- Provide auditability and observability for compliance

- Enforce least privilege when assistants call tools or external
  services

See also: :ref:`Sovereign AI Explained <Sovereign AI Explained>`  .

Where enforcement happens
-------------------------

- Hybrid Manager (recommended): AI Factory is integrated into the Hybrid
  Manager control plane for governance and observability. Use the spoke
  for day‑to‑day operations:

- :ref:`AI Factory in Hybrid Manager </edb-postgres-ai/1.3/hybrid-manager/ai-factory>` 

- :ref:`Gen AI in Hybrid Manager </edb-postgres-ai/1.3/hybrid-manager/ai-factory/gen-ai>` 

- :ref:`Pipelines models reference <Pipelines models reference>` 

- Outside Hybrid Manager (selected components): When you run components
  such as :ref:`Pipelines and data preparation <Pipelines and data preparation>`  or :ref:`Vector Engine (PGvector) <Vector Engine (PGvector)>`  outside Hybrid Manager, you own
  cluster lifecycle, access control, and observability.

Who is responsible
------------------

- Platform teams: Control cluster access, secrets, and network policy;
  provision GPUs for :ref:`Model Serving Reference Manual <Model Serving Reference Manual>`  .

- Data engineers: Govern sources, :ref:`Knowledge Bases <Knowledge Bases>`  , and embedding
  freshness.

- Application/Gen AI builders: Define :ref:`Rulesets Manual <Rulesets Manual>`  , :ref:`Retrievers Reference Manual <Retrievers Reference Manual>`  ,
  :ref:`Tools Reference Manual <Tools Reference Manual>`  , and Assistants with least privilege.

How to apply governance in AI Factory
-------------------------------------

- Keep data in governed stores

- Use Postgres and object storage managed by your platform; control
  access with your standard policies.

- Use :ref:`Vector Engine (PGvector) <Vector Engine (PGvector)>`  to keep embeddings close to data and reuse existing
  database controls.

- Govern retrieval and behavior

- Express retrieval scope and limits with Retrievers and Knowledge
  Bases.

- Enforce acceptable behavior with Rulesets; external calls should be
  explicit through Tools.

- Secure model endpoints

- Expose private endpoints via :ref:`Model Serving Reference Manual <Model Serving Reference Manual>`  with controlled access.

- Right‑size GPU resources; see :ref:`How-To Update GPU Resources for an InferenceService <How-To Update GPU Resources for an InferenceService>`  .

- Observe and audit

- In Hybrid Manager, use integrated observability to monitor model
  serving, Gen AI runs, and pipelines.

- Treat Threads, logs, and metrics as audit artifacts as required by
  your policy.

Prerequisites
-------------

- Hybrid Manager: A project with AI Factory enabled and entitlements for
  Gen AI, Pipelines, Vector search, and Model Serving. See the
  :ref:`AI Factory spoke </edb-postgres-ai/1.3/hybrid-manager/ai-factory>`  .

- Outside Hybrid Manager: Ensure an enterprise secrets store, network
  policy, and audit logging are in place for the components you run.

Use cases
---------

- Regulated knowledge assistants: Use governed Knowledge Bases,
  constrained Retrievers, and Rulesets; run models via private
  endpoints.

- Document intelligence for PII: Keep content in Postgres/object
  storage, process via Pipelines, and restrict Tools to non‑destructive
  operations.

- Internal automation: Gate write‑actions behind Tools with explicit
  scopes; review Threads and logs periodically.

Related topics
--------------

- :ref:`Gen AI Builder Concepts (Assistants, Knowledge, Tools, Threads) <Gen AI Builder Concepts (Assistants, Knowledge, Tools, Threads)>` 

- :ref:`Sovereign AI Explained <Sovereign AI Explained>` 

- :ref:`Gen AI Builder <Gen AI Builder>` 

- :ref:`Pipelines and data preparation <Pipelines and data preparation>` 

- :ref:`Vector Engine (PGvector) <Vector Engine (PGvector)>` 

- :ref:`Model Serving Reference Manual <Model Serving Reference Manual>`