AI Model Platform

Open AI → Model Platform to manage the org-scoped LLM gateway: provider catalog, model approval, routing rules, feature bindings, custom prompts, document libraries (RAG), runtime policies, quotas, and usage analytics.

ⓘ Requires ai_gateway:manage. Connect credentials under Administration → Integrations → AI Providers before models become available to theAI Assistant and other gateway consumers.

Architecture

▥ Assistant (copilot) · Intelligence · Playground · other gateway features

↓

AiGatewayService.complete() / embed() / ragQuery()

↓

▣

Quota enforcement

Org spend and rate limits before routing

↝

Route select

Explicit model → feature binding → route rules → cost/latency → default approved model

♧

Runtime policies

Require approved models, deny by feature/task/provider, security rule signals

♢

Security scan + redact

Secrets, PII redaction; injection and moderation blocked on user turns

▤

RAG + memory

Knowledge-base retrieval and scoped conversation memory

♆

Provider adapter

OpenAI-compatible and vendor-specific APIs with failover

▥

AiRequestLog

Tokens, cost, latency, status, security flags, feedback

Dashboard & tabs

A Model Health dashboard (availability, utilization charts, token/cost KPIs) appears above the tabs on every visit. On first open, Axio bootstraps the built-in provider and model catalog for the organization when none exists yet.

◔

Overview

Configured providers, model health, request/token/cost KPIs, success rate, prompt count

⌂

Providers

Configured integrations (from AI Providers), health check, credential storage

◇

Models

Chat and multimodal models from configured providers — approve or revoke for Assistant use

↯

Routing

Priority rules: task, privacy, geography, compliance, cost/latency preference, failover model ids

⊞

Features

Bind platform feature keys to preferred models (seeded defaults on bootstrap)

▤

Prompts

Custom org templates — version, test, approve (DRAFT → APPROVED)

▦

Documents

Knowledge bases, document ingest, embedding model, gateway RAG

♢

Policy & Quotas

Runtime AI policies and org request/token/cost quotas

▥

Usage

Volume by provider/feature, recent requests, feedback, routing playground

Provider catalog (gateway bootstrap)

Bootstrap seeds 21 provider kinds with reference models into the org catalog. Categories: COMMERCIAL SELF_HOSTED ENTERPRISE. Seeded modalities are primarily CHAT and EMBEDDING (the platform enum also supports COMPLETION and MULTIMODAL).

▥

Commercial

OpenAI, Anthropic, Google Gemini, AWS Bedrock, Cohere, Mistral AI, xAI (Grok), Together AI, Fireworks AI

⌂

Self-hosted

Ollama, vLLM, Hugging Face TGI, NVIDIA NIM, LM Studio, Open WebUI, llama.cpp

♢

Enterprise

Azure AI Foundry, Databricks Mosaic AI, Red Hat OpenShift AI, IBM watsonx.ai, Vertex AI

♧ Administration → Integrations → AI Providers exposes configure cards for OpenAI, Google Gemini, AWS Bedrock, Ollama, Vertex AI, and Azure AI Foundry. Other catalog kinds can be enabled via Model Platform providers API or future admin cards once credentials are stored.

Feature bindings (defaults)

Seeded on bootstrap. Used as the preferred model when callers pass a featureKey but no explicit modelId. The AI Assistant uses feature key copilot and typically passes an explicit model from the picker, which wins over routing rules.

Feature key	Label	Preferred model slug
explain-terraform-plans	Explain Terraform Plans	gpt-5
iac-generation	IaC Generation	claude-sonnet
kubernetes-troubleshooting	Kubernetes Troubleshooting	llama-3-3
cost-optimization	Cost Optimization	gemini-2-flash
compliance-analysis	Compliance Analysis	bedrock-claude
air-gapped-deployment	Air-Gapped Deployment	llama-3-3
copilot	AI Assistant	— (explicit model from Assistant picker)
ai_intelligence	AI Intelligence scans	— (routed via rules / defaults)

♧ Routing order: explicit modelId (Assistant picker) → feature-bound model id → enabled route rule (priority) → cost/latency preference → cheapest approved non-embedding model. Failover model ids are tried when the primary provider call fails.

♢

Security scan

Shared gateway security engine applied to every completion request.

• Secrets (AWS keys, OpenAI sk-, GitHub PAT, generic api_key/password/token)
• PII (email, SSN, card, phone) → redacted placeholders
• Prompt injection / jailbreak patterns (user messages only)
• Moderation (violence/malware phrasing)

Injection and moderation are BLOCKED on user turns by default. Secrets and PII are redacted in outbound content. Request log statuses: SUCCESS, FAILED, BLOCKED, RATE_LIMITED, FAILOVER.

♧

Governance

Custom prompt lifecycle: DRAFT → APPROVED (new versions reset to DRAFT until re-approved). Built-in library prompts are read-only — seePrompt Library.

Enum also supports PENDING_APPROVAL, REJECTED, ARCHIVED.

Policy actions: ALLOW, DENY, REQUIRE_APPROVAL, REDACT, REROUTE

Install the AI Governance Pack from Security & Governance → Policy Library for infrastructure-level AI controls. Model Platform policies enforce at LLM request time.

Memory scopes: SESSION, PROJECT, ORGANIZATION, KNOWLEDGE