Create an AI API¶

What happens

An API creator exposes an AI service, such as OpenAI, Azure OpenAI or Mistral, through APIM. They pick an LLM provider, which an admin has registered with its API definition and models. APIM creates a normal API (AM_API plus resources) and links it to the provider through AM_API_AI_CONFIGURATION. Subscription tiers can then limit tokens, not just requests.

Who: Admin (registers providers), API creator (Publisher) · Tables written: AM_LLM_PROVIDER + AM_LLM_PROVIDER_MODEL (at startup or by an admin), AM_API, AM_API_LC_EVENT, AM_API_URL_MAPPING, AM_API_AI_CONFIGURATION, AM_API_PRIMARY_EP_MAPPING, GOV_*, REG_* · Tables read: AM_POLICY_SUBSCRIPTION

How the tables connect¶

An AI API is a normal API with one extra link to the provider it fronts.

erDiagram
    AM_LLM_PROVIDER ||--o{ AM_LLM_PROVIDER_MODEL : "offers"
    AM_LLM_PROVIDER ||--o{ AM_API_AI_CONFIGURATION : "used by"
    AM_API ||--o{ AM_API_AI_CONFIGURATION : "configured by"
    AM_API ||--o{ AM_API_PRIMARY_EP_MAPPING : "primary endpoint"

There's one AI configuration row per API per revision, so revisions snapshot the provider choice too.

The flow at a glance¶

sequenceDiagram
    actor Admin
    actor Creator as API creator
    participant Pub as Publisher / Admin Portal
    participant DB as APIM DB
    Admin->>Pub: Register provider (built-ins are seeded at startup)
    Pub->>DB: insert AM_LLM_PROVIDER + AM_LLM_PROVIDER_MODEL
    Creator->>Pub: Create AI API from OpenAI
    Pub->>DB: insert AM_API (API_SUBTYPE = AIAPI) + URL mappings
    Pub->>DB: insert AM_API_AI_CONFIGURATION
    Pub->>DB: insert AM_API_PRIMARY_EP_MAPPING

After this, the API is revisioned, deployed, published and subscribed to exactly like any other API (flows 4–10).

Step by step¶

Provider registration (admin, or built in).
- AM_LLM_PROVIDER holds UUID, NAME, API_VERSION and ORGANIZATION.
- BUILT_IN_SUPPORT is true for providers that ship with APIM.
- CONFIGURATIONS holds how to read token counts, the model name and so on from responses. API_DEFINITION holds the provider's OpenAPI.
- (NAME, API_VERSION, ORGANIZATION) is unique.
UUID NAME API_VERSION BUILT_IN_SUPPORT ORGANIZATION

4d78… OpenAI 2.0.0 true carbon.super
Models. AM_LLM_PROVIDER_MODEL lists the models the provider offers, with MODEL_NAME, MODEL_FAMILY_NAME and LLM_PROVIDER_UUID (a real FK).

MODEL_ID MODEL_NAME MODEL_FAMILY_NAME LLM_PROVIDER_UUID

18 gpt-4o OpenAI 4d78…

19 gpt-4o-mini OpenAI 4d78…

20 o3-mini OpenAI 4d78…
The API itself. An AM_API row is created with API_TYPE = 'HTTP' and API_SUBTYPE = 'AIAPI', which is what marks it as an AI API. Its resources, e.g. POST /chat/completions, are stored in AM_API_URL_MAPPING like any other API's.

API_ID API_UUID API_NAME CONTEXT API_TYPE API_SUBTYPE STATUS

4 e6c6… ChatAPI /chat/1.0.0 HTTP AIAPI CREATED
The link. AM_API_AI_CONFIGURATION holds AI_CONFIGURATION_UUID, API_UUID (FK), API_REVISION_UUID and LLM_PROVIDER_UUID (FK). For the working copy API_REVISION_UUID is NULL. When a revision is created, a copy with the revision's UUID is expected, though revisions of the AI API weren't tested.

AI_CONFIGURATION_UUID API_UUID API_REVISION_UUID LLM_PROVIDER_UUID

796b… e6c6… (null) 4d78… (OpenAI 2.0.0)
Endpoints and secrets. AM_API_PRIMARY_EP_MAPPING records which endpoint is primary: (API_UUID, ENDPOINT_UUID = 'default_production_endpoint', REVISION_UUID = 'Current API'). The default endpoint defined at creation time stayed in the registry artifact; no AM_API_ENDPOINTS row was written. Extra named endpoints, used for model routing and failover, go into AM_API_ENDPOINTS with their URL and API key in the ENDPOINT_CONFIG blob. Per-environment keys can go into AM_API_ENVIRONMENT_KEYS.
Token-based limits. A subscription tier can cap AI usage through the AM_POLICY_SUBSCRIPTION columns TOTAL_TOKEN_COUNT, PROMPT_TOKEN_COUNT and COMPLETION_TOKEN_COUNT, with QUOTA_TYPE = 'aiApiQuota'. The seeded tiers AIGold, AISilver and AIBronze are of this type (e.g. AIGold: 500 requests and 50000 total tokens per minute).

What gets cleaned up¶

The FKs from AM_API_AI_CONFIGURATION and AM_LLM_PROVIDER_MODEL have no ON DELETE rule. You can't delete a provider while APIs use it, or while it still has model rows.
APIM's code removes the AI configuration when the API is deleted.

Try it¶

This query lists the AI APIs, the provider behind each one, and the provider's models.

SELECT a.API_NAME, a.API_VERSION, p.NAME AS PROVIDER, p.API_VERSION AS PROVIDER_VERSION,
       m.MODEL_NAME
FROM AM_API_AI_CONFIGURATION c
JOIN AM_API a ON a.API_UUID = c.API_UUID
JOIN AM_LLM_PROVIDER p ON p.UUID = c.LLM_PROVIDER_UUID
LEFT JOIN AM_LLM_PROVIDER_MODEL m ON m.LLM_PROVIDER_UUID = p.UUID
WHERE c.API_REVISION_UUID IS NULL OR c.API_REVISION_UUID = 'Current API';

Different in 3.x

AI APIs don't exist in 3.x. There are no LLM provider or AI configuration tables. The 3.x flows have no equivalent page.

Related domains: AI APIs · API definition · Throttling policies

MODEL_ID	MODEL_NAME	MODEL_FAMILY_NAME	LLM_PROVIDER_UUID
18	gpt-4o	OpenAI	`4d78…`
19	gpt-4o-mini	OpenAI	`4d78…`
20	o3-mini	OpenAI	`4d78…`