Documentation
¶
Overview ¶
Package modelsgateway discovers the models served by an OpenAI-compatible models gateway by querying its /v1/models endpoint.
Index ¶
Constants ¶
This section is empty.
Variables ¶
This section is empty.
Functions ¶
func ListModels ¶
ListModels queries <gatewayURL>/v1/models and returns the IDs of the models the gateway serves. Query parameters present in gatewayURL are forwarded. When the gateway targets a trusted Docker domain, the Docker Desktop token is required and sent as a Bearer token, mirroring how provider clients authenticate gateway traffic.
An error is returned when the gateway is unreachable, responds with a non-200 status (e.g. it doesn't implement /v1/models), or returns a body that isn't a valid OpenAI-style model list. Callers are expected to fall back to their non-discovery behavior in that case.
Types ¶
type Metadata ¶ added in v1.118.0
type Metadata struct {
// Family is the model family from the catalog (e.g. "text-embedding").
Family string
// OutputModalities lists the output modalities the catalog declares.
OutputModalities []string
}
Metadata is the subset of catalog information NormalizeIDs consults to filter out models unsuitable for chat.
type MetadataFunc ¶ added in v1.118.0
MetadataFunc looks up catalog metadata for a provider/model pair. It returns ok=false when the model is unknown to the catalog, in which case NormalizeIDs keeps the model (no signal to exclude it). Callers typically back it with a models.dev store; a nil MetadataFunc skips metadata-based filtering entirely.
type Ref ¶ added in v1.118.0
Ref is a normalized provider/model reference served by a models gateway. It is deliberately a small neutral type so both the runtime (ModelChoice) and the CLI (modelRow) can adapt it without depending on each other.
func NormalizeIDs ¶ added in v1.118.0
func NormalizeIDs(ctx context.Context, ids []string, lookup MetadataFunc) []Ref
NormalizeIDs transforms the raw model IDs returned by a gateway's /v1/models endpoint into provider/model references:
- IDs already prefixed with a provider ("anthropic/claude-x") keep it; bare IDs are routed through the openai provider, matching how the gateway serves them over its OpenAI-compatible endpoint.
- Providers are lowercased ("OpenAI/gpt-4o" → "openai/gpt-4o") so a gateway's casing matches the canonical provider IDs used for catalog lookups and deduplication; the model part keeps its case.
- Malformed references (empty provider or model) are dropped.
- Embedding models are excluded, identified by the model ID or by the catalog family when metadata is available.
- Models whose catalog metadata declares output modalities without "text" are excluded (not suitable for chat); models without metadata or without declared modalities are kept.
- Duplicates are removed; the first occurrence wins, so the result order deterministically follows the gateway's order.
Source Files