#67: Experimenting with dynamic lang provider config loading & valida…

…tion

pkg/models/config.go

@@ -2,17 +2,15 @@ package models
 
 import (
 	"fmt"
-
-	"github.com/EinStack/glide/pkg/providers"
-
 	"github.com/EinStack/glide/pkg/clients"
+	"github.com/EinStack/glide/pkg/provider"
 	"github.com/EinStack/glide/pkg/resiliency/health"
 	"github.com/EinStack/glide/pkg/routers/latency"
 	"github.com/EinStack/glide/pkg/telemetry"
 )
 
 // Config defines an extra configuration for a model wrapper around a provider
-type Config[P providers.ProviderFactory] struct {
+type Config[P provider.ProviderConfig] struct {
 	ID          string                `yaml:"id" json:"id" validate:"required"`           // Model instance ID (unique in scope of the router)
 	Enabled     bool                  `yaml:"enabled" json:"enabled" validate:"required"` // Is the model enabled?
 	ErrorBudget *health.ErrorBudget   `yaml:"error_budget" json:"error_budget" swaggertype:"primitive,string"`
@@ -23,15 +21,15 @@ type Config[P providers.ProviderFactory] struct {
 	Provider P `yaml:"provider" json:"provider"`
 }
 
-func NewConfig[P providers.ProviderFactory](ID string) *Config[P] {
+func NewConfig[P provider.ProviderConfig](ID string) *Config[P] {
 	config := DefaultConfig[P]()
 
 	config.ID = ID
 
 	return &config
 }
 
-func DefaultConfig[P providers.ProviderFactory]() Config[P] {
+func DefaultConfig[P provider.ProviderConfig]() Config[P] {
 	return Config[P]{
 		Enabled:     true,
 		Client:      clients.DefaultClientConfig(),

pkg/models/lang.go

@@ -2,11 +2,10 @@ package models
 
 import (
 	"context"
+	"github.com/EinStack/glide/pkg/provider"
 	"io"
 	"time"
 
-	"github.com/EinStack/glide/pkg/providers"
-
 	"github.com/EinStack/glide/pkg/clients"
 	health2 "github.com/EinStack/glide/pkg/resiliency/health"
 
@@ -32,14 +31,14 @@ type LangModel interface {
 type LanguageModel struct {
 	modelID               string
 	weight                int
-	client                providers.LangProvider
+	client                provider.LangProvider
 	healthTracker         *health2.Tracker
 	chatLatency           *latency.MovingAverage
 	chatStreamLatency     *latency.MovingAverage
 	latencyUpdateInterval *fields.Duration
 }
 
-func NewLangModel(modelID string, client providers.LangProvider, budget *health2.ErrorBudget, latencyConfig latency.Config, weight int) *LanguageModel {
+func NewLangModel(modelID string, client provider.LangProvider, budget *health2.ErrorBudget, latencyConfig latency.Config, weight int) *LanguageModel {
 	return &LanguageModel{
 		modelID:               modelID,
 		client:                client,

pkg/provider/config.go

@@ -0,0 +1,12 @@
+package provider
+
+import (
+	"github.com/EinStack/glide/pkg/clients"
+	"github.com/EinStack/glide/pkg/telemetry"
+)
+
+// TODO: ProviderConfig should be more generic, not tied to LangProviders
+type ProviderConfig interface {
+	UnmarshalYAML(unmarshal func(interface{}) error) error
+	ToClient(tel *telemetry.Telemetry, clientConfig *clients.ClientConfig) (LangProvider, error)
+}

pkg/providers/provider.go → pkg/provider/provider.go

@@ -1,4 +1,4 @@
-package providers
+package provider
 
 import (
 	"context"
@@ -10,9 +10,11 @@ import (
 
 var ErrProviderNotFound = errors.New("provider not found")
 
+type ProviderID = string
+
 // ModelProvider exposes provider context
 type ModelProvider interface {
-	Provider() string
+	Provider() ProviderID
 	ModelName() string
 }
 
@@ -25,3 +27,12 @@ type LangProvider interface {
 	Chat(ctx context.Context, params *schemas.ChatParams) (*schemas.ChatResponse, error)
 	ChatStream(ctx context.Context, params *schemas.ChatParams) (clients.ChatStream, error)
 }
+
+// EmbeddingProvider defines an interface a provider should fulfill to be able to generate embeddings
+type EmbeddingProvider interface {
+	ModelProvider
+
+	SupportEmbedding() bool
+
+	Embed(ctx context.Context, params *schemas.ChatParams) (*schemas.ChatResponse, error)
+}