feat(go/plugins): Generic OpenAI plugin #2440
Conversation
feat(openai plugin): Initial openai plugin
Co-authored-by: Yuki Nagae <[email protected]>
feat(go): new plugin creation way + dep issues fixed
feat(go/openai plugin): added function calling
| openaiGo.ChatModelGPT4oMini: { | ||
| Label: "GPT-4o-mini", | ||
| Supports: compat_oai.Multimodal.Supports, | ||
| }, |
There was a problem hiding this comment.
can add more supported models here
|
|
||
| // Ensure response has required fields | ||
| if resp == nil { | ||
| resp = &ai.ModelResponse{} | ||
| } | ||
| if resp.Message == nil { | ||
| resp.Message = &ai.Message{ | ||
| Role: ai.RoleModel, | ||
| } | ||
| } | ||
| if resp.Usage == nil { | ||
| resp.Usage = &ai.GenerationUsage{} | ||
| } | ||
|
|
There was a problem hiding this comment.
could migrate to within builder methods
| var data openaiGo.EmbeddingNewParamsInputArrayOfStrings | ||
| for _, doc := range input.Documents { | ||
| for _, p := range doc.Content { | ||
| data = append(data, p.Text) | ||
| } | ||
| } | ||
|
|
||
| params := openaiGo.EmbeddingNewParams{ | ||
| Input: openaiGo.F[openaiGo.EmbeddingNewParamsInputUnion](data), | ||
| Model: openaiGo.F(name), | ||
| EncodingFormat: openaiGo.F(openaiGo.EmbeddingNewParamsEncodingFormatFloat), | ||
| } | ||
|
|
||
| embeddingResp, err := o.client.Embeddings.New(ctx, params) | ||
| if err != nil { | ||
| return nil, err | ||
| } | ||
|
|
||
| resp := &ai.EmbedResponse{} | ||
| for _, emb := range embeddingResp.Data { | ||
| embedding := make([]float32, len(emb.Embedding)) | ||
| for i, val := range emb.Embedding { | ||
| embedding[i] = float32(val) | ||
| } | ||
| resp.Embeddings = append(resp.Embeddings, &ai.DocumentEmbedding{Embedding: embedding}) | ||
| } |
There was a problem hiding this comment.
could migrate into within generate.go - isolate testing for openai client and compatibility with genkit
| o.mu.Lock() | ||
| defer o.mu.Unlock() | ||
| if o.initted { | ||
| panic("compat_oai.Init already called") |
There was a problem hiding this comment.
Return error here instead of panicking.
| } | ||
|
|
||
| // DefineModel defines a model in the registry | ||
| func (o *OpenAICompatible) DefineModel(g *genkit.Genkit, name string, info ai.ModelInfo, provider string) (ai.Model, error) { |
There was a problem hiding this comment.
| func (o *OpenAICompatible) DefineModel(g *genkit.Genkit, name string, info ai.ModelInfo, provider string) (ai.Model, error) { | |
| func (o *OpenAICompatible) DefineModel(g *genkit.Genkit, provider, name string, info ai.ModelInfo) (ai.Model, error) { |
| o.mu.Lock() | ||
| defer o.mu.Unlock() | ||
| if !o.initted { | ||
| panic("OpenAICompatible.Init not called") |
There was a problem hiding this comment.
Return error here instead of panicking.
| o.mu.Lock() | ||
| defer o.mu.Unlock() | ||
| if !o.initted { | ||
| panic("OpenAICompatible.Init not called") |
There was a problem hiding this comment.
Return error here instead of panicking.
| ctx := context.Background() | ||
| g, err := genkit.Init(ctx) | ||
| if err != nil { | ||
| log.Fatalf("failed to create Genkit: %v", err) | ||
| } | ||
|
|
||
| apiKey := os.Getenv("OPENAI_API_KEY") | ||
| apiKeyOption := option.WithAPIKey(apiKey) | ||
| oai := oai.OpenAI{ | ||
| Opts: []option.RequestOption{apiKeyOption}, | ||
| } | ||
|
|
||
| oai.Init(ctx, g) | ||
| genkit.WithPlugins(&oai) |
There was a problem hiding this comment.
| ctx := context.Background() | |
| g, err := genkit.Init(ctx) | |
| if err != nil { | |
| log.Fatalf("failed to create Genkit: %v", err) | |
| } | |
| apiKey := os.Getenv("OPENAI_API_KEY") | |
| apiKeyOption := option.WithAPIKey(apiKey) | |
| oai := oai.OpenAI{ | |
| Opts: []option.RequestOption{apiKeyOption}, | |
| } | |
| oai.Init(ctx, g) | |
| genkit.WithPlugins(&oai) | |
| ctx := context.Background() | |
| g, err := genkit.Init(ctx, genkit.WithPlugins(oai.OpenAI{ | |
| Opts: []option.RequestOption{option.WithAPIKey(os.Getenv("OPENAI_API_KEY"))}, | |
| }) | |
| if err != nil { | |
| log.Fatalf("failed to create Genkit: %v", err) | |
| } |
Curious, is there any reason why you didn't do it like this? I wonder if it's not obvious that it should be called this way. The way you had genkit.WithPlugins(&oai) did nothing -- you had already initialized it yourself and the option isn't passed anywhere.
| openaiGo.ChatModelGPT4oMini: { | ||
| Label: "GPT-4o-mini", | ||
| Supports: compat_oai.Multimodal.Supports, | ||
| }, |
| } | ||
|
|
||
| func (o *OpenAI) Init(ctx context.Context, g *genkit.Genkit) error { | ||
| err := o.openAICompatible.Init(ctx, g) |
There was a problem hiding this comment.
Nit: here and elsewhere, can do this which is more succinct and preferred:
if err := o.openAICompatible.Init(ctx, g); err != nil {
return err
}
| Multiturn: true, | ||
| Tools: true, | ||
| SystemRole: true, | ||
| Media: true, |
| initted bool | ||
| client *openaiGo.Client | ||
| Opts []option.RequestOption | ||
| Provider string |
There was a problem hiding this comment.
Some comments here would be good. What is provider?
| }, | ||
| { | ||
| name: "float and int fields", | ||
| config: &openai.ChatCompletionNewParams{ |
There was a problem hiding this comment.
We should try not to depend directly on the config type defined elsewhere (like in 3P SDKs) so that the SDKs are implementation details and we could in theory swap out to another or even call the REST API directly. We sometimes implement configuration via native Genkit features so if we have it both natively and in the config, it will be confusing as to what happens when you combine them.
| t.Log("genkit initialized") | ||
|
|
||
| // Initialize the OpenAI plugin | ||
| apiKeyOption := option.WithAPIKey(apiKey) |
There was a problem hiding this comment.
Which other options do we (want to) support? Maybe APIKey should be a top level field in OpenAI like it is in GoogleAI? Better to err on the side of being simple and complete rather than supporting everything that the SDK supports.
Adding support for an openai go plugin, along with additional generic support for any providers compatible with the openai (from feature request #2204). Collaboration between @xavidop @yukinagae @kekoawong.
Checklist: