fix(go): exclude reasoning part when return value from Text() function

[jcooky]

I’ve noticed that the Text() function is meant to operate on TextPart or DataPart only. If a developer wants to retrieve the reasoning part as text, they should use Reasoning() instead.

However, Text() currently returns a merged value that includes the reasoning part, which does not match the intended behavior.

Checklist (if applicable):

• PR title is following https://www.conventionalcommits.org/en/v1.0.0/
• Tested (manually, unit tested, etc.)
• Docs updated (updated docs or a docs bug required)

[huangjeff5]

Good catch! But should we be including Data() parts in Text()?

[jcooky]

@huangjeff5 The Output() function currently calls Text() for JSON parsing, but I believed the intended design is to use the Data part for that.

For backward compatibility, Text() still includes the Data part. In both the OpenAI and Gemini codebases, JSON output is written into the Text part.

You can find the relevant code here:

• genkit/go/plugins/compat_oai/generate.go

  Lines 338 to 385 in 310f458

  	completion, err := g.client.Chat.Completions.New(ctx, *g.request)
  	if err != nil {
  	return nil, fmt.Errorf("failed to create completion: %w", err)
  	}
  	resp := &ai.ModelResponse{
  	Request: &ai.ModelRequest{},
  	Usage: &ai.GenerationUsage{
  	InputTokens: int(completion.Usage.PromptTokens),
  	OutputTokens: int(completion.Usage.CompletionTokens),
  	TotalTokens: int(completion.Usage.TotalTokens),
  	},
  	Message: &ai.Message{
  	Role: ai.RoleModel,
  	},
  	}
  	choice := completion.Choices[0]
  	switch choice.FinishReason {
  	case openai.ChatCompletionChoicesFinishReasonStop, openai.ChatCompletionChoicesFinishReasonToolCalls:
  	resp.FinishReason = ai.FinishReasonStop
  	case openai.ChatCompletionChoicesFinishReasonLength:
  	resp.FinishReason = ai.FinishReasonLength
  	case openai.ChatCompletionChoicesFinishReasonContentFilter:
  	resp.FinishReason = ai.FinishReasonBlocked
  	case openai.ChatCompletionChoicesFinishReasonFunctionCall:
  	resp.FinishReason = ai.FinishReasonOther
  	default:
  	resp.FinishReason = ai.FinishReasonUnknown
  	}
  	// handle tool calls
  	var toolRequestParts []*ai.Part
  	for _, toolCall := range choice.Message.ToolCalls {
  	toolRequestParts = append(toolRequestParts, ai.NewToolRequestPart(&ai.ToolRequest{
  	Name: toolCall.Function.Name,
  	Input: jsonStringToMap(toolCall.Function.Arguments),
  	}))
  	}
  	if len(toolRequestParts) > 0 {
  	resp.Message.Content = toolRequestParts
  	return resp, nil
  	}
  	resp.Message.Content = []*ai.Part{
  	ai.NewTextPart(completion.Choices[0].Message.Content),
  	}

• genkit/go/plugins/googlegenai/gemini.go

  Lines 682 to 702 in 310f458

  	func translateResponse(resp *genai.GenerateContentResponse) *ai.ModelResponse {
  	var r *ai.ModelResponse
  	if len(resp.Candidates) > 0 {
  	r = translateCandidate(resp.Candidates[0])
  	} else {
  	r = &ai.ModelResponse{}
  	}
  	if r.Usage == nil {
  	r.Usage = &ai.GenerationUsage{}
  	}
  	if u := resp.UsageMetadata; u != nil {
  	r.Usage.InputTokens = int(u.PromptTokenCount)
  	r.Usage.OutputTokens = int(u.CandidatesTokenCount)
  	r.Usage.TotalTokens = int(u.TotalTokenCount)
  	r.Usage.CachedContentTokens = int(u.CachedContentTokenCount)
  	r.Usage.ThoughtsTokens = int(u.ThoughtsTokenCount)
  	}
  	return r
  	}

These implementations should use the Data part instead of the Text part when the output mode is JSON. If we remove the Data part from Text(), we need to update them to support the OutputMode and write output to the Data part instead of Text.

[apascal07]

We call Text() for JSON parsing because JSON is usually (always?) returned as markdown which is text not valid JSON data, which we then strip the markdown delimiters to turn it into "data".

[jcooky]

@apascal07 Thank you for your insight. I've added a comment about you mentioned.

Can you review it?