Workaround openai issue with temperature: 0 being omitted from request

This PR adds a workaround for [this GitHub discussion][gh], describing
how chat completion requests which explicitly set temperature = 0 are
marshalled to JSON incorrectly (due to the 0 being indistinguishable
from the zero value of the field).

To do so we add a custom unmarshal method which checks whether the field
was defined and set to zero, and explicitly set the temperature to
`math.SmallestNonzeroFloat32` in such cases. This isn't perfect but is
likely to get very similar results in practice.

[gh]: sashabaranov/go-openai#9 (comment)

Author: sd2k

packages/grafana-llm-app/pkg/plugin/llm_provider.go

@@ -5,6 +5,7 @@ import (
 	"encoding/json"
 	"errors"
 	"fmt"
+	"math"
 	"math/rand"
 	"strings"
 
@@ -67,6 +68,32 @@ type ChatCompletionRequest struct {
 	Model Model `json:"model"`
 }
 
+// UnmarshalJSON implements json.Unmarshaler.
+// We have a custom implementation here to check whether temperature is being
+// explicitly set to `0` in the incoming request, because the `openai.ChatCompletionRequest`
+// struct has `omitempty` on the Temperature field and would omit it when marshaling.
+// If there is an explicit 0 value in the request, we set it to `math.SmallestNonzeroFloat32`,
+// a workaround mentioned in https://github.com/sashabaranov/go-openai/issues/9#issuecomment-894845206.
+func (c *ChatCompletionRequest) UnmarshalJSON(data []byte) error {
+	// Create a wrapper type alias to avoid recursion, otherwise the
+	// subsequent call to UnmarshalJSON would call this method forever.
+	type Alias ChatCompletionRequest
+	var a Alias
+	if err := json.Unmarshal(data, &a); err != nil {
+		return err
+	}
+	// Also unmarshal to a map to check if temperature is being set explicitly in the request.
+	r := map[string]any{}
+	if err := json.Unmarshal(data, &r); err != nil {
+		return err
+	}
+	if t, ok := r["temperature"].(float64); ok && t == 0 {
+		a.ChatCompletionRequest.Temperature = math.SmallestNonzeroFloat32
+	}
+	*c = ChatCompletionRequest(a)
+	return nil
+}
+
 type ChatCompletionStreamResponse struct {
 	openai.ChatCompletionStreamResponse
 	// Random padding used to mitigate side channel attacks.

packages/grafana-llm-app/pkg/plugin/llm_provider_test.go

@@ -2,9 +2,11 @@ package plugin
 
 import (
 	"encoding/json"
+	"math"
 	"testing"
 
 	"github.com/sashabaranov/go-openai"
+	"github.com/stretchr/testify/assert"
 )
 
 func TestModelFromString(t *testing.T) {
@@ -83,7 +85,7 @@ func TestModelFromString(t *testing.T) {
 	}
 }
 
-func TestUnmarshalJSON(t *testing.T) {
+func TestModelUnmarshalJSON(t *testing.T) {
 	tests := []struct {
 		input    []byte
 		expected Model
@@ -164,6 +166,48 @@ func TestUnmarshalJSON(t *testing.T) {
 	}
 }
 
+func TestChatCompletionRequestUnmarshalJSON(t *testing.T) {
+	for _, tt := range []struct {
+		input    []byte
+		expected ChatCompletionRequest
+	}{
+		{
+			input: []byte(`{"model":"base"}`),
+			expected: ChatCompletionRequest{
+				Model: ModelBase,
+				ChatCompletionRequest: openai.ChatCompletionRequest{
+					Temperature: 0,
+				},
+			},
+		},
+		{
+			input: []byte(`{"model":"base", "temperature":0.5}`),
+			expected: ChatCompletionRequest{
+				Model: ModelBase,
+				ChatCompletionRequest: openai.ChatCompletionRequest{
+					Temperature: 0.5,
+				},
+			},
+		},
+		{
+			input: []byte(`{"model":"base", "temperature":0}`),
+			expected: ChatCompletionRequest{
+				Model: ModelBase,
+				ChatCompletionRequest: openai.ChatCompletionRequest{
+					Temperature: math.SmallestNonzeroFloat32,
+				},
+			},
+		},
+	} {
+		t.Run(string(tt.input), func(t *testing.T) {
+			var req ChatCompletionRequest
+			err := json.Unmarshal(tt.input, &req)
+			assert.NoError(t, err)
+			assert.Equal(t, tt.expected, req)
+		})
+	}
+}
+
 func TestChatCompletionStreamResponseMarshalJSON(t *testing.T) {
 	resp := ChatCompletionStreamResponse{
 		ChatCompletionStreamResponse: openai.ChatCompletionStreamResponse{