Skip to content

Add support for reasoning_content field in chat completion messages for DeepSeek R1 #925

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 14 commits into from
Apr 29, 2025
Merged

Add support for reasoning_content field in chat completion messages for DeepSeek R1 #925

Show file tree
Hide file tree
Changes from 9 commits
Commits
Show all changes
14 commits
Select commit Hold shift + click to select a range
5297ba3
support deepseek field "reasoning_content"
goodenough227 Jan 21, 2025
1d922c6
support deepseek field "reasoning_content"
goodenough227 Jan 21, 2025
5fef03f
Comment ends in a period (godot)
goodenough227 Jan 23, 2025
e025915
add comment on field reasoning_content
goodenough227 Feb 5, 2025
1e313ed
Merge branch 'feat/reasoning-content'
goodenough227 Feb 5, 2025
1105cfc
fix go lint error
goodenough227 Feb 5, 2025
09ab25f
chore: trigger CI
goodenough227 Feb 5, 2025
c31b505
Merge branch 'sashabaranov:master' into master
goodenough227 Feb 7, 2025
5273861
make field "content" in MarshalJSON function omitempty
goodenough227 Feb 7, 2025
e222662
Merge branch 'sashabaranov:master' into master
goodenough227 Mar 12, 2025
fd8334a
remove reasoning_content in TestO1ModelChatCompletions func
goodenough227 Mar 12, 2025
a450fb8
feat: Add test and handler for deepseek-reasoner chat model completio…
goodenough227 Mar 13, 2025
6ef63e2
feat: Add test and handler for deepseek-reasoner chat model completio…
goodenough227 Mar 13, 2025
319a8ea
feat: Add test and handler for deepseek-reasoner chat model completio…
goodenough227 Mar 13, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
74 changes: 42 additions & 32 deletions chat.go
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -103,6 +103,12 @@ type ChatCompletionMessage struct {
// - https://github.com/openai/openai-cookbook/blob/main/examples/How_to_count_tokens_with_tiktoken.ipynb
Name string `json:"name,omitempty"`

// This property is used for the "reasoning" feature supported by deepseek-reasoner
// which is not in the official documentation.
// the doc from deepseek:
// - https://api-docs.deepseek.com/api/create-chat-completion#responses
ReasoningContent string `json:"reasoning_content,omitempty"`

FunctionCall *FunctionCall `json:"function_call,omitempty"`

// For Role=assistant prompts this may be set to the tool calls generated by the model, such as function calls.
Expand All @@ -118,56 +124,60 @@ func (m ChatCompletionMessage) MarshalJSON() ([]byte, error) {
}
if len(m.MultiContent) > 0 {
msg := struct {
Role string `json:"role"`
Content string `json:"-"`
Refusal string `json:"refusal,omitempty"`
MultiContent []ChatMessagePart `json:"content,omitempty"`
Name string `json:"name,omitempty"`
FunctionCall *FunctionCall `json:"function_call,omitempty"`
ToolCalls []ToolCall `json:"tool_calls,omitempty"`
ToolCallID string `json:"tool_call_id,omitempty"`
Role string `json:"role"`
Content string `json:"-"`
Refusal string `json:"refusal,omitempty"`
MultiContent []ChatMessagePart `json:"content,omitempty"`
Name string `json:"name,omitempty"`
ReasoningContent string `json:"reasoning_content,omitempty"`
FunctionCall *FunctionCall `json:"function_call,omitempty"`
ToolCalls []ToolCall `json:"tool_calls,omitempty"`
ToolCallID string `json:"tool_call_id,omitempty"`
}(m)
return json.Marshal(msg)
}

msg := struct {
Role string `json:"role"`
Content string `json:"content,omitempty"`
Refusal string `json:"refusal,omitempty"`
MultiContent []ChatMessagePart `json:"-"`
Name string `json:"name,omitempty"`
FunctionCall *FunctionCall `json:"function_call,omitempty"`
ToolCalls []ToolCall `json:"tool_calls,omitempty"`
ToolCallID string `json:"tool_call_id,omitempty"`
Role string `json:"role"`
Content string `json:"content,omitempty"`
Refusal string `json:"refusal,omitempty"`
MultiContent []ChatMessagePart `json:"-"`
Name string `json:"name,omitempty"`
ReasoningContent string `json:"reasoning_content,omitempty"`
FunctionCall *FunctionCall `json:"function_call,omitempty"`
ToolCalls []ToolCall `json:"tool_calls,omitempty"`
ToolCallID string `json:"tool_call_id,omitempty"`
}(m)
return json.Marshal(msg)
}

func (m *ChatCompletionMessage) UnmarshalJSON(bs []byte) error {
msg := struct {
Role string `json:"role"`
Content string `json:"content,omitempty"`
Refusal string `json:"refusal,omitempty"`
MultiContent []ChatMessagePart
Name string `json:"name,omitempty"`
FunctionCall *FunctionCall `json:"function_call,omitempty"`
ToolCalls []ToolCall `json:"tool_calls,omitempty"`
ToolCallID string `json:"tool_call_id,omitempty"`
Role string `json:"role"`
Content string `json:"content"`
Refusal string `json:"refusal,omitempty"`
MultiContent []ChatMessagePart
Name string `json:"name,omitempty"`
ReasoningContent string `json:"reasoning_content,omitempty"`
FunctionCall *FunctionCall `json:"function_call,omitempty"`
ToolCalls []ToolCall `json:"tool_calls,omitempty"`
ToolCallID string `json:"tool_call_id,omitempty"`
}{}

if err := json.Unmarshal(bs, &msg); err == nil {
*m = ChatCompletionMessage(msg)
return nil
}
multiMsg := struct {
Role string `json:"role"`
Content string
Refusal string `json:"refusal,omitempty"`
MultiContent []ChatMessagePart `json:"content"`
Name string `json:"name,omitempty"`
FunctionCall *FunctionCall `json:"function_call,omitempty"`
ToolCalls []ToolCall `json:"tool_calls,omitempty"`
ToolCallID string `json:"tool_call_id,omitempty"`
Role string `json:"role"`
Content string
Refusal string `json:"refusal,omitempty"`
MultiContent []ChatMessagePart `json:"content"`
Name string `json:"name,omitempty"`
ReasoningContent string `json:"reasoning_content,omitempty"`
FunctionCall *FunctionCall `json:"function_call,omitempty"`
ToolCalls []ToolCall `json:"tool_calls,omitempty"`
ToolCallID string `json:"tool_call_id,omitempty"`
}{}
if err := json.Unmarshal(bs, &multiMsg); err != nil {
return err
Expand Down
6 changes: 6 additions & 0 deletions chat_stream.go
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,6 +11,12 @@ type ChatCompletionStreamChoiceDelta struct {
FunctionCall *FunctionCall `json:"function_call,omitempty"`
ToolCalls []ToolCall `json:"tool_calls,omitempty"`
Refusal string `json:"refusal,omitempty"`

// This property is used for the "reasoning" feature supported by deepseek-reasoner
// which is not in the official documentation.
// the doc from deepseek:
// - https://api-docs.deepseek.com/api/create-chat-completion#responses
ReasoningContent string `json:"reasoning_content,omitempty"`
}

type ChatCompletionStreamChoiceLogprobs struct {
Expand Down
5 changes: 3 additions & 2 deletions chat_test.go
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -420,8 +420,9 @@ func TestO1ModelChatCompletions(t *testing.T) {
MaxCompletionTokens: 1000,
Messages: []openai.ChatCompletionMessage{
{
Role: openai.ChatMessageRoleUser,
Content: "Hello!",
Role: openai.ChatMessageRoleUser,
Content: "Hello!",
ReasoningContent: "hi!",
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Does this deserve its own test in isolation?

This is currently targeting a specific niche model.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I have removed the ReasoningContent section here but did not add the constants and other test functions for the deepseek model, as it seems that, from your perspective, go-openai is primarily intended for OpenAI models. Please share your thoughts on this so I can align it with my subsequent changes.

Copy link
Contributor

@kmesiab kmesiab Mar 12, 2025 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

go-openai is primarily intended for OpenAI models.

This is my personal stance, yes. The addition of support for DeepSeek feels like a knee-jerk reaction. While I appreciate the flexibility your addition provides, it is still a divergence from the spirit of the repo. The o1 models, and the future path of OpenAI's API may very well introduce a competing field.

The compromise, if you will, is to limit the coupling, like here in the tests.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I understand. I have written test functions specifically for deepseek r1 using literal model names to minimize the invasiveness to the main codebase.

Additionally, I'd like to know if there are plans to support non-OpenAI official models in the future, as many new models (like deepseek open-source models) are using OpenAI's data structures. We are currently using our forked branch of go-openai in this way.

},
},
})
Expand Down
2 changes: 1 addition & 1 deletion openai_test.go
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,7 +29,7 @@ func setupAzureTestServer() (client *openai.Client, server *test.ServerTest, tea

// numTokens Returns the number of GPT-3 encoded tokens in the given text.
// This function approximates based on the rule of thumb stated by OpenAI:
// https://beta.openai.com/tokenizer/
// https://beta.openai.com/tokenizer.
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Why is this changing?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is to pass the golint check. When using golint for inspection, golint will require this to be changed to "."

//
// TODO: implement an actual tokenizer for GPT-3 and Codex (once available).
func numTokens(s string) int {
Expand Down
Loading