How to Consistently Retrieve Valid JSON from Claude 3.5 in Go
Extract structured data from any text with a single function
When working with LLMs as a developer, you often want to receive data in a structured format. While this didn't always work well in the early GPT-3 era, the current models that support function calling are much better at it.
In this post, I want to share a simple function called CallClaudeForceTool
, which you can use to call Claude to receive structured data/JSON output. In my case, it is written in Golang and returns any type of struct I need.
First, you need to define the structure and provide it to the LLM in the format of a JSON schema. In my example, I’m defining a struct type that serves as the input to a function that queries for some data.
Here is the struct that is the input to the query function:
type QueryArgs struct {
SearchQuery string `json:"searchQuery,omitempty"`
TypeFilter string `json:"typeFilter,omitempty"`
}
The following is the JSON Schema definition. Adding good names & descriptions is actually important here, as the LLM will use this information when generating the JSON data:
var QueryFnSchema = JSONSchema{
Type: ai.DataTypeObject,
Properties: map[string]ai.JSONSchema{
"searchQuery": {
Type: ai.DataTypeString,
Description: "The search text query to use. Any data that contains this text will be searched. If not provided, then no text comparison will be performed.",
},
"typeFilter": {
Type: ai.DataTypeString,
Description: `The type to filter on. If not provided, data of any type will be searched. Can only be one of: "location", "event", "person", "organization".`,
Enum: []string{"location", "event", "person", "organization"},
},
},
}
type JSONSchema struct {
Type DataType `json:"type,omitempty"`
Description string `json:"description,omitempty"`
Enum []string `json:"enum,omitempty"`
Properties map[string]JSONSchema `json:"properties,omitempty"`
Required []string `json:"required,omitempty"`
Items *JSONSchema `json:"items,omitempty"`
}
func (d JSONSchema) MarshalJSON() ([]byte, error) {
if d.Properties == nil {
d.Properties = make(map[string]JSONSchema)
}
type Alias JSONSchema
return json.Marshal(struct {
Alias
}{
Alias: (Alias)(d),
})
}
type DataType string
const (
DataTypeObject DataType = "object"
DataTypeNumber DataType = "number"
DataTypeInteger DataType = "integer"
DataTypeBoolean DataType = "boolean"
DataTypeString DataType = "string"
DataTypeArray DataType = "array"
DataTypeNull DataType = "null"
)
I’m using a simple JSON Schema struct. You could use a dependency for that, but this struct is usually sufficient for LLM function calling. It works great with Anthropic’s Claude & OpenAI’s ChatGPT.
You would define your own schema instead of QueryFnSchema
, based on your equivalent to QueryArgs
or whatever data you need and want the LLM to respond with.
Now it’s time to use the CallClaudeForceTool
function (which will be explained at the end of the post):
func main() {
// ...
queryArgs, claudeResponse := CallClaudeForceTool[QueryArgs]([]ClaudeLLMMessage{
{Role: "user", Content: getQueryPrompt(inputMessage)},
}, ToolDefinition{
Name: "search",
Description: "Search for data in the database",
InputSchema: queryFnSchema,
}, ClaudeConfig{
MaxTokens: 4000,
Temperature: 0.8,
System: SYSTEM_PROMPT,
})
result := db.Query(queryArgs)
// ...
}
You’d replace the getQueryPrompt(inputMessage)
, SYSTEM_PROMPT
and any config with your own needs.
The queryArgs
variable now contains the LLM-generated data in the format you defined, which is the QueryArgs
struct in my case. At this point, I use it as a parameter for the actual function call to query my database. However, you can also generate any structured data from some input or prompt and then use it in your code.
Finally, here is the CallClaudeForceTool
function and any auxiliary types. Its goal is to call Claude, force it to use the single tool provided, and end without responding to Claude with any subsequent tool/function output or receiving any final response text message. We just want Claude to respond once with the structured data.
func CallClaudeForceTool[T any](messages []ClaudeLLMMessage, tool ToolDefinition, config ClaudeConfig) (T, *ClaudeCreateMessageResponse, error) {
payload := ClaudeCreateMessagePayload{
Model: "claude-3-5-sonnet-20240620",
Messages: messages,
Tools: []ToolDefinition{tool},
ToolChoice: ClaudeToolChoice{
Type: ToolChoiceTypeTool,
Name: tool.Name,
},
MaxTokens: config.MaxTokens,
Temperature: config.Temperature,
System: config.System,
}
payloadBytes, err := json.Marshal(payload)
if err != nil {
return *new(T), nil, fmt.Errorf("impossible to marshal payload: %w", err)
}
req, err := http.NewRequest(
http.MethodPost,
"https://api.anthropic.com/v1/messages",
bytes.NewReader(payloadBytes),
)
if err != nil {
return *new(T), nil, fmt.Errorf("impossible to create request: %w", err)
}
req.Header.Set("x-api-key", os.Getenv("ANTHROPIC_API_KEY"))
req.Header.Set("content-type", "application/json")
req.Header.Set("anthropic-version", "2023-06-01")
// send the request
httpClient := http.Client{Timeout: 60 * time.Second}
res, err := httpClient.Do(req)
if err != nil {
return *new(T), nil, fmt.Errorf("impossible to send request: %w", err)
}
defer res.Body.Close()
resBody, err := io.ReadAll(res.Body)
if err != nil {
return *new(T), nil, fmt.Errorf("impossible to read all body of response: %w", err)
}
// unmarshal the response
var response ClaudeCreateMessageResponse
err = json.Unmarshal(resBody, &response)
if err != nil {
return *new(T), nil, fmt.Errorf("impossible to unmarshal response: %w", err)
}
// Get the tool response
var toolInput T
for _, content := range response.Content {
if content.Type != ClaudeResponseContentTypeToolUse || content.Name != tool.Name {
continue
}
inputBytes, err := json.Marshal(content.Input)
if err != nil {
return *new(T), &response, fmt.Errorf("impossible to marshal tool response: %w", err)
}
err = json.Unmarshal(inputBytes, &toolInput)
if err != nil {
return *new(T), &response, fmt.Errorf("impossible to unmarshal tool response: %w", err)
}
return toolInput, &response, nil
}
return toolInput, &response, fmt.Errorf("impossible to find tool response")
}
// Auxiliary types
type ClaudeLLMMessage struct {
Role string `json:"role"`
Content string `json:"content"`
}
type ClaudeConfig struct {
MaxTokens int
Temperature float64
System string
}
type ClaudeCreateMessagePayload struct {
Model string `json:"model"`
MaxTokens int `json:"max_tokens"`
Messages []ClaudeLLMMessage `json:"messages"`
Stream bool `json:"stream,omitempty"`
System string `json:"system,omitempty"`
Temperature float64 `json:"temperature,omitempty"`
ToolChoice ClaudeToolChoice `json:"tool_choice,omitempty"`
Tools []ToolDefinition `json:"tools,omitempty"`
}
type ClaudeCreateMessageResponse struct {
Content []struct {
Type ClaudeResponseContentType `json:"type"`
Text string `json:"text,omitempty"`
ID string `json:"id,omitempty"`
Name string `json:"name,omitempty"`
Input map[string]interface{} `json:"input,omitempty"`
} `json:"content"`
Id string `json:"id"`
Model string `json:"model"`
Role string `json:"role"`
StopReason string `json:"stop_reason"`
StopSequence string `json:"stop_sequence"`
Type string `json:"type"`
Usage struct {
InputTokens int `json:"input_tokens"`
OutputTokens int `json:"output_tokens"`
} `json:"usage"`
}
type ClaudeToolChoice struct {
Type ToolChoiceType `json:"type"`
Name string `json:"name,omitempty"`
}
type ToolChoiceType string
const (
ToolChoiceTypeAuto ToolChoiceType = "auto"
ToolChoiceTypeCode ToolChoiceType = "any"
ToolChoiceTypeTool ToolChoiceType = "tool"
)
type ToolDefinition struct {
Name string `json:"name"`
Description string `json:"description"`
InputSchema JSONSchema `json:"input_schema"`
}
Some examples for what this can be used:
Calling a function, as shown in my example, which is probably the most obvious use case of LLM “function calling”.
- The full flow of the function calling feature would include actually responding to Claude with the tool/function’s output and generally having a back & forth until Claude responds with its final message.
Generating structured data for API requests: You can use the LLM to generate JSON payloads for API requests based on user input, ensuring that the data conforms to the required schema.
Data validation and transformation: The LLM can be used to validate and transform input data into a structured format, which can then be used for further processing or storage in a database.
Generally whenever you have some text & want to extract or transform it into a specific structure, without wanting to train your own custom classifier.
That’s it! Please let me know if this was useful, or maybe share how you are using LLMs in your applications.
Bis demnächst!
~ Martin
Notes:
Any code is provided as-is, under the MIT license.
The cover image is AI generated.