Baike.dev
Log in
Back to tool/Back to issues
#1484·langchaingo

llms/openai: `llms.WithMetadata` causes HTTP 400 from OpenAI: "metadata parameter is only allowed when store is enabled"

Author: pedro-a-n-moreiraCreated Mar 16, 2026Updated Mar 16, 2026

Summary

When using llms.WithMetadata(...) to pass request-level metadata to the OpenAI Chat Completions API, the library sends the metadata field without the required top-level "store": true parameter. OpenAI's API rejects the request with a 400 error.

Environment

  • langchaingo version: v0.1.14
  • Go version: 1.25.6
  • OS: macOS
  • OpenAI model: gpt-5-nano (also reproducible on other models)

Steps to Reproduce

go
package main

import (
	"context"
	"log"

	"github.com/tmc/langchaingo/llms"
	openai "github.com/tmc/langchaingo/llms/openai"
)

func main() {
	ctx := context.Background()

	llm, err := openai.New(
		openai.WithModel("gpt-5-nano"),
	)
	if err != nil {
		log.Fatal(err)
	}

	resp, err := llm.Call(
		ctx,
		"Summarize this customer support ticket.",
		llms.WithMetadata(map[string]interface{}{
			"feature":     "support",
			"environment": "local",
			"team":        "ai",
		}),
	)
	if err != nil {
		log.Fatal(err) // <-- error occurs here
	}

	log.Println(resp)
}

Actual Result

API returned unexpected status code: 400: The 'metadata' parameter is only allowed when 'store' is enabled.

Expected Result

The call succeeds. Metadata is forwarded to OpenAI alongside "store": true, which is the required flag for request-level metadata per OpenAI's documentation.

Root Cause

llms.WithMetadata values (after filtering out internal keys like openai:* and thinking_config) are passed directly into the ChatRequest.Metadata field in llms/openai/internal/openaiclient/chat.go. However, ChatRequest has no Store field and never sets "store": true in the serialized JSON body.

OpenAI treats metadata and store as co-dependent: store: true must be present for metadata to be accepted. Sending metadata without store results in an immediate 400.

Relevant code in llms/openai/openaillm.go:

go
// Filter out internal metadata that shouldn't be sent to API
apiMetadata := make(map[string]any)
if opts.Metadata != nil {
    for k, v := range opts.Metadata {
        if k == "thinking_config" || strings.HasPrefix(k, "openai:") {
            continue
        }
        apiMetadata[k] = v
    }
}
if len(apiMetadata) == 0 {
    apiMetadata = nil
}

req := &openaiclient.ChatRequest{
    // ...
    Metadata: apiMetadata, // sent without store: true
}

ChatRequest in llms/openai/internal/openaiclient/chat.go has no Store field:

go
type ChatRequest struct {
    Model               string         `json:"model"`
    Messages            []*ChatMessage `json:"messages"`
    // ... all other fields ...
    Metadata            map[string]any `json:"metadata,omitempty"`
    // Store bool is missing
}

Proposed Fix

  1. Add a Store bool field to ChatRequest:
go
type ChatRequest struct {
    // ... existing fields ...
    Store    bool           `json:"store,omitempty"`
    Metadata map[string]any `json:"metadata,omitempty"`
}
  1. In openaillm.go, set Store: true whenever non-empty metadata is being sent:
go
req := &openaiclient.ChatRequest{
    // ...
    Store:    len(apiMetadata) > 0,
    Metadata: apiMetadata,
}

Alternatively, expose a dedicated WithStore(bool) call option so callers can opt in to storage explicitly, and Store is automatically set to true when metadata is provided.

Workaround

Until this is fixed, call the OpenAI Chat Completions API directly with "store": true in the request body:

go
type chatRequest struct {
    Model    string            `json:"model"`
    Messages []chatMessage     `json:"messages"`
    Store    bool              `json:"store"`
    Metadata map[string]string `json:"metadata,omitempty"`
}
// ... build and POST to https://api.openai.com/v1/chat/completions manually

References

Source: tmc/langchaingo

View original on GitHubView discussion on GitHub