[BOT ISSUE] LangChain4j instrumentation only supports OpenAI-backed chat models

Bắt đầu với BraintrustLangchain.java, đặc biệt là các kiểm tra model và phần wrapping dành riêng cho OpenAI, sau đó đọc LangchainInstrumentationModule.java và WrappedHttpClient.java. Theo dõi cách advice hiện có của OpenAI builder và HTTP client wrapper hoạt động. Hoàn thành có nghĩa là các provider LangChain4j không phải OpenAI được liệt kê nhận được instrumentation tương đương ở cấp LLM trong cả đường dẫn wrapping và auto-instrumentation.

Summary

The LangChain4j instrumentation only wraps OpenAiChatModel and OpenAiStreamingChatModel. All other LangChain4j model providers — including Anthropic, Google Gemini, Azure OpenAI, and Mistral — are explicitly skipped with a warning, producing no LLM-level span data for those calls.

Unlike the Spring AI integration (where unsupported providers can fall through to direct SDK auto-instrumentation, e.g. the GenAI module intercepts Client$Builder.build()), LangChain4j model providers use their own internal HTTP clients rather than the upstream SDK clients. This means no other auto-instrumentation module catches these calls.

What is missing

BraintrustLangchain.wrap(OpenTelemetry, AiServices) (lines 36-51) performs instanceof checks against only two model classes:

if (chatModel instanceof OpenAiChatModel oaiModel) {
    aiServices.chatModel(wrap(openTelemetry, oaiModel));
} else {
    log.warn("unsupported model: {}. LLM calls will not be instrumented",
            chatModel.getClass().getName());
}

The auto-instrumentation module (LangchainInstrumentationModule) similarly only registers ByteBuddy advice for OpenAiChatModel$OpenAiChatModelBuilder.build() and OpenAiStreamingChatModel$OpenAiStreamingChatModelBuilder.build().

The most impactful missing model providers are:

Since the existing OpenAI instrumentation works by wrapping LangChain4j's internal HttpClient interface (extracted via reflection from the model's internal client), the same pattern could apply to other providers — each uses the same dev.langchain4j.http.client.HttpClient interface internally.

Braintrust docs status

Braintrust docs do not have a LangChain4j-specific page. The Java SDK is listed as supported for OpenAI, Anthropic, and Gemini providers at https://www.braintrust.dev/docs/integrations/ai-providers, but LangChain4j as a framework integration is not_found in docs.

Upstream sources

• LangChain4j supports 20+ model providers as first-party integrations: https://github.com/langchain4j/langchain4j
• LangChain4j model provider modules: langchain4j-anthropic, langchain4j-google-ai-gemini, langchain4j-vertex-ai-gemini, langchain4j-azure-open-ai, langchain4j-mistral-ai, etc.
• All LangChain4j model providers use their own dev.langchain4j.http.client.HttpClient internally (not the upstream provider SDK), so upstream SDK auto-instrumentation does not cover them.

Local files inspected

• braintrust-sdk/instrumentation/langchain_1_8_0/src/main/java/dev/braintrust/instrumentation/langchain/v1_8_0/BraintrustLangchain.java — lines 36-51 (instanceof check limits to OpenAI only), lines 88-152 (OpenAI-specific wrapping via internal HttpClient)
• braintrust-sdk/instrumentation/langchain_1_8_0/src/main/java/dev/braintrust/instrumentation/langchain/v1_8_0/auto/LangchainInstrumentationModule.java — lines 55-68, 85-100 (auto-instrumentation only targets OpenAI builders)
• braintrust-sdk/instrumentation/langchain_1_8_0/src/main/java/dev/braintrust/instrumentation/langchain/v1_8_0/WrappedHttpClient.java — the HTTP client wrapper that could be reused for other providers