[BOT ISSUE] LangChain4j EmbeddingModel calls are not instrumented

[braintrust-bot]

Summary

The LangChain4j instrumentation wraps OpenAiChatModel, OpenAiStreamingChatModel, and AiServices, but does not instrument any EmbeddingModel implementation. Calls to embeddingModel.embed() or embeddingModel.embedAll() through LangChain4j produce no Braintrust spans, no input/output capture, and no usage metrics.

This is distinct from #60 (which covers missing non-OpenAI chat model providers) and from #71 (which covers Spring AI embedding gaps). LangChain4j's EmbeddingModel is a separate interface from ChatLanguageModel with its own implementations, builders, and HTTP transport.

What is missing

No wrapper methods for EmbeddingModel

BraintrustLangchain.java (lines 22–152) only handles:

• AiServices<T> wrapping (lines 22–85)
• OpenAiChatModel wrapping (lines 87–118)
• OpenAiStreamingChatModel wrapping (lines 120–152)

There is no wrap(OpenTelemetry, EmbeddingModel) method or any reference to embedding models.

No auto-instrumentation for embedding builders

LangchainInstrumentationModule.java (lines 44–145) registers ByteBuddy advice for exactly 3 builder classes:

• OpenAiChatModel$OpenAiChatModelBuilder
• OpenAiStreamingChatModel$OpenAiStreamingChatModelBuilder
• AiServices

No embedding model builder (e.g., OpenAiEmbeddingModel$OpenAiEmbeddingModelBuilder) is targeted.

No fallback to upstream SDK auto-instrumentation

LangChain4j embedding model implementations use LangChain4j's own internal dev.langchain4j.http.client.HttpClient, not the upstream provider SDKs (com.openai:openai-java, etc.). This means the OpenAI auto-instrumentation module (openai_2_8_0) does not intercept embedding calls made through LangChain4j's OpenAiEmbeddingModel.

The WrappedHttpClient class in the LangChain4j module already wraps LangChain4j's internal HTTP client for chat models — the same pattern could be applied to embedding models.

Affected LangChain4j embedding models

LangChain4j class	Provider
OpenAiEmbeddingModel	OpenAI
AzureOpenAiEmbeddingModel	Azure OpenAI
GoogleAiEmbeddingModel	Google AI
VertexAiEmbeddingModel	Google Vertex AI
OllamaEmbeddingModel	Ollama
MistralAiEmbeddingModel	Mistral AI

Braintrust docs status

• LangChain4j is not mentioned on the Braintrust integrations page at https://www.braintrust.dev/docs/integrations/ai-providers: not_found
• No LangChain4j-specific documentation exists for the Java SDK

Upstream sources

• LangChain4j EmbeddingModel interface: https://github.com/langchain4j/langchain4j — EmbeddingModel is a core interface with embed(String), embed(TextSegment), and embedAll(List<TextSegment>) methods
• LangChain4j embedding model implementations: https://docs.langchain4j.dev/integrations/embedding-models — documents 15+ embedding model providers
• LangChain4j OpenAI embedding model: langchain4j-open-ai module provides OpenAiEmbeddingModel as a first-party integration

Local files inspected

• braintrust-sdk/instrumentation/langchain_1_8_0/src/main/java/dev/braintrust/instrumentation/langchain/v1_8_0/BraintrustLangchain.java — lines 22–152 (only chat model and AiServices wrapping; no embedding model handling)
• braintrust-sdk/instrumentation/langchain_1_8_0/src/main/java/dev/braintrust/instrumentation/langchain/v1_8_0/auto/LangchainInstrumentationModule.java — lines 44–145 (only chat model builders and AiServices targeted; no embedding builder)
• braintrust-sdk/instrumentation/langchain_1_8_0/src/main/java/dev/braintrust/instrumentation/langchain/v1_8_0/WrappedHttpClient.java — existing HTTP client wrapper that could be reused for embedding models
• braintrust-sdk/instrumentation/langchain_1_8_0/src/test/java/dev/braintrust/instrumentation/langchain/v1_8_0/BraintrustLangchainTest.java — no embedding model tests exist

[starfolkai]

Drop into this session: sfk devbox session 76

---

Summary

Issue #73: LangChain4j EmbeddingModel calls are not instrumented

Root Cause

Two missing pieces in the instrumentation pipeline:

1. BraintrustLangchain.java had no wrap(OpenTelemetry, OpenAiEmbeddingModel) method — only OpenAiChatModel and OpenAiStreamingChatModel were handled.

2. LangchainInstrumentationModule.java had no ByteBuddy advice targeting OpenAiEmbeddingModel$OpenAiEmbeddingModelBuilder.build() — the agent never intercepted embedding model construction.

As a secondary gap, InstrumentationSemConv.tagOpenAIResponse didn't handle the data field in embeddings responses (only choices/output for chat), so braintrust.output_json would be missing.

Verification

Before fix:

testEmbeddingModelProducesSpan() FAILED
Expected exactly one span for embedding call. Zero spans means EmbeddingModel 
is not instrumented (issue #73). ==> expected: <1> but was: <0>

After fix:

testEmbeddingModelProducesSpan() PASSED
testStreamingChatCompletion() PASSED
testAiServicesWithTools() PASSED
testSyncChatCompletion() PASSED
TracingToolExecutorTest > typeToolJsonCorrect() PASSED
BUILD SUCCESSFUL

Changes (commit 80fea63)

File	Change
BraintrustLangchain.java	Added wrap(otel, OpenAiEmbeddingModel) using same reflection pattern as chat models
LangchainInstrumentationModule.java	Added OpenAiEmbeddingModelBuilderInstrumentation + OpenAiEmbeddingModelBuilderAdvice
InstrumentationSemConv.java	Added data field handling in tagOpenAIResponse for embedding output capture
BraintrustLangchainTest.java	Added testEmbeddingModelProducesSpan() regression test
cassettes/openai/...	VCR cassette for mocked OpenAI /embeddings endpoint

The fix works because OpenAiEmbeddingModel uses the same DefaultOpenAiClient → httpClient internal field path as the chat models. The same HTTP client wrapping pattern applies directly.