opentelemetry-instrumentation-genai-langchain gaps vs openinference instrumentation

Ready to Fix

Missing Spans

• Add nested workflow span on on_chain_start (langchain_core.callbacks.BaseCallbackHandler.on_chain_start) - maps to GenAI invoke_workflow (#578)
• Classify agent chains by name on on_chain_start (langchain_core.callbacks.BaseCallbackHandler.on_chain_start) - maps to GenAI invoke_agent (#579)

Missing Attributes & Features

• Capture audio input tokens on chat (OpenInference: llm.token_count.prompt_details.audio, semconv: gen_ai.usage.audio.input_tokens) (#580)
• Capture audio output tokens on chat (OpenInference: llm.token_count.completion_details.audio, semconv: gen_ai.usage.audio.output_tokens) (#580)
• Capture top_k on chat (OpenInference: llm.invocation_parameters, semconv: gen_ai.request.top_k) (#581)
• Capture choice count on chat (OpenInference: llm.invocation_parameters, semconv: gen_ai.request.choice.count) (#581)
• Capture request model from ls_model_name metadata on chat (OpenInference: llm.model_name, semconv: gen_ai.request.model) (#582)
• Capture token usage from llm_output and generation_info fallbacks on chat (OpenInference: llm.token_count.prompt / llm.token_count.completion, semconv: gen_ai.usage.input_tokens / gen_ai.usage.output_tokens) (#583)
• Capture document score on retrieval (OpenInference: document.score, semconv: gen_ai.retrieval.documents) (#584)
• Capture prompt template name and variables on chat (OpenInference: llm.prompt_template.template / llm.prompt_template.variables, semconv: gen_ai.prompt.name / gen_ai.prompt.variable) (#585)

Blocked on Semantic Conventions

Missing Spans

• Add workflow node / chain step span on on_chain_start (langchain_core.callbacks.BaseCallbackHandler.on_chain_start) (OpenInference: openinference.span.kind = "CHAIN") (https://github.com/open-telemetry/semantic-conventions-genai/issues/188)

Missing Attributes & Features

• Capture total token count on chat (OpenInference: llm.token_count.total)
• Capture raw invocation parameters blob on chat (OpenInference: llm.invocation_parameters)
• Capture raw metadata JSON blob on all spans (OpenInference: metadata)

Gaps for Legacy APIs

• Add legacy on_llm_start span (langchain_core.callbacks.BaseCallbackHandler.on_llm_start) - maps to GenAI text_completion
• Support legacy prompts on text_completion (OpenInference: llm.prompts)

Start with the LangChain callback entry points named in the issue: on_chain_start, chat, retrieval, and on_llm_start, and compare their current instrumentation with the listed OpenInference mappings. Done means the unblocked checklist items are implemented for the relevant spans and attributes, while items blocked on semantic conventions remain pending.