Respan IBM watsonx.ai Instrumentation

Native Respan instrumentation for the official ibm-watsonx-ai Python SDK.

Installation

pip install respan-ai respan-instrumentation-watsonx ibm-watsonx-ai

Usage

from respan import Respan
from respan_instrumentation_watsonx import WatsonxInstrumentor

respan = Respan(
    api_key="RESPAN_API_KEY",
    instrumentations=[WatsonxInstrumentor()],
)

The instrumentor traces:

• ModelInference.generate()
• ModelInference.generate_text()
• ModelInference.generate_text_stream()
• ModelInference.chat()
• ModelInference.chat_stream()
• ModelInference.agenerate()
• ModelInference.agenerate_stream()
• ModelInference.achat()
• ModelInference.achat_stream()
• Embeddings.generate()
• Embeddings.embed_query()
• Embeddings.embed_documents()
• Embeddings.agenerate()
• Embeddings.aembed_query()
• Embeddings.aembed_documents()

Chat and text-generation spans use canonical GenAI fields such as gen_ai.prompt.*, gen_ai.completion.*, llm.request.functions, and token usage attributes. Stream calls are finalized on normal completion, early close, or failure and carry the canonical streaming flag. Embedding spans retain the complete numeric vector payload, as required by the Respan span contract.

Create the model/client inside application code and close it according to the IBM SDK lifecycle. Always call respan.shutdown() in an outer finally block so buffered spans are exported even when the provider call fails.