unknown type for gen_ai spans

[xrmx]

When tracing a python langchain application with opentelemetry-instrumentation-langchain (provided by openllmetry) we may produce spans like this:

{
    "name": "AzureChatOpenAI.chat",
    "context": {
        "trace_id": "0xdf0874b5202669e839de54311930359c",
        "span_id": "0x731e8d3b3a82268f",
        "trace_state": "[]"
    },
    "kind": "SpanKind.CLIENT",
    "parent_id": "0xdf313e3a840d43d3",
    "start_time": "2024-10-15T08:04:04.522445Z",
    "end_time": "2024-10-15T08:04:06.125744Z",
    "status": {
        "status_code": "UNSET"
    },
    "attributes": {
        "traceloop.workflow.name": "ConversationalRetrievalChain",
        "traceloop.entity.path": "StuffDocumentsChain.LLMChain",
        "gen_ai.system": "Langchain",
        "llm.request.type": "chat",
        "gen_ai.request.model": "gpt-4o-mini",
        "gen_ai.response.model": "gpt-4o-mini",
        "gen_ai.request.temperature": 0.7,
        "gen_ai.prompt.0.role": "user",
        "gen_ai.prompt.0.content": "somecontent",
        "gen_ai.completion.0.role": "assistant"
    },
    "events": [],
    "links": [],
    "resource": {
        "attributes": {
            "telemetry.sdk.language": "python",
            "telemetry.sdk.name": "opentelemetry",
            "telemetry.sdk.version": "1.27.0",
            "service.name": "foobar",
            "service.version": "1.0",
            "deployment.environment": "production",
            "service.instance.id": "gcpvmlls01",
            "telemetry.auto.version": "0.48b0"
        },
        "schema_url": ""
    }
}

that ends up having an unknown span type in kibana which does not look great.

Please note that this span is not following the OTel genai convention but if we want to add a special type for genai stuff the presence of the gen_ai.system attribute is in the semconv while this particular value (langchain) is not.

Here's the current enrichment code setting the span.type:

opentelemetry-lib/enrichments/trace/internal/elastic/span.go

Lines 342 to 372 in 5f809d4

	func (s *spanEnrichmentContext) setSpanTypeSubtype(span ptrace.Span) {
	var spanType, spanSubtype string
	switch {
	case s.isDB:
	spanType = "db"
	spanSubtype = s.dbSystem
	case s.isMessaging:
	spanType = "messaging"
	spanSubtype = s.messagingSystem
	case s.isRPC:
	spanType = "external"
	spanSubtype = s.rpcSystem
	case s.isHTTP:
	spanType = "external"
	spanSubtype = "http"
	default:
	switch span.Kind() {
	case ptrace.SpanKindInternal:
	spanType = "app"
	spanSubtype = "internal"
	default:
	spanType = "unknown"
	}
	}
	span.Attributes().PutStr(AttributeSpanType, spanType)
	if spanSubtype != "" {
	span.Attributes().PutStr(AttributeSpanSubtype, spanSubtype)
	}
	}

[xrmx]

It looks like gen_ai.system is not a good bet anymore (open-telemetry/semantic-conventions#1613), maybe the presence of anything that starts with gen_ai. is enough.

[xrmx]

Also note that this happens with our openai instrumentation too.