Description

When using LangChain with non-OpenAI providers (e.g., ChatSnowflake, ChatAnthropic), the gen_ai.request.model span attribute shows unknown_model instead of the actual model name. This makes it impossible to track which model was used for cost attribution, performance analysis, etc.

Environment

• splunk-otel-instrumentation-langchain: 0.1.x
• langchain-snowflake: 0.2.x (or langchain-anthropic)
• Provider: Snowflake Cortex (ChatSnowflake)

Steps to Reproduce

from langchain_snowflake import ChatSnowflake

llm = ChatSnowflake(model="claude-3-5-sonnet", ...)
response = llm.invoke("Hello")
# Check span attributes in trace

Expected Behavior

Span name: chat claude-3-5-sonnet
gen_ai.request.model: claude-3-5-sonnet

Actual Behavior

Span name: chat unknown_model
gen_ai.request.model: unknown_model

Evidence

Observed in production trace from Snowflake SPCS deployment:

• Trace ID: 90a9d9e471e7ba2a1c88f96e2bc46293
• Multiple chat unknown_model spans despite using claude-3-5-sonnet

{
  "spanId": "ac8f798541d3812c",
  "operationName": "chat unknown_model",
  "tags": {
    "gen_ai.request.model": "unknown_model",
    "gen_ai.response.model": "claude-3-5-sonnet",
    "gen_ai.provider.name": "snowflake",
    "gen_ai.usage.input_tokens": 1108,
    "gen_ai.usage.output_tokens": 99
  }
}

Full trace JSON: trace_flat_hierarchy_bug.json

Root Cause Analysis

The callback handler only checks invocation_params.model_name:

# Current code - only checks model_name (OpenAI convention)
model = invocation_params.get("model_name", "unknown_model")

But ChatSnowflake and ChatAnthropic use invocation_params.model:

# What ChatSnowflake/ChatAnthropic provides
invocation_params = {"model": "claude-3-5-sonnet", ...}

Provider Compatibility

Impact

• ❌ Cost tracking broken - can't attribute costs to specific models
• ❌ Performance analysis impossible - can't compare model latencies
• ❌ Span names misleading - chat unknown_model provides no context
• ❌ Dashboard filters broken - can't filter by model in Splunk O11y