Package (Required)
- [x] langchain
- [x] langchain-core

Feature Description
Current fallback logic in LangChain is reactive, only triggering when a model crashes or times out. I propose a TieredSemanticRouter that intelligently routes traffic between small, fast models like GPT-4o-mini or Llama-3-8B and larger, more complex models based on a confidence check.

Motivation
Enterprise users often default to expensive models to avoid edge-case failures, leading to much higher costs than necessary. A proactive router allows simple traffic to be handled at a lower cost while maintaining a high-quality fallback for complex queries.

Proposed API
```python
from langchain_core.runnable import TieredSemanticRouter

small_model = ChatOpenAI(model="gpt-4o-mini")
large_model = ChatOpenAI(model="gpt-4o")

router = TieredSemanticRouter(
    primary=small_model, 
    fallback=large_model,
    threshold=0.85
)

chain = prompt | router | parser
```

Key Use Cases
1. Cost Reduction: Offloading basic tasks to smaller models.
2. Latency Optimization: Responding faster to simple queries.
3. Profitability: Improving the unit economics of AI applications.