التوجيه والسلاسل الشرطية

Sequential chains always run the same steps in order. But real workflows branch: a billing question and a coding question need different handling.

This lesson covers routing chains that choose a path based on the input.

A router inspects the input, decides which destination chain best fits, and forwards the input there. It is a dispatcher in front of several specialist chains.

Each destination is a normal chain tuned for one kind of task, with its own prompt and configuration.

The router needs to know what each destination is for. Provide a name and a short description so it can match an input to the right one.

One approach asks the LLM itself to classify the input and name the destination. It is flexible and handles fuzzy intent, but adds a model call and some unpredictability.

When intent is clear from structure, a deterministic function can route faster and cheaper than an LLM.

Some inputs match no specialist. Always provide a default destination so the chain never fails on an unexpected request.

LangChain's routing chain ties the router, the destination map, and the default together into a single callable that picks the path automatically.

Routing is a prediction and can be wrong. Log the chosen destination, let a destination decline and fall back, and monitor misroute rates so you can refine descriptions over time.

Choose your router by the situation:

• Rules: fast, cheap, deterministic; brittle for fuzzy intent
• LLM: flexible, handles nuance; slower, costs a call, less predictable

Hybrids use rules first and fall back to the LLM.

Putting it together:

• Build specialist destination chains
• Describe each route clearly
• Choose rule-based, LLM-based, or hybrid routing
• Always include a default and monitor misroutes

You learned to build branching, conditional chains.

• A router dispatches input to specialist destination chains
• Routing can be rule-based, LLM-based, or hybrid
• Clear route descriptions improve accuracy
• Always include a default and monitor misroutes