langgraph-routing（语言图路由）

LangGraph Conditional Routing

Route workflow execution dynamically based on state.

Basic Conditional Edge

from langgraph.graph import StateGraph, END

def route_based_on_quality(state: WorkflowState) -> str:
    """Decide next step based on quality score."""
    if state["quality_score"] >= 0.8:
        return "publish"
    elif state["retry_count"] < 3:
        return "retry"
    else:
        return "manual_review"

workflow.add_conditional_edges(
    "quality_check",
    route_based_on_quality,
    {
        "publish": "publish_node",
        "retry": "generator",
        "manual_review": "review_queue"
    }
)

Quality Gate Pattern

def route_after_quality_gate(state: AnalysisState) -> str:
    """Route based on quality gate result."""
    if state["quality_passed"]:
        return "compress_findings"
    elif state["retry_count"] < 2:
        return "supervisor"  # Retry
    else:
        return END  # Return partial results

workflow.add_conditional_edges(
    "quality_gate",
    route_after_quality_gate,
    {
        "compress_findings": "compress_findings",
        "supervisor": "supervisor",
        END: END
    }
)

Retry Loop Pattern

def llm_call_with_retry(state):
    """Retry failed LLM calls."""
    try:
        result = call_llm(state["input"])
        state["output"] = result
        state["retry_count"] = 0
        return state
    except Exception as e:
        state["retry_count"] += 1
        state["error"] = str(e)
        return state

def should_retry(state) -> str:
    if state.get("output"):
        return "success"
    elif state["retry_count"] < 3:
        return "retry"
    else:
        return "failed"

workflow.add_conditional_edges(
    "llm_call",
    should_retry,
    {
        "success": "next_step",
        "retry": "llm_call",  # Loop back
        "failed": "error_handler"
    }
)

Routing Patterns

Sequential:    A → B → C              (simple edges)
Branching:     A → (B or C)           (conditional edges)
Looping:       A → B → A              (retry logic)
Convergence:   (A or B) → C           (multiple inputs)
Diamond:       A → (B, C) → D         (parallel then merge)

State-Based Router

def dynamic_router(state: WorkflowState) -> str:
    """Route based on multiple state conditions."""
    if state.get("error"):
        return "error_handler"
    if not state.get("validated"):
        return "validator"
    if state["confidence"] < 0.5:
        return "enhance"
    return "finalize"

Command vs Conditional Edges (2026 Best Practice)

from langgraph.types import Command
from typing import Literal

# Use CONDITIONAL EDGES when: Pure routing, no state updates
def simple_router(state: WorkflowState) -> str:
    if state["score"] > 0.8:
        return "approve"
    return "reject"

workflow.add_conditional_edges("evaluate", simple_router)

# Use COMMAND when: Updating state AND routing together
def router_with_state(state: WorkflowState) -> Command[Literal["approve", "reject"]]:
    if state["score"] > 0.8:
        return Command(
            update={"route_reason": "high score", "routed_at": time.time()},
            goto="approve"
        )
    return Command(
        update={"route_reason": "low score", "routed_at": time.time()},
        goto="reject"
    )

workflow.add_node("evaluate", router_with_state)
# No conditional edges needed - Command handles routing

Semantic Routing Implementation

from sentence_transformers import SentenceTransformer
import numpy as np

embedder = SentenceTransformer("all-MiniLM-L6-v2")

# Pre-compute route embeddings
ROUTE_EMBEDDINGS = {
    "technical": embedder.encode("technical implementation code programming engineering"),
    "business": embedder.encode("business strategy revenue customers sales marketing"),
    "support": embedder.encode("help troubleshoot error problem fix support issue"),
    "creative": embedder.encode("design creative writing content marketing copy"),
}

def semantic_router(state: WorkflowState) -> str:
    """Route based on semantic similarity."""
    query = state["query"]
    query_embedding = embedder.encode(query)

    # Calculate cosine similarities
    similarities = {}
    for route, route_embedding in ROUTE_EMBEDDINGS.items():
        similarity = np.dot(query_embedding, route_embedding) / (
            np.linalg.norm(query_embedding) * np.linalg.norm(route_embedding)
        )
        similarities[route] = similarity

    # Return highest similarity route
    best_route = max(similarities, key=similarities.get)

    # Optional: threshold check
    if similarities[best_route] < 0.3:
        return "general"  # Fallback

    return best_route

workflow.add_conditional_edges(
    "classifier",
    semantic_router,
    {
        "technical": "tech_agent",
        "business": "business_agent",
        "support": "support_agent",
        "creative": "creative_agent",
        "general": "general_agent"
    }
)

Key Decisions

Common Mistakes

• No END fallback (workflow hangs)
• Infinite loops (no max retry)
• Side effects in router (hard to debug)
• Missing edge mappings (runtime error)

Evaluations

See references/evaluations.md for test cases.

Related Skills

• langgraph-state - State design for routing decisions
• langgraph-supervisor - Supervisor pattern with dynamic routing
• langgraph-parallel - Route to parallel branches
• langgraph-human-in-loop - Route based on human decisions
• langgraph-tools - Route after tool execution results
• agent-loops - ReAct loop patterns with conditional routing

Capability Details

conditional-routing

Keywords: conditional, branch, decision, if-else Solves:

• Route based on conditions
• Implement branching logic
• Create decision nodes

semantic-routing

Keywords: semantic, embedding, similarity, intent Solves:

• Route by semantic similarity
• Intent-based routing
• Embedding-based decisions

router-template

Keywords: template, router, semantic, implementation Solves:

• Semantic router template
• Production router code
• Copy-paste implementation