Ai Langchain Patterns

LangChain Patterns Agent

Purpose

Designs LangChain architectures with LCEL pipelines, multi-strategy retrieval, tool-calling agents, memory management, streaming, and production guardrails — production-grade LLM chains.

Agent Protocol

Trigger

User request includes: LangChain, LlamaIndex, chains, LCEL, retriever, agent, tool, callback, Runnable, document loader, text splitter, vector store, memory, LangSmith, deep LangChain patterns.

Protocol

1. Clarify task type (RAG, agent, chain, tool-use) and LLM provider.
2. Select LCEL composition pattern (sequence, parallel, streaming, branching).
3. Design retriever strategy (base, multi-query, ensemble, compression, parent-document).
4. Choose agent type (tool-calling, ReAct, custom) and define tools with validated schemas.
5. Configure memory (conversation buffer, summary, window, entity, token-bounded).
6. Set up document pipeline (loaders, splitters, transformers).
7. Wire streaming and event handlers.
8. Add production guardrails: retry, circuit breaker, cost tracking, observability.

Output

LangChain architecture with chain/retriever/agent patterns, memory strategy, streaming setup, production config.

Response Format

## LangChain Architecture
### LCEL Pipeline
Type: {RunnableSequence/RunnableParallel/RunnableBranch}
Steps: [{step1}, {step2}, ...]
Streaming: {async/event handlers} | Batch Size: {N}

### Retriever
Strategy: {base/multi-query/ensemble/compression/parent-document}
Retrievers: [{type, top-K, params}]
Final Top-K: {N}

### Agent
Type: {tool-calling/ReAct/custom}
Tools: [{name, description, schema}]
Max Iterations: {N} | Max Execution Time: {seconds}
Early Stopping: {generate/error}

### Memory
Type: {conversation/summary/buffer/window/entity}
Window Size: {N turns} | Summary LLM: {model}
Entity Store: {enabled/disabled}

### Document Pipeline
Loaders: [{source, type}]
Splitter: {strategy} | Chunk Size: {tokens} | Overlap: {tokens}
Transformer: {type, params}

### Production
Retry: {strategy} | Cache: {type}
Tracing: {LangSmith project}
Cost Tracking: {enabled}

No preamble. No postamble. No explanations. No filler/hedging/transitions. Compress output.

Completion Criteria

• [ ] LCEL pipeline uses RunnableSequence/RunnableParallel/RunnableBranch as appropriate.
• [ ] Retriever strategy matches data distribution and query patterns.
• [ ] Agent tools have validated schemas and error handling.
• [ ] Memory strategy bounds token consumption.
• [ ] Document pipeline handles source heterogeneity.
• [ ] Streaming is wired for production (async generators or event handlers).
• [ ] Error handling and retry logic are specified.
• [ ] LangSmith tracing is configured for observability.

---

Decision Trees

Chain Type Decision Tree

Is it a single-turn Q&A?
  ├─ With external knowledge ──► RAG chain (retriever + prompt + llm)
  ├─ No external knowledge ────► LLM chain (prompt + llm + parser)
  └─ Multiple docs ───────────► Stuff/MapReduce/Refine chain
Is it multi-turn conversation?
  ├─ With memory ──► ConversationChain + memory
  ├─ With retrieval ──► ConversationalRetrievalChain
  └─ With tools ──► Agent with memory
Is it structured output?
  ├─ JSON ──► LCEL + JsonOutputParser
  ├─ Pydantic ──► LCEL + PydanticOutputParser
  └─ Enum/category ──► RunnableBranch or RouterChain
Is it long document processing?
  ├─ Single doc ──► Stuff chain (if fits context)
  ├─ Summarize ──► MapReduce / Refine chain
  └─ Extract ──► LCEL + structured extraction

Memory Type Decision Tree

Does the conversation need full history?
  ├─ Yes, short ( ChainArch:
    if has_knowledge_base and not is_multi_turn:
        return "rag"
    if has_knowledge_base and is_multi_turn:
        return "agent"
    if needs_structured_output:
        return "structured"
    if is_multi_turn:
        return "conversational"
    if document_count > 1:
        return "map_reduce" if document_size_tokens > 4000 else "stuff"
    return "llm"

def select_memory_type(
    expected_turns: int,
    needs_entity_tracking: bool,
    max_token_budget: int = 4000,
) -> str:
    if needs_entity_tracking:
        return "entity"
    if expected_turns > 20 and max_token_budget  10:
        return "window"
    return "buffer"

def select_retriever_strategy(
    ambiguity_level: float,
    has_sparse_retriever: bool,
    has_dense_retriever: bool,
    noise_tolerance: float,
) -> str:
    if ambiguity_level > 0.7:
        return "multi_query"
    if has_sparse_retriever and has_dense_retriever:
        return "ensemble"
    if noise_tolerance  str:
    parts = []
    for src, docs in sources.items():
        for d in docs:
            d.metadata["source_type"] = src
            parts.append(d)
    return format_docs(sorted(parts, key=lambda x: x.metadata.get("score", 0), reverse=True)[:10])

fused_rag = (
    {"context": parallel_retrieval | fuse_sources, "question": RunnablePassthrough()}
    | prompt | llm | StrOutputParser()
)

Pattern 5: Conditional Routing (RunnableBranch)

from langchain_core.runnables import RunnableBranch

tech_prompt = ChatPromptTemplate.from_template("Answer tech question: {query}")
general_prompt = ChatPromptTemplate.from_template("Answer: {query}")
code_prompt = ChatPromptTemplate.from_template("Write code for: {query}")

branch_chain = RunnableBranch(
    (lambda x: any(tag in x["query"].lower() for tag in ["python", "java", "code"]),
     code_prompt | llm | StrOutputParser()),
    (lambda x: any(tag in x["query"].lower() for tag in ["server", "api", "database"]),
     tech_prompt | llm | StrOutputParser()),
    general_prompt | llm | StrOutputParser(),
)

Pattern 6: Tool-Calling Agent

from langchain.agents import create_tool_calling_agent, AgentExecutor
from langchain_core.tools import tool
from langchain_core.prompts import ChatPromptTemplate

@tool
def search_kb(query: str, top_k: int = 5) -> str:
    """Search the knowledge base. Use for factual questions."""
    results = vectorstore.similarity_search(query, k=top_k)
    return "\n\n".join(d.page_content for d in results)

@tool
def calculator(expression: str) -> str:
    """Evaluate a math expression. Use for calculations."""
    try:
        import ast, operator
        ops = {ast.Add: operator.add, ast.Sub: operator.sub,
               ast.Mult: operator.mul, ast.Div: operator.truediv}
        tree = ast.parse(expression, mode="eval")
        return str(eval(compile(tree, "", "eval")))
    except Exception as e:
        return f"Error: {e}"

tools = [search_kb, calculator]
prompt = ChatPromptTemplate.from_messages([
    ("system", "You are a helpful assistant with tools. Use them when needed."),
    ("placeholder", "{chat_history}"),
    ("human", "{input}"),
    ("placeholder", "{agent_scratchpad}"),
])

agent = create_tool_calling_agent(llm, tools, prompt)
agent_executor = AgentExecutor(
    agent=agent,
    tools=tools,
    max_iterations=10,
    max_execution_time=30,
    early_stopping_method="generate",
    handle_parsing_errors=True,
    return_intermediate_steps=True,
)

Pattern 7: Streaming with astream_events

from typing import AsyncIterator

async def stream_response(chain, input_data: dict) -> AsyncIterator[dict]:
    async for event in chain.astream_events(input_data, version="v2"):
        kind = event["event"]
        if kind == "on_chat_model_stream":
            chunk = event["data"]["chunk"]
            if hasattr(chunk, "content") and chunk.content:
                yield {"type": "token", "content": chunk.content}
        elif kind == "on_retriever_end":
            docs = event["data"]["output"]
            yield {"type": "sources", "documents": [
                {"content": d.page_content[:200], "source": d.metadata.get("source", ""),
                 "score": d.metadata.get("score", 0.0)} for d in docs
            ]}
        elif kind == "on_tool_end":
            yield {"type": "tool_result", "name": event.get("name", ""),
                   "output": str(event["data"]["output"])[:200]}
        elif kind == "on_chain_end":
            yield {"type": "done"}

Pattern 8: Callback for Cost Tracking

from langchain.callbacks.base import BaseCallbackHandler
from collections import defaultdict

class CostTrackingHandler(BaseCallbackHandler):
    MODEL_COSTS = {
        "gpt-4o": {"input": 2.50 / 1e6, "output": 10.00 / 1e6},
        "gpt-4o-mini": {"input": 0.15 / 1e6, "output": 0.60 / 1e6},
        "claude-sonnet-4-20250514": {"input": 3.00 / 1e6, "output": 15.00 / 1e6},
        "claude-haiku-3-5-20241022": {"input": 0.80 / 1e6, "output": 4.00 / 1e6},
    }

    def __init__(self):
        self.total_input_tokens = 0
        self.total_output_tokens = 0
        self.total_cost = 0.0
        self.model_name = "unknown"
        self.run_counts = defaultdict(int)

    def on_llm_start(self, serialized, prompts, **kwargs):
        self.model_name = serialized.get("kwargs", {}).get("model_name", "unknown")
        self.run_counts[self.model_name] += 1

    def on_llm_end(self, response, **kwargs):
        usage = response.llm_output.get("token_usage", {}) if response.llm_output else {}
        inp = usage.get("prompt_tokens", 0)
        out = usage.get("completion_tokens", 0)
        self.total_input_tokens += inp
        self.total_output_tokens += out
        costs = self.MODEL_COSTS.get(self.model_name, {"input": 0, "output": 0})
        self.total_cost += inp * costs["input"] + out * costs["output"]

    def get_report(self) -> dict:
        return {
            "model": self.model_name,
            "calls": self.run_counts[self.model_name],
            "input_tokens": self.total_input_tokens,
            "output_tokens": self.total_output_tokens,
            "total_tokens": self.total_input_tokens + self.total_output_tokens,
            "estimated_cost_usd": round(self.total_cost, 6),
        }

Pattern 9: Production RAG with Caching and Tracing

from langchain.globals import set_llm_cache
from langchain.cache import RedisCache
from langchain.callbacks.tracers import LangChainTracer
from langchain_core.runnables import RunnableConfig

set_llm_cache(RedisCache(redis_client))

tracer = LangChainTracer(project_name="my-rag-prod")

def invoke_with_observability(question: str, user_id: str) -> str:
    config = RunnableConfig(
        callbacks=[CostTrackingHandler(), tracer],
        metadata={"user_id": user_id, "env": "production"},
        tags=["rag", "production"],
    )
    return rag_chain.invoke({"question": question}, config=config)

---

Production Considerations

Error Handling Strategy

| Error Type | Detection | Recovery | Retry Strategy | |---|---|---|---| | RateLimitError | Exception from provider | Exponential backoff + jitter | 3 attempts, 1s-10s | | APITimeoutError | Request timeout | Circuit breaker after 5 failures | 30s cooldown | | ContextLengthExceeded | Token count in callback | Truncate oldest messages | Immediate retry | | BadRequestError | Exception from provider | Validate input schema | No retry | | ToolExecutionError | Tool raises exception | Return structured error string | No retry |

Circuit Breaker Pattern

class CircuitBreaker:
    def __init__(self, threshold: int = 5, recovery_timeout: float = 30.0):
        self.threshold = threshold
        self.recovery_timeout = recovery_timeout
        self.failures = 0
        self.last_failure = 0.0
        self.state = "closed"

    def __call__(self, func):
        async def wrapper(*args, **kwargs):
            if self.state == "open":
                if (time.time() - self.last_failure) > self.recovery_timeout:
                    self.state = "half-open"
                else:
                    return {"error": "service_unavailable",
                            "message": "Circuit breaker open. Please retry later."}
            try:
                result = await func(*args, **kwargs)
                if self.state == "half-open":
                    self.state = "closed"
                    self.failures = 0
                return result
            except Exception:
                self.failures += 1
                self.last_failure = time.time()
                if self.failures >= self.threshold:
                    self.state = "open"
                raise
        return wrapper

LangSmith Trace Configuration

from langsmith import Client
from langchain.callbacks.tracers import LangChainTracer
from langchain_core.tracers.context import tracing_v2_enabled

client = Client()

# Option 1: Per-call tracing
tracer = LangChainTracer(
    project_name="my-project",
    client=client,
)
result = chain.invoke(input, config={"callbacks": [tracer]})

# Option 2: Context manager
with tracing_v2_enabled(project_name="my-project"):
    result = chain.invoke(input)

# Option 3: Environment variables
# LANGCHAIN_TRACING_V2=true
# LANGCHAIN_PROJECT=my-project
# LANGCHAIN_ENDPOINT=https://api.smith.langchain.com

Cost Budget Manager

class BudgetManager:
    def __init__(self, monthly_budget_usd: float):
        self.monthly_budget = monthly_budget_usd
        self.spent = 0.0
        self.daily_log: list[dict] = []

    def track(self, cost: float, operation: str):
        self.spent += cost
        self.daily_log.append({
            "timestamp": time.time(),
            "cost": cost,
            "operation": operation,
        })
        if self.spent > self.monthly_budget:
            logger.warning(f"Monthly budget ${self.monthly_budget} exceeded: ${self.spent:.2f}")
            return False
        return True

    def remaining(self) -> float:
        return max(0.0, self.monthly_budget - self.spent)

    def daily_spend(self) -> float:
        today = time.time() - 86400
        return sum(e["cost"] for e in self.daily_log if e["timestamp"] > today)

---

Anti-Patterns

Anti-Pattern 1: Over-Abstraction

# BAD: Heavy inheritance for simple transformations
class QueryProcessor(BaseRunnable):
    def __init__(self, rules: list[Callable]):
        self.rules = rules
    def invoke(self, input: str, config: RunnableConfig | None = None) -> str:
        result = input
        for rule in self.rules:
            result = rule(result)
        return result

# GOOD: Direct RunnableLambda composition
chain = RunnableLambda(str.strip) | RunnableLambda(str.lower) | llm

Anti-Pattern 2: Ignoring Async

# BAD: Synchronous calls blocking async context
def handle_request(query: str):
    return chain.invoke(query)  # Blocks event loop

# GOOD: Async throughout
async def handle_request(query: str):
    return await chain.ainvoke(query)

Anti-Pattern 3: Memory Leaks via Unbounded History

# BAD: No bounds on memory
memory = ConversationBufferMemory(return_messages=True)  # Grows forever

# GOOD: Bounded memory
memory = ConversationBufferWindowMemory(k=10, return_messages=True)
# OR with token limit
from langchain.memory import ConversationTokenBufferMemory
memory = ConversationTokenBufferMemory(llm=llm, max_token_limit=4000)

Anti-Pattern 4: Callback Spaghetti

# BAD: Inline callbacks doing too much
chain.invoke(input, config={"callbacks": [
    handler_a, handler_b, handler_c, handler_d,
    handler_e, handler_f, handler_g,
]})

# GOOD: Composite handler
class CompositeHandler(BaseCallbackHandler):
    def __init__(self):
        self.logger = LoggingHandler()
        self.cost = CostTrackingHandler()
        self.latency = LatencyTrackingHandler()

    def on_llm_end(self, response, **kwargs):
        self.logger.on_llm_end(response, **kwargs)
        self.cost.on_llm_end(response, **kwargs)
        self.latency.on_llm_end(response, **kwargs)

chain.invoke(input, config={"callbacks": [CompositeHandler()]})

Anti-Pattern 5: No Max Iterations on Agents

# BAD: Agent can loop forever
executor = AgentExecutor(agent=agent, tools=tools)

#

…

## Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

- **Author:** [j4flmao](https://github.com/j4flmao)
- **Source:** [j4flmao/agent-skills](https://github.com/j4flmao/agent-skills)
- **License:** MIT

Install and usage instructions live in the source repository linked above.