Componer cadenas LCEL en LangChain para workflows LLM

LangChain Expression Language (LCEL) es una forma declarativa de componer workflows LLM usando el operador pipe (|). Las cadenas se vuelven componibles, streamables y reintentables sin boilerplate. A continuacion: pipelines multi-paso con prompts, modelos, parsers, ramas paralelas y componentes runnable personalizados.

Cuando Usar Esto

• For alternatives, see Complete Guide to LangChain in Production.

• Workflows LLM multi-paso (resumir luego traducir, extraer luego clasificar)

• Pipelines que necesitan streaming, batching o logica de reintento

• Componer prompts, herramientas y parsers en cadenas reutilizables

Requisitos Previos

• Python 3.10+
• Paquetes langchain y langchain-openai

Solucion

1. Instalar dependencias

pip install langchain langchain-openai

2. Cadena basica — Prompt | Model | Parser

from langchain_core.prompts import ChatPromptTemplate
from langchain_openai import ChatOpenAI
from langchain_core.output_parsers import StrOutputParser

prompt = ChatPromptTemplate.from_messages([
    ("system", "You are a concise technical writer."),
    ("user", "Explain {topic} in under 100 words."),
])

model = ChatOpenAI(model="gpt-4o-mini", temperature=0.3)
parser = StrOutputParser()

# El operador pipe de LCEL compone la cadena
chain = prompt | model | parser

result = chain.invoke({"topic": "vector databases"})
print(result)

3. Cadena multi-paso — resumir luego traducir

from langchain_core.runnables import RunnablePassthrough

summarize_prompt = ChatPromptTemplate.from_template(
    "Summarize this text in 3 bullet points:\n\n{text}"
)

translate_prompt = ChatPromptTemplate.from_template(
    "Translate to Spanish:\n\n{summary}"
)

summarize_chain = summarize_prompt | model | StrOutputParser()
translate_chain = translate_prompt | model | StrOutputParser()

# Encadenar la salida de summarize hacia translate
full_chain = (
    {"summary": summarize_chain}
    | translate_chain
)

result = full_chain.invoke({
    "text": "LangChain is a framework for building applications powered by LLMs..."
})

4. Ramas paralelas con RunnableParallel

from langchain_core.runnables import RunnableParallel

# Ejecutar multiples cadenas simultaneamente sobre la misma entrada
classification_prompt = ChatPromptTemplate.from_template(
    "Classify this text as 'technical', 'marketing', or 'other':\n\n{text}"
)

sentiment_prompt = ChatPromptTemplate.from_template(
    "What is the sentiment of this text? (positive, neutral, negative):\n\n{text}"
)

parallel_chain = RunnableParallel(
    classification=classification_prompt | model | StrOutputParser(),
    sentiment=sentiment_prompt | model | StrOutputParser(),
)

result = parallel_chain.invoke({
    "text": "Our new product is amazing and uses cutting-edge AI!"
})
# {'classification': 'marketing', 'sentiment': 'positive'}

5. Componentes Runnable personalizados

from langchain_core.runnables import RunnableLambda

def extract_keywords(text: str) -> list[str]:
    """Extract keywords from text (simplified)."""
    stop_words = {"the", "a", "an", "is", "are", "and", "or", "in", "on"}
    words = [w.lower().strip(".,!?") for w in text.split()]
    return [w for w in words if w not in stop_words and len(w) > 3][:10]

def format_output(data: dict) -> str:
    """Format the final output."""
    return f"Keywords: {', '.join(data['keywords'])}\nSummary: {data['summary']}"

custom_chain = (
    RunnableParallel(
        keywords=RunnableLambda(extract_keywords),
        summary=summarize_chain,
    )
    | RunnableLambda(format_output)
)

result = custom_chain.invoke({
    "text": "Vector databases enable semantic search over embeddings..."
})

6. Agregar reintento y fallback

from langchain_core.runnables import RunnableWithFallbacks

fast_model = ChatOpenAI(model="gpt-4o-mini", temperature=0)
strong_model = ChatOpenAI(model="gpt-4o", temperature=0)

fast_chain = prompt | fast_model | StrOutputParser()
strong_chain = prompt | strong_model | StrOutputParser()

# Intentar modelo rapido primero, fallback a modelo fuerte en error
chain_with_fallback = fast_chain.with_fallbacks([strong_chain])

# Agregar logica de reintento
chain_with_retry = fast_chain.with_retry(
    stop_after_attempt=3,
    wait_exponential_jitter=True,
)

7. Streaming de salida

# Stream tokens mientras llegan
streaming_chain = prompt | model | StrOutputParser()

for chunk in streaming_chain.stream({"topic": "GraphQL federation"}):
    print(chunk, end="", flush=True)

8. Procesamiento batch

# Procesar multiples entradas en paralelo
results = chain.batch([
    {"topic": "Redis caching"},
    {"topic": "PostgreSQL indexing"},
    {"topic": "Docker multi-stage builds"},
])

for r in results:
    print(r)
    print("---")

Como Funciona

1. Operador pipe (|) pasa la salida de un runnable como entrada al siguiente, similar a los pipes de Unix.
2. RunnableParallel ejecuta multiples cadenas sobre la misma entrada simultaneamente, retornando un dict con el resultado de cada rama.
3. RunnableLambda envuelve cualquier funcion Python como un runnable, habilitando logica personalizada dentro de las cadenas.
4. with_fallbacks intenta la cadena principal primero; si lanza una excepcion, prueba cada fallback en orden.
5. with_retry reintentar automaticamente ante fallos con intentos y backoff configurables.
6. stream produce chunks de salida a medida que se generan, habilitando actualizaciones de UI en tiempo real.
7. batch procesa multiples entradas concurrentemente, aprovechando la API batch del proveedor LLM para throughput.

Variantes

Enrutamiento condicional

from langchain_core.runnables import RunnableBranch

def is_technical(input_dict):
    return "technical" in input_dict.get("text", "").lower()

branch_chain = RunnableBranch(
    (is_technical, technical_summary_chain),
    (lambda x: "marketing" in x["text"].lower(), marketing_summary_chain),
    default_chain,  # fallback
)

result = branch_chain.invoke({"text": "This technical guide covers..."})

Salida JSON con parser estructurado

from langchain_core.output_parsers import JsonOutputParser
from pydantic import BaseModel

class ProductReview(BaseModel):
    rating: int
    summary: str
    pros: list[str]
    cons: list[str]

json_parser = JsonOutputParser(pydantic_object=ProductReview)

review_prompt = ChatPromptTemplate.from_template(
    "Analyze this review and extract structured data.\n{format_instructions}\n\nReview: {review}"
).partial(format_instructions=json_parser.get_format_instructions())

review_chain = review_prompt | model | json_parser

result = review_chain.invoke({"review": "Great product, fast delivery..."})
# {'rating': 5, 'summary': 'Positive review', 'pros': [...], 'cons': [...]}

Cadena conversacional con memoria

from langchain_core.chat_history import InMemoryChatMessageHistory
from langchain_core.runnables.history import RunnableWithMessageHistory

conversational_prompt = ChatPromptTemplate.from_messages([
    ("system", "You are a helpful assistant."),
    ("placeholder", "{history}"),
    ("human", "{input}"),
])

base_chain = conversational_prompt | model | StrOutputParser()

chain_with_history = RunnableWithMessageHistory(
    base_chain,
    lambda session_id: InMemoryChatMessageHistory(),
    input_messages_key="input",
    history_messages_key="history",
)

result = chain_with_history.invoke(
    {"input": "What is Redis?"},
    config={"configurable": {"session_id": "user123"}},
)

Mejores Practicas

• Usa StrOutputParser para salida de texto simple — extrae el contenido string del mensaje AI
• Divide cadenas complejas en componentes mas pequenos — cada componente debe hacer una sola cosa
• Usa with_fallbacks para produccion — las APIs LLM pueden rate-limit o timeout
• Stream para salida orientada al usuario — mejora la latencia percibida considerablemente

Errores Comunes

• Olvidar pasar las claves de entrada correctas — cada runnable espera claves de entrada especificas; los desajustes causan errores en runtime
• No usar RunnableParallel para ramas independientes — la ejecucion secuencial pierde tiempo cuando las ramas son independientes
• Encadenar sin manejo de errores — un solo fallo de API rompe toda la cadena
• Usar llamadas sincronas en contextos async — usa ainvoke, astream y abatch en codigo async

Troubleshooting

• Model outputs are inconsistent: set temperature to 0 for deterministic tasks, use seed where supported, and version the prompt.
• Prompt injection leaks context: separate user input from system instructions.
• High token costs: cache embeddings, summarize long context, and choose smaller models for simple tasks.
• Retrieval returns irrelevant chunks: tune chunk size, overlap, and metadata filters. Evaluate retrieval metrics separately from generation.
• Evaluation scores do not match human judgment: Human review is still the ground truth.