Compresión contextual con LLM
Optimice el contexto que se proporciona al LLM filtrando y comprimiendo dinámicamente los documentos recuperados para centrarse en la relevancia.
Compresión contextual con LLM es una lección gratuita de LangChain / RAG / Vector DBs en CoddyKit. Esta es la lección 2 de 4. Puedes leer la lección completa abajo gratuitamente — luego la practicas en el navegador con un editor de código integrado y un tutor de IA 24/7. Forma parte de la ruta de aprendizaje de LangChain / RAG / Vector DBs, y tu progreso se sincroniza en la web y la app de CoddyKit. El curso de LangChain / RAG / Vector DBs incluye 4 lecciones en total.
Partes de esta lección aún no han sido traducidas y se muestran en inglés.
Why Compress Context?
When building RAG (Retrieval Augmented Generation) systems, Large Language Models (LLMs) have a limited context window. Feeding them too much irrelevant information can lead to several problems:
- Performance issues: LLMs might get confused by noise.
- Higher costs: More tokens mean higher API bills.
- Slower responses: More text takes longer to process.
This is where contextual compression comes in!
What is Contextual Compression?
Contextual compression is a technique used to refine the documents retrieved by your RAG system before they are passed to the LLM. It acts as a smart filter and extractor.
- Filter: Remove entire documents that are less relevant.
- Extract: From the remaining documents, identify and keep only the most pertinent sentences or paragraphs related to the user's query.
Think of it like highlighting the most important parts of a long article.
The Base Retriever's Role
Contextual compression doesn't replace your initial retrieval mechanism. Instead, it works on top of it.
- You still need a base retriever (e.g., a vector store retriever) to fetch an initial set of potentially relevant documents.
- The compressor then takes these documents and applies its logic to narrow down or condense their content.
It's a refinement step, not a substitute for finding the initial information.
Introducing LangChain's Tools
LangChain provides excellent tools for implementing contextual compression:
- The
ContextualCompressionRetrieveris the main component. It wraps your base retriever and a document compressor. - A document compressor (e.g., an LLM-based one) is the actual logic that filters or extracts information from the documents.
Let's see how to set up a dummy base retriever first.
Setting Up a Base Retriever
Here's a simple dummy retriever. In a real application, this would typically connect to a vector database like Pinecone or Chroma.
from langchain_core.documents import Document
class SimpleBaseRetriever:
def get_relevant_documents(self, query: str):
if "programming" in query.lower() or "python" in query.lower():
return [
Document(page_content="Python is a versatile programming language. Used in AI & web dev."),
Document(page_content="Java is popular for enterprise apps."),
Document(page_content="Data science uses Python for analysis & ML."),
Document(page_content="History of programming languages dates back centuries.")
]
return [
Document(page_content="The quick brown fox jumps over the lazy dog."),
Document(page_content="Cats enjoy napping in sunny spots, especially in the sun."),
Document(page_content="A computer processes data very quickly and efficiently.")
]
base_retriever = SimpleBaseRetriever()
print("Base retriever initialized.")
# Example usage:
# docs = base_retriever.get_relevant_documents("python")
# for doc in docs: print(doc.page_content)LLMs as Document Compressors
One of the most powerful ways to compress context is by using another LLM! An LLM can intelligently understand the query and the retrieved documents, then extract only the most relevant sentences.
- This goes beyond simple keyword matching.
- It leverages the LLM's understanding of semantics.
LangChain provides specific compressors that use LLMs for this task.
Using LLMChainExtractor (Code)
The LLMChainExtractor uses an LLM to extract relevant sections from documents. Here's how to set it up along with the ContextualCompressionRetriever.
from langchain_core.documents import Document
from langchain.retrievers.document_compressors import LLMChainExtractor
from langchain.retrievers import ContextualCompressionRetriever
from langchain_openai import OpenAI # pip install langchain-openai
# --- Dummy Base Retriever (for standalone runnable) ---
class SimpleBaseRetriever:
def get_relevant_documents(self, query: str):
if "programming" in query.lower() or "python" in query.lower():
return [
Document(page_content="Python is a versatile programming language. Used in AI & web dev."),
Document(page_content="Java is popular for enterprise apps."),
Document(page_content="Data science uses Python for analysis & ML."),
Document(page_content="History of programming languages dates back centuries.")
]
return [
Document(page_content="The quick brown fox jumps over the lazy dog."),
Document(page_content="Cats enjoy napping in sunny spots, especially in the sun."),
Document(page_content="A computer processes data very quickly and efficiently.")
]
base_retriever = SimpleBaseRetriever()
# --- LLM for Compression ---
# NOTE: You need to set your OpenAI API key as an environment variable
# e.g., import os; os.environ["OPENAI_API_KEY"] = "YOUR_API_KEY"
# Ensure 'langchain-openai' is installed (pip install langchain-openai).
llm = OpenAI(temperature=0.1) # Low temp for focused extraction
# --- Create the Extractor and Compression Retriever ---
compressor = LLMChainExtractor.from_llm(llm)
compression_retriever = ContextualCompressionRetriever(
base_compressor=compressor,
base_retriever=base_retriever
)
print("LLMChainExtractor and ContextualCompressionRetriever initialized!")Running Compressed Retrieval (Code)
Now, let's run a query and observe how the ContextualCompressionRetriever (using LLMChainExtractor) processes the documents. Pay attention to the length of the document content!
from langchain_core.documents import Document
from langchain.retrievers.document_compressors import LLMChainExtractor
from langchain.retrievers import ContextualCompressionRetriever
from langchain_openai import OpenAI # pip install langchain-openai
# --- Dummy Base Retriever (re-defined for standalone runnable) ---
class SimpleBaseRetriever:
def get_relevant_documents(self, query: str):
if "programming" in query.lower() or "python" in query.lower():
return [
Document(page_content="Python is a versatile programming language. Used in AI & web dev."),
Document(page_content="Java is popular for enterprise apps."),
Document(page_content="Data science uses Python for analysis & ML."),
Document(page_content="History of programming languages dates back centuries.")
]
return [
Document(page_content="The quick brown fox jumps over the lazy dog."),
Document(page_content="Cats enjoy napping in sunny spots, especially in the sun."),
Document(page_content="A computer processes data very quickly and efficiently.")
]
base_retriever = SimpleBaseRetriever()
# --- LLM for Compression (re-defined for standalone runnable) ---
# NOTE: You need to set your OpenAI API key as an environment variable
# Ensure 'langchain-openai' is installed.
llm = OpenAI(temperature=0.1)
compressor = LLMChainExtractor.from_llm(llm)
compression_retriever = ContextualCompressionRetriever(
base_compressor=compressor,
base_retriever=base_retriever
)
# --- Run a Query ---
query = "What is Python used for?"
print(f"Query: '{query}'\n")
print("--- Original (simulated) documents ---")
original_docs = base_retriever.get_relevant_documents(query)
for i, doc in enumerate(original_docs):
print(f"Doc {i+1} (Chars: {len(doc.page_content)}): {doc.page_content}")
print(f"Total original documents: {len(original_docs)}\n")
print("--- Content after compression ---")
compressed_docs = compression_retriever.get_relevant_documents(query)
for i, doc in enumerate(compressed_docs):
# The LLMChainExtractor modifies the page_content of existing docs
print(f"Compressed Doc {i+1} (Chars: {len(doc.page_content)}): {doc.page_content}")
print(f"Total compressed documents: {len(compressed_docs)}")Other Compression Options
LangChain offers more than just LLMChainExtractor:
LLMChainFilter: Uses an LLM to decide if an entire document is relevant enough to keep, rather than extracting parts.EmbeddingsFilter: Filters documents based on the semantic similarity of their embeddings to the query. This is often faster but less nuanced than LLM-based filtering.DocumentCompressorPipeline: Allows you to chain multiple compressors together for complex logic!
Why Compression Matters
Contextual compression is a powerful technique for optimizing RAG systems:
- Improved Relevance: LLMs receive more focused, pertinent information.
- Reduced Cost: Fewer tokens are sent to the LLM API.
- Faster Responses: LLMs process less text, leading to quicker answers.
- Better Accuracy: Less irrelevant noise reduces the chance of LLM hallucinations.
It helps you get more out of your LLMs while saving resources!
Compression Check
Test your understanding of contextual compression!
Contextual Compression Recap
We've explored how contextual compression refines retrieved documents for LLMs:
- It's a post-retrieval step that filters and extracts key information.
- LangChain's
ContextualCompressionRetrieverwraps a base retriever and a document compressor. LLMChainExtractoruses an LLM to intelligently extract relevant snippets.- Benefits include improved relevance, reduced costs, faster responses, and better accuracy.
This technique is vital for optimizing RAG systems, especially with large knowledge bases.
Preguntas frecuentes
¿La lección «Compresión contextual con LLM» es gratis?
Sí — el texto completo de «Compresión contextual con LLM» es gratis para leer aquí en la web. Para practicarla de forma interactiva (editor de código integrado y tutor de IA 24/7) y desbloquear el resto del curso de LangChain / RAG / Vector DBs, actualiza a CoddyKit PRO. El curso de LangChain / RAG / Vector DBs incluye 4 lecciones en total.
¿Qué aprenderé en «Compresión contextual con LLM»?
Optimice el contexto que se proporciona al LLM filtrando y comprimiendo dinámicamente los documentos recuperados para centrarse en la relevancia. Practicas LangChain / RAG / Vector DBs con código real que ejecutas directamente en el navegador, y un tutor de IA 24/7 responde tus preguntas mientras trabajas en la lección.
¿Necesito experiencia previa para empezar LangChain / RAG / Vector DBs?
No se requiere experiencia previa. LangChain / RAG / Vector DBs en CoddyKit está estructurado para principiantes hasta estudiantes avanzados, así que puedes empezar aquí o desde el inicio y avanzar a tu ritmo. Esta es la lección 2 de 4.
¿Cuánto tiempo toma la lección «Compresión contextual con LLM»?
La mayoría de las lecciones de CoddyKit toman alrededor de 5–10 minutos. Cada una es compacta e interactiva, así que avanzas constantemente y retomas exactamente por donde dejaste en la web y la app.
¿Puedo escribir y ejecutar código en esta lección de LangChain / RAG / Vector DBs?
Sí. Cada lección de LangChain / RAG / Vector DBs incluye un editor de código integrado, así que escribes y ejecutas código real directamente en tu navegador y obtienes retroalimentación instantánea de IA — sin configuración local necesaria.
Todas las lecciones de este curso
- Estrategias de recuperación multiconsulta
- Compresión contextual con LLM
- Búsqueda híbrida y reordenación
- Recuperación por documento principal y ventana de oraciones