Construir un pipeline RAG local con ChromaDB y Sentence

RAG (Retrieval-Augmented Generation) combina recuperacion de documentos con generacion LLM. ChromaDB es una base de datos vectorial local que corre en proceso — sin servicio externo. Combinado con sentence-transformers para embeddings locales, puedes construir un pipeline RAG completo sin API keys ni dependencias cloud. A continuacion: chunking, embedding, recuperacion y generacion.

Cuando Usar Esto

• For alternatives, see Run LLMs Locally with Ollama for Private Inference.

• Q&A de documentos privados donde los datos no pueden salir de tu maquina

• Prototipado de pipelines RAG antes de escalar a bases de datos vectoriales gestionadas

• Entornos offline o air-gapped

Requisitos Previos

• Python 3.10+
• Paquete chromadb (pip install chromadb)
• Paquete sentence-transformers (pip install sentence-transformers)
• Un LLM (API de OpenAI, Ollama o cualquier proveedor LLM)

Solucion

1. Instalar dependencias

pip install chromadb sentence-transformers

2. Chunking de documentos

from dataclasses import dataclass

@dataclass
class Chunk:
    id: str
    text: str
    metadata: dict

def chunk_text(
    text: str,
    chunk_size: int = 500,
    overlap: int = 50,
) -> list[Chunk]:
    """Split text into overlapping chunks.

    Args:
        text: Input text to chunk.
        chunk_size: Target characters per chunk.
        overlap: Number of characters to overlap between chunks.

    Returns:
        List of Chunk objects.
    """
    chunks = []
    start = 0
    chunk_idx = 0

    while start < len(text):
        end = start + chunk_size
        chunk_text_str = text[start:end]

        chunks.append(Chunk(
            id=f"chunk_{chunk_idx}",
            text=chunk_text_str,
            metadata={"chunk_index": chunk_idx, "char_start": start},
        ))

        start = end - overlap
        chunk_idx += 1

    return chunks

def chunk_documents(documents: list[dict]) -> list[Chunk]:
    """Chunk multiple documents."""
    all_chunks = []
    for doc in documents:
        chunks = chunk_text(doc["text"])
        for chunk in chunks:
            chunk.metadata.update({
                "source": doc.get("source", "unknown"),
                "title": doc.get("title", ""),
            })
        all_chunks.extend(chunks)
    return all_chunks

3. Inicializar ChromaDB con embeddings locales

import chromadb

class LocalRAG:
    def __init__(self, persist_dir: str = "./chroma_db"):
        self.client = chromadb.PersistentClient(path=persist_dir)
        self.collection = self.client.get_or_create_collection(
            name="documents",
            metadata={"hnsw:space": "cosine"},
        )

    def add_chunks(self, chunks: list[Chunk]) -> None:
        """Add chunks to the ChromaDB collection."""
        self.collection.add(
            ids=[c.id for c in chunks],
            documents=[c.text for c in chunks],
            metadatas=[c.metadata for c in chunks],
        )

    def query(
        self,
        query_text: str,
        n_results: int = 5,
    ) -> list[dict]:
        """Retrieve relevant chunks for a query.

        Args:
            query_text: User query.
            n_results: Number of chunks to retrieve.

        Returns:
            List of {text, metadata, distance} dicts.
        """
        results = self.collection.query(
            query_texts=[query_text],
            n_results=n_results,
        )

        return [
            {
                "text": doc,
                "metadata": meta,
                "distance": dist,
            }
            for doc, meta, dist in zip(
                results["documents"][0],
                results["metadatas"][0],
                results["distances"][0],
            )
        ]

4. Construir el pipeline RAG

from openai import OpenAI

class RAGPipeline:
    def __init__(self, persist_dir: str = "./chroma_db"):
        self.rag = LocalRAG(persist_dir)
        self.llm = OpenAI()

    def ingest(self, documents: list[dict]) -> None:
        """Ingest documents into the vector store."""
        chunks = chunk_documents(documents)
        self.rag.add_chunks(chunks)
        print(f"Ingested {len(chunks)} chunks from {len(documents)} documents")

    def retrieve(self, query: str, n_results: int = 5) -> list[dict]:
        """Retrieve relevant chunks for a query."""
        return self.rag.query(query, n_results)

    def generate(
        self,
        query: str,
        context_chunks: list[dict],
    ) -> str:
        """Generate an answer using retrieved context.

        Args:
            query: User question.
            context_chunks: Retrieved chunks from retrieve().

        Returns:
            LLM-generated answer.
        """
        context = "\n\n".join(
            f"[Source: {c['metadata'].get('source', 'unknown')}]\n{c['text']}"
            for c in context_chunks
        )

        response = self.llm.chat.completions.create(
            model="gpt-4o-mini",
            messages=[
                {
                    "role": "system",
                    "content": (
                        "Answer the question based on the provided context. "
                        "If the context does not contain the answer, say "
                        "'I don't have enough information to answer this.'"
                    ),
                },
                {
                    "role": "user",
                    "content": f"Context:\n{context}\n\nQuestion: {query}",
                },
            ],
            temperature=0.2,
        )

        return response.choices[0].message.content

    def ask(self, query: str, n_results: int = 5) -> dict:
        """Full RAG pipeline: retrieve + generate.

        Returns:
            Dict with answer, sources, and retrieved chunks.
        """
        chunks = self.retrieve(query, n_results)
        answer = self.generate(query, chunks)

        return {
            "answer": answer,
            "sources": [
                {"source": c["metadata"].get("source"), "text": c["text"][:200]}
                for c in chunks
            ],
            "num_chunks": len(chunks),
        }

5. Usar el pipeline

pipeline = RAGPipeline()

# Ingerir documentos
pipeline.ingest([
    {
        "title": "Redis Guide",
        "source": "redis-docs.md",
        "text": "Redis is an in-memory data structure store. It supports strings, hashes, lists, sets, sorted sets, streams, and more. Redis is commonly used for caching, session management, real-time analytics, and message brokering.",
    },
    {
        "title": "PostgreSQL Guide",
        "source": "postgres-docs.md",
        "text": "PostgreSQL is a capable open-source relational database. It supports ACID compliance, JSON columns, full-text search, and geospatial data via PostGIS.",
    },
])

# Hacer preguntas
result = pipeline.ask("What data structures does Redis support?")
print(result["answer"])
print(f"Sources: {[s['source'] for s in result['sources']]}")

6. Usar con Ollama para RAG totalmente local

import requests

class LocalLLMRAG(RAGPipeline):
    def __init__(self, persist_dir: str = "./chroma_db", ollama_url: str = "http://localhost:11434"):
        self.rag = LocalRAG(persist_dir)
        self.ollama_url = ollama_url

    def generate(self, query: str, context_chunks: list[dict]) -> str:
        context = "\n\n".join(c["text"] for c in context_chunks)

        response = requests.post(
            f"{self.ollama_url}/api/generate",
            json={
                "model": "llama3",
                "prompt": f"Context:\n{context}\n\nQuestion: {query}\n\nAnswer:",
                "stream": False,
            },
        )

        return response.json()["response"]

# Totalmente local — sin API keys
pipeline = LocalLLMRAG()
pipeline.ingest(documents)
result = pipeline.ask("How does Redis handle caching?")

Como Funciona

1. Chunking divide documentos en segmentos de ~500 caracteres con 50 caracteres de overlap. El overlap asegura continuidad de contexto entre chunks para que informacion relevante no se divida entre limites.
2. ChromaDB usa sentence-transformers (all-MiniLM-L6-v2 por defecto) para embedder chunks y queries. Los embeddings se generan localmente en proceso — sin llamadas a API.
3. Recuperacion embedde la query y encuentra los chunks mas cercanos por similitud coseno. ChromaDB usa HNSW (Hierarchical Navigable Small World) para busqueda rapida de nearest neighbor aproximado.
4. Generacion concatena los chunks recuperados como contexto y los envia al LLM con la pregunta del usuario. El LLM genera una respuesta basada en el contexto proporcionado.
5. PersistentClient almacena la base de datos vectorial en disco, para que los embeddings persistan entre reinicios sin re-ingerir.

Variantes

Filtrado por metadatos

def query_with_filter(
    self,
    query_text: str,
    n_results: int = 5,
    where: dict | None = None,
) -> list[dict]:
    results = self.collection.query(
        query_texts=[query_text],
        n_results=n_results,
        where=where,
    )
    return results

# Buscar solo dentro de un source especifico
results = pipeline.rag.query_with_filter(
    "caching strategies",
    where={"source": "redis-docs.md"},
)

Modelo de embedding personalizado

from sentence_transformers import SentenceTransformer

class CustomEmbeddingRAG(LocalRAG):
    def __init__(self, model_name: str = "BAAI/bge-large-en-v1.5"):
        self.embedder = SentenceTransformer(model_name)
        self.client = chromadb.PersistentClient(path="./chroma_db")
        self.collection = self.client.get_or_create_collection(
            name="documents",
            embedding_function=self._embed_fn,
        )

    def _embed_fn(self, input: list[str]) -> list[list[float]]:
        return self.embedder.encode(input).tolist()

Reranking de chunks recuperados

from sentence_transformers import CrossEncoder

class RerankedRAG(RAGPipeline):
    def __init__(self):
        super().__init__()
        self.reranker = CrossEncoder("cross-encoder/ms-marco-MiniLM-L-6-v2")

    def retrieve(self, query: str, n_results: int = 20, top_k: int = 5) -> list[dict]:
        # Recuperar mas chunks de los necesarios
        chunks = self.rag.query(query, n_results)

        # Rerankear con cross-encoder
        pairs = [(query, c["text"]) for c in chunks]
        scores = self.reranker.predict(pairs)

        # Ordenar por score del reranker y retornar top_k
        ranked = sorted(zip(chunks, scores), key=lambda x: x[1], reverse=True)
        return [c for c, _ in ranked[:top_k]]

Mejores Practicas

• Usa overlap en chunking — 10-20% de overlap previene perder contexto en los limites de chunks
• Incluye metadatos de source — permite citar fuentes en la respuesta generada
• Usa temperatura baja para generacion — 0.1-0.3 reduce alucinacion en Q&A factual
• Rerankear para mejor precision — recupera 3-5x mas chunks y rerankear con cross-encoder para top-k

Errores Comunes

• Chunks demasiado grandes — el LLM puede perder detalles en contexto largo; mantén chunks de 200-500 palabras
• Sin overlap — informacion relevante dividida en limites de chunks se pierde
• No persistir ChromaDB — EphemeralClient pierde datos al reiniciar; usa PersistentClient
• Recuperar muy pocos chunks — 3-5 es un buen default; muy pocos pierden contexto relevante