Skip to main content
🤖 Running agents for a team, not just yourself?Get an independent review of identity, secrets, failover, observability and governance.Assess your agent platform
AI

Enterprise RAG Architecture Patterns at Scale

Production-grade RAG architectures for enterprise. Vector database selection, chunking strategies, hybrid search, and scaling patterns on Kubernetes.

Luca Berton

· The Production AI Expert

Published ·2 min read

RAG Beyond the Demo

Every AI demo uses RAG. Few do it well in production. The gap between a prototype that queries a vector database and an enterprise system that serves thousands of users with accurate, sourced answers is enormous.

Here’s what I’ve learned deploying RAG systems at scale.

Architecture Overview

Documents → Ingestion Pipeline → Vector DB + Metadata Store
                                        ↓
User Query → Query Planner → Hybrid Search → Reranker → LLM → Response
                                                              ↓
                                                        Citation Check

Chunking Strategies That Actually Work

The most impactful decision in RAG isn’t your model or vector database — it’s how you chunk documents.

Semantic Chunking

Instead of fixed-size chunks, split on semantic boundaries:

from langchain.text_splitter import RecursiveCharacterTextSplitter

# Bad: Fixed-size chunks break context
bad_splitter = RecursiveCharacterTextSplitter(chunk_size=500)

# Better: Respect document structure
good_splitter = RecursiveCharacterTextSplitter(
    chunk_size=1000,
    chunk_overlap=200,
    separators=["\n## ", "\n### ", "\n\n", "\n", ". ", " "],
)

Hierarchical Chunking

Store chunks at multiple granularities:

class HierarchicalChunker:
    def chunk(self, document):
        # Level 1: Full sections (for context)
        sections = self.split_by_headings(document)
        
        # Level 2: Paragraphs (for retrieval)
        paragraphs = []
        for section in sections:
            for para in self.split_paragraphs(section):
                para.metadata["parent_section"] = section.id
                paragraphs.append(para)
        
        return sections, paragraphs

When a paragraph matches, retrieve its parent section for full context. This dramatically improves answer quality.

Metadata Enrichment

Every chunk needs rich metadata:

chunk_metadata = {
    "source": "architecture-guide-v3.pdf",
    "page": 42,
    "section": "Security Requirements",
    "author": "Platform Team",
    "last_updated": "2026-01-15",
    "document_type": "technical_spec",
    "access_level": "internal",
    "chunk_index": 7,
    "total_chunks": 23,
}

Vector Database Selection

DatabaseBest ForKubernetes-NativeHybrid Search
MilvusLarge scale (100M+ vectors)✅ (Helm chart)✅
QdrantMid-scale, filtering✅ (Operator)✅
WeaviateMulti-modal, GraphQL✅ (Helm)✅
pgvectorSmall scale, existing PostgresVia operatorLimited
ChromaDBPrototypes only❌❌

For enterprise Kubernetes deployments, I recommend Milvus for scale or Qdrant for simplicity.

Hybrid Search: The Secret Weapon

Pure vector search misses exact matches. Pure keyword search misses semantics. Combine both:

async def hybrid_search(query: str, top_k: int = 20) -> list[Chunk]:
    # Semantic search
    embedding = await embed(query)
    vector_results = await vector_db.search(embedding, top_k=top_k)
    
    # Keyword search (BM25)
    keyword_results = await search_engine.search(query, top_k=top_k)
    
    # Reciprocal Rank Fusion
    combined = reciprocal_rank_fusion(
        [vector_results, keyword_results],
        k=60,
    )
    
    return combined[:top_k]

def reciprocal_rank_fusion(result_lists, k=60):
    scores = {}
    for results in result_lists:
        for rank, doc in enumerate(results):
            if doc.id not in scores:
                scores[doc.id] = 0
            scores[doc.id] += 1 / (k + rank + 1)
    
    return sorted(scores.items(), key=lambda x: x[1], reverse=True)

Reranking

Initial retrieval casts a wide net. A reranker picks the best results:

from sentence_transformers import CrossEncoder

reranker = CrossEncoder("cross-encoder/ms-marco-MiniLM-L-12-v2")

def rerank(query: str, chunks: list[Chunk], top_k: int = 5) -> list[Chunk]:
    pairs = [(query, chunk.text) for chunk in chunks]
    scores = reranker.predict(pairs)
    
    ranked = sorted(zip(chunks, scores), key=lambda x: x[1], reverse=True)
    return [chunk for chunk, score in ranked[:top_k]]

Scaling on Kubernetes

apiVersion: apps/v1
kind: Deployment
metadata:
  name: rag-api
spec:
  replicas: 3
  template:
    spec:
      containers:
      - name: rag-service
        resources:
          requests:
            memory: "2Gi"
            cpu: "1"
          limits:
            memory: "4Gi"
        env:
        - name: VECTOR_DB_HOST
          value: "milvus.vector-db.svc.cluster.local"
        - name: RERANKER_MODEL
          value: "cross-encoder/ms-marco-MiniLM-L-12-v2"
        - name: CACHE_TTL_SECONDS
          value: "3600"
---
# HPA for automatic scaling
apiVersion: autoscaling/v2
kind: HorizontalPodAutoscaler
metadata:
  name: rag-api-hpa
spec:
  scaleTargetRef:
    apiVersion: apps/v1
    kind: Deployment
    name: rag-api
  minReplicas: 2
  maxReplicas: 10
  metrics:
  - type: Resource
    resource:
      name: cpu
      target:
        type: Utilization
        averageUtilization: 70

Evaluation: How to Know Your RAG Works

You can’t improve what you don’t measure:

class RAGEvaluator:
    async def evaluate(self, test_set: list[QAPair]):
        results = []
        for qa in test_set:
            response = await rag_pipeline.query(qa.question)
            
            results.append({
                "faithfulness": self.check_faithfulness(response, qa.context),
                "relevance": self.check_relevance(response, qa.question),
                "correctness": self.check_correctness(response, qa.answer),
                "has_citations": bool(response.citations),
                "latency_ms": response.latency_ms,
            })
        
        return aggregate_metrics(results)

Track these metrics weekly. RAG quality degrades as your document corpus changes.

Key Takeaways

  1. Chunking is everything — invest time in semantic, hierarchical chunking with rich metadata
  2. Hybrid search beats vector-only — combine BM25 and embeddings with reciprocal rank fusion
  3. Always rerank — a lightweight cross-encoder dramatically improves precision
  4. Evaluate continuously — build a test set and track faithfulness, relevance, and correctness
  5. Cache aggressively — similar queries get similar results; cache at the embedding and response level

Building an enterprise RAG system? I help organizations design production-grade retrieval architectures. Get in touch.

Free 30-min Production AI consultation

Book Now