Production RAG Pipeline Design, Evaluation, and Optimization

Note: When a RAG pipeline returns irrelevant results, the problem is often in the retrieval architecture. A basic RAG 'works' in a day, but a production-ready system with reliable retrieval, monitoring, and controlled costs requires careful design. We have designed and deployed over 30 RAG systems for various industries — from legal documents to technical support. Our experience shows that 70% of success lies in the pipeline architecture. For example, a FinTech client used dense-only search and achieved a recall of 0.65. After implementing hybrid search with a reranker, recall increased to 0.92, with p99 latency remaining under 200 ms — this significantly reduced the cost of re-queries to the LLM (by 40%, saving $5,000 per month). Retrieval-Augmented Generation is not just a buzzword but a mature technique that requires sound engineering implementation.

Components of a Modern RAG Pipeline

┌─────────────────────────────────────────────────────┐
│                 INGESTION PIPELINE                   │
│  Sources → Loaders → Parsers → Chunkers → Embedder  │
│           → Metadata Extractor → Vector Store        │
└─────────────────────────────────────────────────────┘

┌─────────────────────────────────────────────────────┐
│                 RETRIEVAL PIPELINE                   │
│  Query → Query Transformer → Multi-Index Search     │
│        → Reranker → Context Assembler               │
└─────────────────────────────────────────────────────┘

┌─────────────────────────────────────────────────────┐
│                GENERATION PIPELINE                   │
│  Context + Query → Prompt Builder → LLM             │
│                  → Response Validator → User         │
└─────────────────────────────────────────────────────┘

How to Choose a Chunking Strategy?

Chunking determines retrieval quality. Fixed size 256-512 tokens is simple but loses context. Structural chunking by section headings yields better results for documents with clear hierarchy. Comparison:

We use a semantic splitter based on sentence boundaries with 10% overlap. This gives a balance of quality and speed. Metadata (source, date, type) is attached to each chunk. Semantic chunking is 1.5 times better than fixed-size for recall on narrative texts.

Why Hybrid Search (Sparse + Dense) Is Necessary?

Dense embeddings are great at finding semantic duplicates but struggle with rare terms or abbreviations. Sparse (BM25) does the opposite. According to Qdrant documentation, hybrid search with Reciprocal Rank Fusion (RRF) improves relevance metrics by 20-30%. Hybrid search combines both approaches and is 1.25 times better than dense-only for recall.

from qdrant_client import QdrantClient
from qdrant_client.models import SparseVector

def hybrid_search(query, top_k=10):
    dense_vector = embedder.embed_query(query)
    sparse_vector = sparse_encoder.encode(query)
    results = client.query_points(
        collection_name="docs",
        prefetch=[
            {"query": dense_vector, "using": "dense", "limit": 30},
            {"query": SparseVector(indices=sparse_vector.indices, values=sparse_vector.values),
             "using": "sparse", "limit": 30},
        ],
        query=rrf_fusion,  # Reciprocal Rank Fusion
        limit=top_k,
    )
    return results

In our projects, hybrid search increases Recall@10 by 25-30% compared to dense-only — a factor of 1.25-1.3x. Additionally, we use a reranker based on a cross-encoder (e.g., ms-marco-MiniLM-L-12-v2), which reorders the top-30 results to top-5. This improves precision by 1.5-2x. Rerankers are 2 times better than no reranker for top-5 relevance.

Evaluating Retrieval Quality

For production, an evaluation framework with a set of metrics is essential: Recall@k, MRR, NDCG. We use a custom dataset of 500+ queries with relevance annotations. Load testing checks p99 latency — typically target < 500 ms. Production monitoring detects data drift and metric degradation. We guarantee our pipelines meet MLOps best practices.

Comparison of Embedding Models

For production, we recommend open-source models (E5, BGE) with self-hosting — this reduces operational costs several times with comparable quality. An open-source model is 3 times cheaper than OpenAI per million tokens — saving $0.07 per million tokens. At scale (1M queries/day, 5M tokens), this saves $10,500 per month.

Process

1. Analytics — studying sources, query types, expected load (RPS).
2. Design — stack selection (vector DB, embedding model, chunking).
3. Implementation — building ingestion pipeline, retrieval pipeline, integration with LLM.
4. Testing — evaluation on custom dataset, A/B testing of different configurations.
5. Deployment and monitoring — deployment via Docker/Kubernetes, alerting on latency and recall.

What's Included

• Architecture documentation (diagrams, component specifications)
• Stack selection and justification (vector DB, embedding, LLM)
• Pipeline prototype with hybrid search and reranker
• Evaluation framework with metric suite
• Integration with existing infrastructure
• Customer team training and code review
• Support during deployment phase (up to 1 month)

Timelines (Approximate)

• Design: 1 week
• Ingestion pipeline: 1–2 weeks
• Retrieval pipeline with reranker: 2–3 weeks
• Evaluation and optimization: 1–2 weeks
• Production hardening: 1–2 weeks
• Total: 6–10 weeks depending on complexity

The cost is calculated individually — contact us for a consultation on RAG architecture for your project. We guarantee the result will comply with MLOps best practices and your SLA. Hybrid search reduces LLM inference costs by 30-50% (saving $5,000-$10,000 per month at 1M queries/day), which is especially noticeable under high loads. Order the design of a RAG pipeline and get a working prototype in as little as 3 weeks. With 5+ years of experience and 30+ deployments, we ensure your pipeline is production-ready.

LLM Development: Fine-Tuning, RAG, Agents, and Production Deployment

Using GPT‑4 or Claude 3.5 Sonnet through a public API is not a solution — it's just a tool. When the requirement is to "make it like ChatGPT, but on our data," there is a real engineering challenge behind it: from prompt engineering to training a 70B model on your own infrastructure. End-to-end LLM solution development is a complex stack, and we have been doing it for over 5 years. During this time, we have completed over 20 projects in generative AI: from RAG systems for legal departments to custom support agents. Where exactly your task falls depends on data, latency requirements, budget, and how critical confidentiality is.

A typical situation: the client has already tried ChatGPT, but results are unstable — sometimes accurate, sometimes hallucinating. Or they need integration into a corporate portal while complying with security policies. Let's break down each layer of the stack in detail — from RAG to production deployment.

Why Do RAG Systems Break and How to Fix It?

RAG (Retrieval-Augmented Generation) looks simple: find relevant documents, put them in context, get an answer. In practice, it fails in several places.

Chunking without overlap. Classic mistake: chunk_size=512, overlap=0. If the answer lies across two chunks, retrieval won't find either with sufficient confidence. Solution: overlap 15–25% of chunk_size, or better yet, sentence-aware splitting with spaCy or NLTK instead of naive character splitting.

Poor embedder. text-embedding-ada-002 is good for general use, but on legal or medical texts, specialized models like E5-large-v2, BGE-M3, or fine-tuned sentence-transformers on domain data outperform it. Recall@5 differences can be 15–25%.

No re-ranking. Vector search optimizes for speed, not relevance. A cross-encoder re-ranker (ms-marco-MiniLM-L-6-v2, bge-reranker-large) after initial retrieval improves top-3 accuracy with acceptable latency (+50–150ms). This is often more impactful than improving the embedding model.

Hybrid search. Dense vectors alone work poorly on exact queries: names, SKUs, codes. BM25 (sparse) finds exact matches but misses semantics. Hybrid via RRF (Reciprocal Rank Fusion) is the optimal compromise. Qdrant, Weaviate, and pgvector 0.7+ support hybrid search natively.

When to Fine-Tune Instead of Prompt Engineering?

Prompt engineering solves ~70% of LLM adaptation tasks for a domain. The remaining 30% require fine-tuning. Three indicators: the model ignores a specific output format even with detailed prompting; the task requires deep knowledge of specialized vocabulary (medicine, law); you need to significantly reduce token costs by replacing a large model with a smaller specialized one.

LoRA and QLoRA are the standard for SFT. LoRA adds trainable low-rank matrices to attention layers. A typical configuration for Llama-3 8B: r=64, lora_alpha=128, target_modules=["q_proj","v_proj","k_proj","o_proj"] yields ~0.8% trainable parameters, training on one A100 40GB. QLoRA adds 4-bit quantization (NF4) and allows fine-tuning 70B models on two A100 40GB, though speed drops by half compared to bf16.

DPO instead of RLHF. Direct Preference Optimization requires only (chosen, rejected) pairs, not scalar reward signals. DPOTrainer from the trl library (Hugging Face) implements it in a few dozen lines.

Common mistake. A dataset of 500 examples, 5 epochs, validation loss 0.8 — seems fine. But on test, the model degrades on general instructions. Cause: catastrophic forgetting. Solution: add 10–20% general instruction-following examples (Alpaca, FLAN) to the training set to preserve original capabilities.

How to Choose a Base Model: 8B or 70B?

For most tasks, fine-tuning an 8B model is sufficient. 70B is needed when deep reasoning is required or the 8B baseline does not reach the required quality even after fine-tuning. Inference cost for Llama-3 8B via vLLM on A100 is efficient; the exact cost depends on volume.

What Does PagedAttention Bring to Production?

vLLM is the first choice for serving open-source models. PagedAttention is the key technical innovation: KV-cache is managed like virtual memory in an OS, without fragmentation. This yields 2–4x higher throughput compared to naive HuggingFace Transformers inference. The vLLM documentation confirms that continuous batching and PagedAttention are the standard for high-load LLM services.

Typical numbers on A100 80GB for Llama-3 8B (bf16): 400–600 req/s, P50 latency 200–400ms, P99 latency 600–900ms at concurrency 64. For 70B on two A100 with tensor parallelism: 80–120 req/s, P99 latency 1.5–2.5s. AWQ or GPTQ quantization reduces memory consumption by 2x with quality loss within 1–3%.

Multi-Agent Systems

Agents are LLMs with access to tools: search, code execution, API calls, database interaction. Common patterns:

• ReAct (Reason + Act): the model reasons → chooses a tool → observes the result → reasons again. LangChain and LlamaIndex implement it out of the box.
• Multi-agent orchestration: multiple specialized agents with a coordinator on top. Example: coordinator → researcher (search + summarization) → coder (code generation and execution) → critic (verification). Tools: AutoGen (Microsoft), CrewAI, custom implementation on LangGraph.

In production, agent systems are non-deterministic. Essential: guardrails, step limits, logging of each step, human-in-the-loop for critical actions.

How We Work: Stages, Timeline, Deliverables

What Is Included

We deliver:

• Technical documentation (model card, configs, deployment instructions)
• Access to infrastructure (code repository, trained weights)
• 1 month of post-deployment support (consultations, bug fixes)
• Customer team training (2–3 sessions on system operation)

Timeline: basic RAG prototype — 1–2 weeks. Fine-tuning with customer data — 3–6 weeks (including data preparation). Production system with monitoring and retraining — 2–4 months. Cost is calculated individually based on data volume, model complexity, and infrastructure requirements.

We guarantee the quality of the final model with performance benchmarks and ongoing monitoring. Our engineers have hands‑on experience with dozens of production LLM systems.

Want to evaluate your project? Leave a request — we will prepare a preliminary summary within 1–2 business days. Or get a consultation on choosing the approach: RAG, fine-tuning, or hybrid — we will tell you what works best for you. Contact us to discuss your LLM development needs. Schedule a free consultation today.