Building RAG with Milvus Vector Database

Building RAG with Milvus Vector Database

When Millions of Chunks No Longer Fit in RAM

Relational databases with cosine distance fail at volumes >100K vectors. We encountered this in a fintech project: 2.5M documentation chunks, 6 languages, requirement P99 latency <500 ms. We chose Milvus — an open-source vector DB with HNSW indexes and GPU acceleration. We'll share how we tuned hybrid search (dense + sparse) in production.

A typical problem is "phantom" relevant results: the model returns semantically similar but contextually irrelevant items. We solved this with hybrid search using RRF reranking. Our experience shows that proper index configuration reduces latency by 40%.

Problems We Solve

Scale: 2.5M chunks is not the limit. Milvus scales horizontally easily: we deployed clusters up to 10 nodes supporting billions of vectors. Choosing the right index is key: for our client we used HNSW (M=32, efConstruction=400) for dense and SPARSE_INVERTED_INDEX for sparse.

Hybrid search: dense vs sparse. Dense vectors capture semantics well but miss exact keyword matches. We combined both approaches with RRF reranking. Relevance increased by 25%.

Multi-tenancy without headaches. Partitioning isolates client data without creating separate collections. We implemented a dynamic partition scheme — easy and secure.

Why Milvus Is the Best Choice for Enterprise RAG?

Milvus outperforms other vector DBs in speed and cost under high load. Compare indexes:

We use HNSW as a universal choice for production, and DISKANN for archival data on SSD.

Why Milvus Is Better Than Pinecone for High Loads?

Milvus is cheaper at volumes >1M vectors since it doesn't charge per query. Compare:

For a typical 10M vector cluster, Milvus is 3-5 times cheaper than Pinecone with the same performance.

How to Configure HNSW Indexes for Minimal Latency?

Key parameters: M (16-64) and efConstruction (200-500). Higher efConstruction means better accuracy but slower building. In production we use M=32, efConstruction=400, and for search ef=100-200. This gives P99 latency <400 ms on 2.5M chunks.

How We Configured Hybrid Search for High Relevance?

Here's the code we applied for our client (fintech, 2.5M chunks):

from pymilvus import connections, Collection, FieldSchema, CollectionSchema, DataType, utility

# Connect to Milvus
connections.connect(
    alias="default",
    host="localhost",
    port="19530"
)

# Or via URI (Milvus Lite for local development)
from pymilvus import MilvusClient
client = MilvusClient("./milvus_local.db")  # SQLite-like file

fields = [
    FieldSchema(name="id", dtype=DataType.INT64, is_primary=True, auto_id=True),
    FieldSchema(name="text", dtype=DataType.VARCHAR, max_length=4096),
    FieldSchema(name="source", dtype=DataType.VARCHAR, max_length=512),
    FieldSchema(name="doc_type", dtype=DataType.VARCHAR, max_length=64),
    FieldSchema(name="page", dtype=DataType.INT32),
    FieldSchema(
        name="dense_vector",
        dtype=DataType.FLOAT_VECTOR,
        dim=1536  # text-embedding-3-small
    ),
    FieldSchema(
        name="sparse_vector",
        dtype=DataType.SPARSE_FLOAT_VECTOR  # BM25
    ),
]

schema = CollectionSchema(fields=fields, description="Corporate Knowledge Base")
collection = Collection(name="knowledge_base", schema=schema)

# Indexes for vector fields
collection.create_index(
    field_name="dense_vector",
    index_params={"metric_type": "COSINE", "index_type": "HNSW", "params": {"M": 16, "efConstruction": 200}}
)
collection.create_index(
    field_name="sparse_vector",
    index_params={"metric_type": "IP", "index_type": "SPARSE_INVERTED_INDEX"}
)

collection.load()

from pymilvus import AnnSearchRequest, RRFRanker

def milvus_hybrid_search(query: str, top_k: int = 5) -> list:
    # Dense vector
    dense_vec = dense_embedder.embed_query(query)

    # Sparse vector (via built-in BM25Encoder)
    sparse_vec = sparse_encoder.encode_queries([query])

    # Two requests for RRF
    dense_req = AnnSearchRequest(
        data=[dense_vec],
        anns_field="dense_vector",
        param={"metric_type": "COSINE", "params": {"ef": 100}},
        limit=30,
    )

    sparse_req = AnnSearchRequest(
        data=sparse_vec,
        anns_field="sparse_vector",
        param={"metric_type": "IP"},
        limit=30,
    )

    # RRF fusion
    results = collection.hybrid_search(
        reqs=[dense_req, sparse_req],
        rerank=RRFRanker(k=60),
        limit=top_k,
        output_fields=["text", "source", "doc_type"],
    )

    return results

Result: 850 QPS with P99 latency <400ms on a 3-node cluster (8 vCPU, 32GB RAM each).

What's Included in the Work

• Data audit and collection schema design
• Milvus cluster deployment (Kubernetes / bare-metal) with GPU acceleration
• Index and parameter tuning (efConstruction, M, ef) for load
• Ingestion pipeline (PySpark / Kafka / direct)
• RAG pipeline with LangChain or LlamaIndex
• Documentation and team training
• 2 weeks of post-launch support

Process

1. Analysis: gather requirements, assess volume and query frequency.
2. Design: select schema, indexes, cluster topology.
3. Implementation: deploy Milvus, write pipelines, integrate with LLM.
4. Testing: load testing, parameter tweaks.
5. Deployment: CI/CD, monitoring (Prometheus + Grafana).

Timeline

• Milvus cluster setup + schema: 3–5 days
• Ingestion pipeline with hybrid indexing: 5–10 days
• RAG pipeline and evaluation: 1–2 weeks
• Total: 3–5 weeks

Company Metrics

Over 7 years working with vector databases, 30+ Milvus production deployments. We guarantee 99.9% uptime on the cluster and expert support. We'll evaluate your project in 2 days — contact us. Get a consultation on RAG architecture: we'll estimate cost and timeline for your data volume.

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.