RAG with FAISS: Local Vector Search Without Cloud

Vector Search Without the Cloud: FAISS as a Solution for RAG

Cloud vector databases are expensive, slow, and insecure for internal documents. Our team regularly encounters requests for local RAG where every millisecond of latency matters. FAISS by Meta is not a database but a high-performance vector search engine that operates in memory or on disk without network interaction. We use it for embedding into applications, offline scenarios, and situations where an external service is unacceptable. Five years of experience in NLP and 15+ implemented RAG projects allow us to guarantee pipeline stability. According to tests, local FAISS on GPU saves up to 70% compared to cloud services like Pinecone or Weaviate. Switching to local FAISS saves significant amounts monthly for 10 million vectors. Get a consultation on implementing FAISS in your project.

Problems Solved by FAISS

High embedding cost and latency. With batch requests to the OpenAI API, delays grow linearly. FAISS on a local GPU gives p99 latency <5 ms for 100K vectors — 50 times faster than cloud solutions. Data confidentiality. Financial reports, medical records, trade secrets — we do not send embeddings to external services. FAISS stores everything locally. Offline mode. Field devices, closed loops — FAISS works without internet. Order an audit of your current pipeline to identify bottlenecks.

Why Is FAISS Faster Than Traditional Databases?

FAISS vector search does not use SQL, B-trees, or metadata filtering — instead, it applies approximate search algorithms: IVF (Inverted File), HNSW (Hierarchical Navigable Small World), and Product Quantization. Compare:

How We Build RAG with FAISS: Stack and Case Study

Typical stack: Python 3.10+, FAISS 1.7+, OpenAI text-embedding-3-small (1536 dim), GPT-4o-mini for generation. For one client with a corpus of 50,000 technical articles, we chose IndexHNSWFlat (M=16, efConstruction=200). This gave 98% recall and 2 ms latency per query.

import faiss
import numpy as np
import pickle
from openai import OpenAI

openai_client = OpenAI()

def build_faiss_index(texts: list[str], dimension: int = 1536) -> tuple:
    """Creates a FAISS index and corresponding list of texts"""

    # Получаем embeddings батчами
    embeddings = []
    batch_size = 100
    for i in range(0, len(texts), batch_size):
        batch = texts[i:i + batch_size]
        response = openai_client.embeddings.create(
            model="text-embedding-3-small",
            input=batch,
        )
        batch_embeddings = [e.embedding for e in response.data]
        embeddings.extend(batch_embeddings)

    # Конвертируем в numpy float32
    vectors = np.array(embeddings, dtype=np.float32)

    # Нормализуем для cosine similarity (через inner product)
    faiss.normalize_L2(vectors)

    # Создаём HNSW индекс
    index = faiss.IndexHNSWFlat(dimension, 16)  # M=16
    index.hnsw.efConstruction = 200
    index.add(vectors)

    return index, texts

# Сохранение на диск
def save_index(index, texts, path_prefix: str):
    faiss.write_index(index, f"{path_prefix}.index")
    with open(f"{path_prefix}_texts.pkl", "wb") as f:
        pickle.dump(texts, f)

# Загрузка
def load_index(path_prefix: str) -> tuple:
    index = faiss.read_index(f"{path_prefix}.index")
    with open(f"{path_prefix}_texts.pkl", "rb") as f:
        texts = pickle.load(f)
    return index, texts

Search and RAG answer:

def faiss_rag_answer(
    question: str,
    index: faiss.Index,
    texts: list[str],
    top_k: int = 5
) -> str:
    # Embedding вопроса
    query_embedding = openai_client.embeddings.create(
        model="text-embedding-3-small",
        input=question,
    ).data[0].embedding

    query_vector = np.array([query_embedding], dtype=np.float32)
    faiss.normalize_L2(query_vector)

    # Поиск
    distances, indices = index.search(query_vector, top_k)

    # Извлечение текстов
    context_texts = [texts[i] for i in indices[0] if i >= 0]
    context = "\n\n---\n\n".join(context_texts)

    # Генерация ответа
    response = openai_client.chat.completions.create(
        model="gpt-4o-mini",
        messages=[
            {"role": "system", "content": "Отвечай строго на основе предоставленного контекста."},
            {"role": "user", "content": f"Контекст:\n{context}\n\nВопрос: {question}"}
        ],
        temperature=0,
    )
    return response.choices[0].message.content

How to Accelerate FAISS on GPU?

Transferring the index to GPU gives up to 100× performance boost. We use faiss.StandardGpuResources:

res = faiss.StandardGpuResources()
gpu_index = faiss.index_cpu_to_gpu(res, 0, index)  # GPU 0
distances, indices = gpu_index.search(query_vectors, top_k)

RAG Pipeline Development Process

1. Analysis: Examine the corpus, requirements for latency, accuracy (recall), update volume. Choose the index type.
2. Design: Define the pipeline — embedder (OpenAI / local), index, LLM (GPT-4o-mini / Claude). Work out the context format.
3. Implementation: Deploy indexing and search code, integrate with existing application (API / embedded).
4. Testing: Measure latency, recall, answer quality. Optimize parameters (M, efSearch, batch size).
5. Deployment: Containerization (Docker), CI/CD, monitoring metrics (p99 latency, QPS, GPU utilization).

What's Included in the Work?

• Pipeline architecture (documentation + diagram)
• Indexing and search code (Python, versioned Git)
• Index configuration tailored to your corpus (recommendations for HNSW/IVF)
• Integration with LLM (OpenAI, local models via vLLM)
• Load testing (report with metrics)
• Post-deployment support (2 weeks warranty maintenance)

Typical Mistakes When Implementing FAISS

1. Normalization not performed. Without L2 normalization, inner product is not equivalent to cosine similarity — results degrade by 10–20%.
2. Choosing the wrong index. IndexFlatL2 on 10M vectors consumes >60 GB RAM — use IVFPQ.
3. efSearch not tuned. For HNSW, efSearch < 100 reduces recall to 80% — raise to 200–400.
4. Ignoring GPU memory. For large indexes, the index does not fit on GPU — switch to IVF with quantization.

Comparison of FAISS Index Types

Timeline and How to Order

Estimated timeline: for a corpus up to 500K vectors — 1–2 weeks turnkey. For larger volumes (>10M) — 3–4 weeks with index optimization. Contact us for an assessment of your scenario — we will select a configuration and calculate the cost individually. Get a consultation on implementing FAISS in your project. Order an audit of your current pipeline to identify bottlenecks.

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.