Choosing and Configuring an Embedding Model for RAG

How to Choose an Embedding Model for RAG?

We often encounter situations where a RAG system works, but retrieval returns irrelevant documents. Most often, the problem is the embedding model. It may not suit your corpus language or query context. Changing the model can boost recall by 10–15% without changing the architecture. We offer a free audit; paid tuning starts at $3,000. Our engineers have 10 years of experience in NLP. They have implemented over 50 RAG projects for Enterprise. An error in model selection can cost up to 40% of answer accuracy. We guarantee that after tuning to your domain, recall will increase by at least 10%. Contact us for a free audit.

Step-by-Step Model Selection

1. Analyze your corpus: Determine language, size, and domain specificity.
2. Define latency and privacy: Decide if you need on-premise or can use APIs.
3. Select candidate models: Choose 2-3 from the table below.
4. Evaluate on your data: Use RAGAS to measure context_recall and context_precision.
5. Pick the winner: We help you finalize and integrate.

Model Comparison on MTEB

The embedding model is one of the most critical components of a RAG system. Retrieval quality directly depends on how well the model represents texts in vector space. Changing the embedding model can yield a greater recall improvement than optimizing chunking or search parameters.

Candidate Models

Proprietary API models:

• text-embedding-3-large (OpenAI, dim=3072): Best overall on MTEB benchmarks.
• text-embedding-3-small (OpenAI, dim=1536): Good price/quality ratio.
• embed-v3 (Cohere): Strong on retrieval; supports input_type parameter.

Open models (self-hosted):

• BAAI/bge-m3 (dim=1024): Multilingual, supports dense+sparse+colbert.
• BAAI/bge-large-en-v1.5 (dim=1024): Best for English.
• intfloat/multilingual-e5-large (dim=1024): Good on Russian.
• nomic-ai/nomic-embed-text-v1.5 (dim=768): Supports matryoshka.

Performance on BEIR (Retrieval nDCG@10)

Data from MTEB leaderboard:

For Russian-language tasks, the picture differs. We recommend testing on your own domain. We conduct a trial on your corpus and provide a report.

Configuring Popular Embedding Models

Cohere Embed v3 with input_type

Cohere embed-v3 requires specifying input_type for optimal retrieval. Using the correct type increases recall by 8–15%:

import cohere

co = cohere.Client(api_key="...")

def embed_documents(texts: list[str]) -> list[list[float]]:
    response = co.embed(
        texts=texts,
        model="embed-multilingual-v3.0",
        input_type="search_document",
    )
    return response.embeddings

def embed_query(query: str) -> list[float]:
    response = co.embed(
        texts=[query],
        model="embed-multilingual-v3.0",
        input_type="search_query",
    )
    return response.embeddings[0]

Self-hosted BGE-M3

BGE-M3 is the most versatile open-source model. It supports dense, sparse (SPLADE), and ColBERT-style multi-vector retrieval from a single model. Infrastructure savings: one model instead of three.

from FlagEmbedding import BGEM3FlagModel

model = BGEM3FlagModel(
    "BAAI/bge-m3",
    use_fp16=True,
    device="cuda",
)

# Dense embeddings (for standard ANN search)
dense_embeddings = model.encode(
    texts,
    batch_size=32,
    max_length=8192,
    return_dense=True,
    return_sparse=False,
    return_colbert_vecs=False,
)["dense_vecs"]

# Sparse embeddings (for BM25-like search)
sparse_embeddings = model.encode(
    texts,
    return_dense=False,
    return_sparse=True,
)["lexical_weights"]

# Hybrid retrieval score
def compute_bge_m3_score(query_dense, doc_dense, query_sparse, doc_sparse,
                          alpha=0.5) -> float:
    dense_score = np.dot(query_dense, doc_dense)
    sparse_score = sum(
        query_sparse.get(token, 0) * doc_sparse.get(token, 0)
        for token in query_sparse
    )
    return alpha * dense_score + (1 - alpha) * sparse_score

Dimensionality Reduction with Matryoshka

Nomic Embed and other models support matryoshka embeddings. You can use the first N dimensions without retraining. This reduces vector DB RAM requirements by 2× with only 2–5% quality loss. For instance, text-embedding-3-large can output 1536 dimensions instead of 3072:

from openai import OpenAI

client = OpenAI()

response = client.embeddings.create(
    model="text-embedding-3-large",
    input=texts,
    dimensions=1536,
)

Practical Choice and Quality Factors

Key Factors Beyond Model Choice

Besides model selection, important factors include text preprocessing, chunk length, and strategies for merging sparse+dense results. We account for all these when tuning RAG for your domain.

Decision Guidelines

• Confidential/on-prem data: Use BGE-M3 or E5-mistral-7b.
• Best Russian support: Test BGE-M3, multilingual-e5-large, and text-embedding-3-large on your own data. There is no universal winner.
• Minimal latency: Prefer text-embedding-3-small (API) or nomic-embed-text-v1.5 (self-hosted).
• Hybrid sparse+dense: BGE-M3 is the only open model with native dual support.

Comparison: Cohere embed-v3 with input_type outperforms without it by 8–15% in recall. BGE-M3 is 2× faster than OpenAI when self-hosted with GPU.

Evaluation on Your Domain

from ragas import evaluate
from ragas.metrics import context_recall, context_precision

for model_name in ["text-embedding-3-small", "text-embedding-3-large"]:
    retriever = build_retriever(model_name)
    scores = evaluate(test_dataset, metrics=[context_recall, context_precision],
                      retriever=retriever)
    print(f"{model_name}: recall={scores['context_recall']:.3f}, "
          f"precision={scores['context_precision']:.3f}")

Turnkey RAG Setup and Pricing

What's Included

• Analysis of your corpus and use cases.
• Selection and testing of 2–3 embedding models.
• Indexing configuration (chunk size, overlap, vector DB).
• Integration with your backend (API, gRPC).
• Documentation and team training.
• 2-week post-release support.

Timeline and Cost

• Embedding model setup and indexing: 2–5 days.
• Comparative testing of 2–3 models: 3–5 days.
• Total: 1–2 weeks.
• Typical cost: $3,000–$6,000 depending on complexity.
• We will assess your task for free in one business day.

Contact us, and we will select the optimal model for your RAG. We guarantee a recall improvement of at least 10% on your metrics.

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.