Fine-Tuning Embedding Models for Your Domain

Fine-Tuning Embedding Models for Your Domain

Imagine you've deployed a RAG pipeline with a state-of-the-art embedding model—BAAI/bge-m3 or OpenAI text-embedding-3. It works well for general queries, but when it comes to domain-specific documents—medical protocols, court rulings, or technical standards—search accuracy drops. The model confuses domain terms, context recall stagnates at 0.6-0.7, and top-K results are filled with semantically similar but thematically irrelevant documents. We solve this by fine-tuning the model on your data. No infrastructure changes required. Our approach: synthetic pair generation via LLM, fine-tuning with MultipleNegativesRankingLoss, and thorough evaluation. The result: NDCG improvement of 15-30%. Over 5 years, we've completed more than 30 projects customizing NLP models.

"Domain-specific fine-tuning of embedding models can improve retrieval metrics by 15-30% without changing infrastructure" — from a project report

When Is Fine-Tuning Necessary?

Symptoms that indicate the need for fine-tuning:

• The general model confuses specific terms: MeSH terms in medicine, legal constructs in law, technical abbreviations.
• Context recall of the RAG system is stuck below 0.75, even after optimizing chunking and search index.
• High false positive rate—semantically similar but topically irrelevant documents appear in top-K.

If you notice these signs, fine-tuning will boost metrics by 15-30% without changing architecture.

Why Fine-Tuning Is Better Than Model Replacement

Switching to a larger model (e.g., from 768 to 1536 dimensions) increases latency and vector storage costs. Fine-tuning the same model on domain data is cheaper and faster. We use MultipleNegativesRankingLoss—it's more effective than triplet loss for retrieval tasks. Infrastructure savings can reach 30%, and API costs for data generation are offset by accuracy gains.

How We Fine-Tune: Stack and Configuration

We use sentence-transformers, PyTorch, Hugging Face Transformers. Base models: BAAI/bge-m3 or intfloat/multilingual-e5-large. We train on A100 (80GB) with batch size 32, learning rate 2e-5, warmup 10%.

Generating Training Pairs with LLM

To create a dataset without manual labeling, we use GPT-4o-mini. Example generation:

from openai import OpenAI
import json

client = OpenAI()

def generate_queries_for_document(doc_text: str, n: int = 5) -> list[str]:
    response = client.chat.completions.create(
        model="gpt-4o-mini",
        messages=[{
            "role": "user",
            "content": f"""Generate {n} search queries..."""
        }],
        response_format={"type": "json_object"},
    )
    return json.loads(response.choices[0].message.content)["queries"]

Typically: 1000 documents × 5 queries = 5000 pairs in ~2 hours and $5-15 on API.

Step-by-Step Fine-Tuning Process

1. Domain analysis and collection of representative documents. Identify key topics and query types.
2. Generation of synthetic query-document pairs. Use LLM (GPT-4o-mini) to create up to 5000 pairs.
3. Manual review and correction (optional). Engage domain experts to improve quality.
4. Fine-tuning. Run training on A100 with MultipleNegativesRankingLoss.
5. Evaluation. Compare metrics on test set before and after.
6. Deployment. Replace weights file—infrastructure unchanged.

Data Volume: Minimum and Optimal

From Our Practice: Legal Documents

We worked with a large law firm. Their task was searching through court rulings and regulations. The baseline model BAAI/bge-m3 gave NDCG@10 = 0.68. We fine-tuned on 8000 pairs (6500 synthetic via GPT-4o-mini, 1500 manual from experts). Results:

+24% NDCG without infrastructure changes—just a weight update. The client got a search system that finds relevant documents twice as accurately.

What’s Included

• Analysis of your data and identification of retrieval issues.
• Preparation of training dataset (synthetic + manual labeling if needed).
• Fine-tuning the model on the chosen stack.
• Evaluation on your test queries.
• Deployment of the fine-tuned model in your infrastructure (Docker, SageMaker, Triton).
• Documentation and team training.

Request an evaluation of your project—we'll design the optimal fine-tuning strategy.

Timeline

• Dataset generation: 3-7 days.
• Fine-tuning: 2-4 hours (on A100) up to a day for large volume.
• Evaluation and comparison: 2-3 days.
• Total: 1 to 3 weeks depending on complexity.

How to Evaluate Quality After Fine-Tuning?

We use InformationRetrievalEvaluator from sentence-transformers:

evaluator = InformationRetrievalEvaluator(
    queries=test_queries,
    corpus=test_corpus,
    relevant_docs=relevance_labels,
    precision_recall_at_k=[1,5,10],
    ndcg_at_k=[10],
)

Compare baseline and fine-tuned model. Visualize results and deliver to client.

To avoid common pitfalls, we use sufficient data volume, separate test sets, and normalize embeddings at inference. With 5 years of NLP experience and over 30 successful fine-tuning projects, we guarantee results.

Contact us—we'll evaluate your project and propose the best solution. We work end-to-end: from data analysis to deployment. Get a consultation—it's free.

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.