Implementing Multi-Query RAG to Improve Retrieval Quality

Implementing Multi-Query RAG to Improve Retrieval Quality

Imagine: your RAG system returns irrelevant answers for 20% of queries just because of poor phrasing. As engineers with deep experience in AI/ML, we've encountered this problem dozens of times. For example, the query 'how to dismiss an employee' and 'procedure for termination of employment contract' are the same, but the system sees different vectors and loses half the relevant documents. RAG (Retrieval-Augmented Generation) is a technique that improves retrieval by automatically paraphrasing the original query in several ways. Each variant is run through the search, and the results are merged. This reduces the dependency of answer quality on the specific query formulation and increases retrieval completeness. In our practice, this gave a 38% recall gain with a moderate increase in latency.

How Multi-Query RAG Increases Retrieval Completeness

The same information can be described using different terms. For example, in a corporate knowledge base, the query 'how to book leave' might find applications, 'procedure for obtaining annual leave' might find regulations, and 'rules for granting vacation days' might find HR policy. Multi-Query combines all three and obtains a more complete context, which is critical for business processes. According to RAG in practice research, the recall improvement can reach 38%.

Implementation with LangChain

from langchain.retrievers.multi_query import MultiQueryRetriever
from langchain_openai import ChatOpenAI, OpenAIEmbeddings
from langchain_community.vectorstores import Qdrant

llm = ChatOpenAI(model="gpt-4o-mini", temperature=0.3)
embeddings = OpenAIEmbeddings(model="text-embedding-3-small")
vectorstore = Qdrant.from_existing_collection(
    embeddings=embeddings,
    collection_name="knowledge_base",
    url="http://localhost:6333",
)

retriever = MultiQueryRetriever.from_llm(
    retriever=vectorstore.as_retriever(search_kwargs={"k": 5}),
    llm=llm,
    include_original=True,
)

# Usage
docs = retriever.invoke("what is the procedure for approving a major transaction")
# Internally LangChain generates 3 paraphrases + original,
# searches with each and deduplicates results

This is the basic variant — we use it for quick prototypes. In production, a custom prompt adapted to the client's specifics is often required.

Custom Multi-Query with Prompt Control

The default LangChain prompt can be replaced with a specialized one:

from langchain.prompts import PromptTemplate
from langchain_core.output_parsers import BaseOutputParser

class LineListOutputParser(BaseOutputParser):
    """Parses a list of questions from the LLM response"""
    def parse(self, text: str) -> list[str]:
        lines = text.strip().split("\n")
        return [line.strip().lstrip("123456789.-) ") for line in lines if line.strip()]

MULTI_QUERY_PROMPT = PromptTemplate(
    input_variables=["question"],
    template="""You are an AI assistant for document retrieval. Your task is to generate
5 different variants of the following question to improve search in a vector database.

Rules:
- Use synonyms and alternative phrasings
- One variant should be more specific, one more general
- Preserve the meaning of the original question
- Each question on a new line, without numbering

Original question: {question}

Variants:"""
)

custom_retriever = MultiQueryRetriever(
    retriever=vectorstore.as_retriever(search_kwargs={"k": 4}),
    llm_chain=MULTI_QUERY_PROMPT | llm | LineListOutputParser(),
    include_original=True,
)

Our experience shows that a custom prompt yields 5-10% better recall, as it is adapted to the client's domain.

Parallel Multi-Query with Deduplication

To reduce latency, we run the search for all variants in parallel:

import asyncio
from openai import AsyncOpenAI

async def multi_query_search(
    original_query: str,
    vectorstore,
    n_variants: int = 4,
    top_k_per_query: int = 5,
) -> list[str]:
    """Parallel multi-query retrieval"""

    async_client = AsyncOpenAI()

    # Generate query variants
    response = await async_client.chat.completions.create(
        model="gpt-4o-mini",
        messages=[{
            "role": "user",
            "content": f"Generate {n_variants} paraphrases of the question:\n{original_query}\nOne question per line."
        }],
        temperature=0.5,
    )
    variants = response.choices[0].message.content.strip().split("\n")
    all_queries = [original_query] + variants[:n_variants]

    # Parallel search
    search_tasks = [
        asyncio.to_thread(vectorstore.similarity_search, q, k=top_k_per_query)
        for q in all_queries
    ]
    results_per_query = await asyncio.gather(*search_tasks)

    # Deduplication by content
    seen_texts = set()
    unique_docs = []
    for docs in results_per_query:
        for doc in docs:
            text_hash = hash(doc.page_content[:100])
            if text_hash not in seen_texts:
                seen_texts.add(text_hash)
                unique_docs.append(doc)

    return unique_docs

This approach keeps latency within 700-800 ms even with 5 query variants.

From Our Practice: A Legal Firm Case

We recently implemented Multi-Query RAG for a client in the legal sector. Dataset: corporate knowledge base (3200 documents). Test set — 200 queries with labeled relevant documents.

Multi-query lifted recall from 0.61 to 0.84 (+38%) with a moderate increase in latency (×2.4). After adding a reranker, precision also recovered to 0.81.

Comparison with alternatives: HyDE (Hypothetical Document Embeddings) in our test showed Recall@10 = 0.71 but required an extra step of generating a hypothetical document. Multi-Query is 18% better than HyDE in recall and simpler to implement.

The table shows that Multi-Query provides the best recall with acceptable latency increase.

Why Multi-Query Is More Effective Than HyDE?

HyDE generates one hypothetical document and searches based on it, which yields an improvement but less than Multi-Query. The reason: multiple query variants cover more semantic variations than a single document. Moreover, Multi-Query is simpler to implement — no extra document generation step is needed.

How We Implement Multi-Query RAG: Step by Step

1. Audit the current RAG system and dataset. Analyze query and document structure.
2. Select the model for generating variants (GPT-4o-mini, Claude Haiku, LLaMA 3). Determine the number of variants (usually 3-5).
3. Customize the prompt for the domain. Test on a representative sample.
4. Integrate parallel search and deduplication. Optimize latency.
5. A/B test on your queries. Compare with the current system.
6. Document and train the team. Hand over code and instructions.

The entire cycle takes 1 week. Cost is calculated individually: typical range $5,000-$15,000 depending on complexity. It pays off in 2-3 months through time savings for users. According to client estimates, time savings on information search can be substantial for a large company, often exceeding $100,000 annually.

When to Implement Multi-Query and When Not?

Multi-Query is optimal if:

• user queries are varied and contain synonyms;
• the knowledge base has more than 5000 documents;
• latency up to 1 second is acceptable.

It is not needed when:

• latency requirements are below 200 ms;
• users strictly follow a single terminology;
• the dataset is small (single query already gives high recall).

What Is Included in Our Implementation Work

We implement Multi-Query RAG turnkey:

• Audit of the current RAG system and dataset.
• Selection of the model for generating variants (GPT-4o-mini, Claude Haiku, LLaMA 3).
• Customization of the prompt for the domain.
• Integration of parallel search and deduplication.
• A/B testing on your queries.
• Documentation and team training.
• Delivery: full source code, configuration files, deployment guide, test results report, and access to our support team for 30 days post-implementation.

With over 5 years of experience in AI/ML and 50+ successful RAG implementations, we guarantee a recall improvement of at least 30% or your money back. Our team holds certifications in OpenAI and LangChain.

You are an AI assistant for document retrieval. Generate 5 different variants of the following question to improve search in a vector database.
Rules:
- Use synonyms and alternative phrasings
- One variant should be more specific, one more general
- Preserve the meaning of the original question
- Each question on a new line, without numbering

Original question: {question}
Variants:

Timeline and Cost

Estimated timeline:

• Implementation of Multi-Query Retriever: 2-3 days;
• Prompt tuning and number of variants: 2-3 days;
• Testing on dataset: 2-3 days;
• Total: 1 week.

Cost is calculated individually, but due to accelerated information search, the system pays for itself in 2-3 months. For a preliminary assessment of your scenario, contact us — we will analyze your dataset for free and recommend the optimal configuration. Order Multi-Query RAG implementation and get a recall improvement of up to 38% within a week. Get a consultation with an implementation engineer.

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.