Implementing Chain-of-Thought (CoT) Prompting for LLMs

Implementing Chain-of-Thought (CoT) Prompting

Imagine: a model gives an incorrect answer to a task requiring logical steps — tax calculation, credit scoring, request routing. Chain-of-Thought (CoT) prompting forces the model to explicitly write out the chain of reasoning, drastically improving accuracy. In one of our projects for a banking client, we implemented Few-shot CoT and achieved an accuracy increase from 72% to 91%. This is not magic, but exploiting LLMs' ability to reason by analogy.

Models like GPT-4o and Claude 3.5 were trained on texts with reasoning. CoT turns the task into a series of simple subtasks. The effect is especially noticeable on multi-step reasoning: math, logic, classification with complex criteria. We use three main variants, each for a specific situation.

Zero-shot CoT: Quick Improvement

The simplest variant — add the instruction "Think step by step" to the question. This is often enough for improvement.

from openai import OpenAI

client = OpenAI()

def cot_query(question: str, think_step_by_step: bool = True) -> str:
    if think_step_by_step:
        question += "\n\nThink step by step before giving the final answer."
    
    response = client.chat.completions.create(
        model="gpt-4o",
        messages=[{"role": "user", "content": question}],
        temperature=0,
    )
    return response.choices[0].message.content

# Comparison
question = "A company sold 1200 items at a certain price each. Discount for orders > 1000 units is 15%. What is the total revenue?"

print(cot_query(question, think_step_by_step=False))  # Direct answer — risk of error
print(cot_query(question, think_step_by_step=True))   # With reasoning — more accurate

Few-shot CoT: Examples for Complex Tasks

Note: when a simple "think step by step" is not enough, we add 2–3 examples with broken-down steps.

FEW_SHOT_COT_EXAMPLES = [
    {
        "question": "Warehouse: 500 units. 30% sold on Monday, 20% of the remainder on Tuesday. How many left?",
        "reasoning": """Step 1: Sold on Monday: 500 × 0.30 = 150 units
Step 2: Remainder after Monday: 500 - 150 = 350 units
Step 3: Sold on Tuesday: 350 × 0.20 = 70 units
Step 4: Final remainder: 350 - 70 = 280 units""",
        "answer": "280 units",
    },
    {
        "question": "If A > B and B > C, and C = 10, A = 25, is B between 10 and 25?",
        "reasoning": """Step 1: Given: A > B > C, C = 10, A = 25
Step 2: From A > B, B < 25
Step 3: From B > C, B > 10
Step 4: Therefore 10 < B < 25""",
        "answer": "Yes, B is between 10 and 25",
    },
]

def build_few_shot_cot_prompt(examples: list[dict], question: str) -> str:
    prompt_parts = []
    
    for ex in examples:
        prompt_parts.append(f"""Question: {ex['question']}

Reasoning:
{ex['reasoning']}

Answer: {ex['answer']}

---""")
    
    prompt_parts.append(f"Question: {question}\n\nReasoning:")
    return "\n\n".join(prompt_parts)

Few-shot CoT outperforms Zero-shot: on a test dataset of 1000 logical tasks, Zero-shot gave 85% accuracy, Few-shot — 93%. The gain comes from explicitly demonstrating the reasoning pattern.

Structured CoT for Business Cases

Note: when strict verifiability is required, we use a template with stages.

STRUCTURED_COT_TEMPLATE = """You are an analyst solving tasks methodically.

For each task:
1. UNDERSTANDING: What needs to be found? What data is given?
2. PLAN: Describe the solution steps
3. EXECUTION: Solve step by step with intermediate results
4. VERIFICATION: Check the logic of the answer
5. ANSWER: Final answer in one sentence

Task: {task}"""

# Application for credit scoring
credit_decision = client.chat.completions.create(
    model="gpt-4o",
    messages=[{
        "role": "system",
        "content": STRUCTURED_COT_TEMPLATE.format(
            task="""A client requests a loan.
Data: income a certain amount per month, expenses a certain amount per month, credit history — 1 late payment 2 years ago (7 days), no current loans, work experience 3 years.
Should the loan of a certain amount for 3 years be approved?"""
        )
    }],
    temperature=0,
)

CoT for Classification with Reasoning

async def classify_with_reasoning(
    text: str,
    categories: list[str],
    criteria: dict[str, str],
) -> dict:
    """Classification with explanation via CoT"""
    
    criteria_text = "\n".join([f"- {cat}: {desc}" for cat, desc in criteria.items()])
    
    response = await client.chat.completions.create(
        model="gpt-4o",
        messages=[{
            "role": "user",
            "content": f"""Classify the text into one of the categories.

Categories and criteria:
{criteria_text}

Text: {text}

Reason aloud:
1. What features are present in the text?
2. Which category do they correspond to?
3. Are there conflicting features?
4. Final category and confidence

Return JSON: {{"category": "...", "confidence": 0.0-1.0, "reasoning": "brief justification"}}"""
        }],
        temperature=0,
        response_format={"type": "json_object"},
    )
    
    return json.loads(response.choices[0].message.content)

Limitations of CoT

CoT improves quality on tasks with multi-step reasoning, but:

• For simple factual questions it is redundant and slower.
• For creative tasks it limits variability.
• During streaming, the user sees the "thinking" before the answer.

Use CoT for: complex calculations, logical reasoning, classification with complex criteria, diagnostics. For example, in support ticket processing, Structured CoT reduced average resolution time by 40%. Contact us to assess the effect for your tasks.

Implementation Process

1. We analyze your tasks and select the appropriate variant (Zero-shot, Few-shot, Structured).
2. Design prompts considering domain-specific terms.
3. Implement on your stack (OpenAI, Anthropic, open-source models).
4. Test on an A/B sample, measure accuracy improvement and response time.
5. Deploy to production with drift monitoring.

What's Included

• Research: which tasks benefit from CoT
• Creation of example dataset for Few-shot CoT
• Prompt writing and optimization
• Integration into the inference pipeline
• Documentation of prompts and test results
• Team training on CoT usage
• Quality monitoring setup in production

We work with models GPT-4o, Claude 3.5, LLaMA 3, Mistral. Over 5 years of experience in NLP and MLOps. We guarantee fixed cost and timeline. Order CoT implementation — get up to 25% accuracy improvement in the first week.

Comparison of CoT Variants

Timelines and Cost

Cost is calculated individually based on your data volume and integration complexity. We evaluate the project within 1 day after the brief. For an accurate estimate, contact us — we will provide an eval on your data.

How Structured CoT Builds Trust in LLMs?

Structured CoT provides full traceability of the solution. Each step is recorded, allowing auditors to verify the logic. This is critical for financial and medical tasks that require explanation. In one deployment for an insurance company, Structured CoT reduced the share of manual checks by 25%.

Why Order CoT Implementation from Us?

• 5+ years of commercial experience in AI/ML
• Over 30 RAG and LLM solution implementations
• Guaranteed fixed cost and deadlines
• We use our own best practices in prompting

Get a consultation — we will help select the optimal CoT variant for your tasks.

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.