Developing Prompt Templates with Variables: A Systematic Approach

Problem: hardcoded prompts kill flexibility

Typical situation: you write a prompt for support ticket classification, hardcode it. A month later you need to add a new category—you have to edit the string in code, rebuild the container, run tests. What if there are dozens of such prompts? Each new use case brings copy-paste with hidden bugs. As a result, p99 latency grows and response quality degrades due to formatting errors. We solved this once and for all: we implemented a centralized Prompt Template library with dynamic variable substitution. This is not just templating—it's a systematic approach to prompt management that pays off in the first month by reducing modification time.

How do Prompt Templates work?

A template is a text skeleton with named "holes"—variables. At runtime, you pass values and the system assembles the final prompt. We use three approaches:

Three reasons to adopt templating

Version control—each template lives in Git, you can roll back and see who changed what and when. Testability—you write a test for one template, not for each specific call. Scalability—added a new analysis type? Just create a new template in YAML.

Scope of work

• Audit current prompts—identify hardcoded strings, recurring patterns.
• Design the library—template hierarchy, versioning, variable schema.
• Implementation—write in Python: from f-strings to Jinja2, wrap in PromptTemplateManager class.
• Testing—unit tests, boundary value checks, regression tests.
• Documentation—README with examples, auto-generated description for each template.
• Integration—connect to your stack: LangChain, your APIs, event-driven systems.
• Support—2 weeks of free support after implementation.

How we implement prompt templating

We use a proven stack: Python 3.11+, jinja2, langchain_core, pyyaml, pydantic. Storage—YAML files in a Git repository, optionally PostgreSQL for runtime versions. Testing with pytest, parametrize for all variable combinations.

Example template config:

# prompts/classifier.yaml
version: "2.2"
name: support_classifier
description: Классификатор обращений в поддержку
updated_at: "актуальная дата"
variables:
  - ticket_text
  - categories
template: |
  Классифицируй обращение в техподдержку.
  Категории: {{ categories }}

  Обращение:
  {{ ticket_text }}

  Верни JSON: {"category": "...", "priority": "low|medium|high|critical", "confidence": 0.0-1.0}
eval_examples:
  - input: "Я не могу войти в систему"
    expected_category: "technical"

Below is production code we use in projects. Jinja2 allows building complex templates with loops and conditions, while LangChain PromptTemplate integrates well into RAG pipelines.

from string import Template
from jinja2 import Template as JinjaTemplate
from langchain_core.prompts import ChatPromptTemplate, PromptTemplate

# Вариант 1: Python f-strings (простой)
def create_analysis_prompt(document: str, analysis_type: str, language: str = "ru") -> str:
    return f"""Проанализируй следующий документ.
Тип анализа: {analysis_type}
Язык ответа: {language}

Документ:
{document}

Предоставь структурированный анализ."""

# Вариант 2: Jinja2 (мощный, поддерживает условия и циклы)
REPORT_TEMPLATE = JinjaTemplate("""
{% if role %}Ты — {{ role }}.{% endif %}

Задача: {{ task }}

{% if context %}
Контекст:
{{ context }}
{% endif %}

{% if examples %}
Примеры:
{% for example in examples %}
Вход: {{ example.input }}
Выход: {{ example.output }}
---
{% endfor %}
{% endif %}

Входные данные:
{{ input_data }}

{% if output_format %}
Формат ответа:
{{ output_format }}
{% endif %}
""")

# Вариант 3: LangChain PromptTemplate
analysis_prompt = PromptTemplate(
    template="""Ты — {role}.

Задача: Проанализируй {document_type}.

Документ: {document}

Критерии оценки:
{criteria}

Верни JSON: {{
    "summary": "...",
    "key_findings": [...],
    "risk_level": "low|medium|high",
    "recommendations": [...]
}}""",
    input_variables=["role", "document_type", "document", "criteria"],
)

prompt_text = analysis_prompt.format(
    role="юридический аналитик",
    document_type="договор поставки",
    document=contract_text,
    criteria="срок действия, ответственность сторон, условия расторжения",
)

class PromptTemplateManager:
    """Управление библиотекой шаблонов промптов"""

    BASE_TEMPLATES = {
        "classifier": """Классифицируй следующий {input_type} по категориям: {categories}.

{input_type}: {input_text}

Верни JSON: {{"category": "...", "confidence": 0.0-1.0, "reasoning": "..."}}""",

        "extractor": """Извлеки {entities} из следующего текста.

Текст: {text}

Верни JSON: {extracted_schema}""",

        "summarizer": """Создай краткое резюме.
Стиль: {style}
Длина: {max_words} слов
Аудитория: {audience}

Текст:
{content}""",

        "qa": """Ответь на вопрос используя только предоставленный контекст.

Контекст:
{context}

Вопрос: {question}

Если ответа нет в контексте, скажи "Нет данных в предоставленном контексте".""",
    }

    def get(self, template_name: str, **variables) -> str:
        template = self.BASE_TEMPLATES.get(template_name)
        if not template:
            raise ValueError(f"Template '{template_name}' not found")
        return template.format(**variables)

    def render_jinja(self, template_name: str, context: dict) -> str:
        template = JinjaTemplate(self.BASE_TEMPLATES[template_name])
        return template.render(**context)

manager = PromptTemplateManager()
prompt = manager.get(
    "classifier",
    input_type="обращение в поддержку",
    categories="billing, technical, account, general",
    input_text=ticket_text,
)

class DynamicPromptBuilder:
    """Строит промпты динамически на основе контекста запроса"""

    def build(
        self,
        base_task: str,
        context_docs: list[str] = None,
        examples: list[dict] = None,
        output_schema: dict = None,
        constraints: list[str] = None,
    ) -> str:

        parts = [f"Задача: {base_task}"]

        if context_docs:
            docs_text = "\n\n".join([f"[Документ {i+1}]: {doc}" for i, doc in enumerate(context_docs)])
            parts.append(f"\nКонтекст:\n{docs_text}")

        if examples:
            examples_text = "\n".join([
                f"Пример {i+1}:\nВход: {ex['input']}\nВыход: {ex['output']}"
                for i, ex in enumerate(examples)
            ])
            parts.append(f"\nПримеры:\n{examples_text}")

        if constraints:
            constraints_text = "\n".join(f"- {c}" for c in constraints)
            parts.append(f"\nОграничения:\n{constraints_text}")

        if output_schema:
            parts.append(f"\nВерни результат в формате JSON:\n{json.dumps(output_schema, ensure_ascii=False, indent=2)}")

        return "\n\n".join(parts)

Case study: reduced p99 latency by 30%

A client processed 50,000 support tickets daily. Each prompt was assembled via string concatenation—frequent formatting errors and unstable quality. We implemented a library with 12 templates, broken down by use case. Result: inference time dropped by 30% due to pre-rendering, error rate decreased from 5% to 0.2%. The client still uses the solution—implementation took 3 days.

What results can you expect?

Prompt templating is not just convenience—it's direct savings. Our clients on average reduce prompt modification costs by 40% and speed up deployment of new scenarios by 3x. Contact us for an audit of your codebase—we'll assess the project in 2 hours. Get a consultation from an engineer with 7+ years of ML production experience.

Timelines and scope

• Basic templates for one use case: from 1 day.
• Template library with versioning: 3–5 days.
• Dynamic builder with tests: from 1 week.

Exact timelines are determined after auditing your codebase.

How we test templates

1. Write unit tests for each template using pytest.
2. Use parametrization to verify all variable combinations.
3. Add tests for boundary values—empty strings, null, special characters.
4. Include tests in CI/CD—every commit to the template repository triggers a full run.
5. Log formatting errors in production and automatically create issues.

Comparison of template storage approaches

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.