LLM App Patterns

🤖 LLM 应用模式

构建 LLM 应用的生产级模式，灵感来自 Dify 和行业最佳实践。

何时使用此技能

在以下场景使用此技能：

• 设计 LLM 驱动的应用
• 实现 RAG（检索增强生成）
• 构建带工具的 AI Agent
• 搭建 LLMOps 监控
• 在不同 Agent 架构之间做选择

1. RAG 流水线架构

概述

RAG（检索增强生成）将 LLM 的回答锚定在你的数据上。

┌─────────────┐     ┌─────────────┐     ┌─────────────┐
│   Ingest    │────▶│   Retrieve  │────▶│   Generate  │
│  Documents  │     │   Context   │     │   Response  │
└─────────────┘     └─────────────┘     └─────────────┘
      │                   │                   │
      ▼                   ▼                   ▼
 ┌─────────┐       ┌───────────┐       ┌───────────┐
 │ Chunking│       │  Vector   │       │    LLM    │
 │Embedding│       │  Search   │       │  + Context│
 └─────────┘       └───────────┘       └───────────┘

1.1 文档摄入

# Chunking strategies
class ChunkingStrategy:
    # Fixed-size chunks (simple but may break context)
    FIXED_SIZE = "fixed_size"  # e.g., 512 tokens

    # Semantic chunking (preserves meaning)
    SEMANTIC = "semantic"      # Split on paragraphs/sections

    # Recursive splitting (tries multiple separators)
    RECURSIVE = "recursive"    # ["\n\n", "\n", " ", ""]

    # Document-aware (respects structure)
    DOCUMENT_AWARE = "document_aware"  # Headers, lists, etc.

# Recommended settings
CHUNK_CONFIG = {
    "chunk_size": 512,       # tokens
    "chunk_overlap": 50,     # token overlap between chunks
    "separators": ["\n\n", "\n", ". ", " "],
}

1.2 Embedding 与存储

# Vector database selection
VECTOR_DB_OPTIONS = {
    "pinecone": {
        "use_case": "Production, managed service",
        "scale": "Billions of vectors",
        "features": ["Hybrid search", "Metadata filtering"]
    },
    "weaviate": {
        "use_case": "Self-hosted, multi-modal",
        "scale": "Millions of vectors",
        "features": ["GraphQL API", "Modules"]
    },
    "chromadb": {
        "use_case": "Development, prototyping",
        "scale": "Thousands of vectors",
        "features": ["Simple API", "In-memory option"]
    },
    "pgvector": {
        "use_case": "Existing Postgres infrastructure",
        "scale": "Millions of vectors",
        "features": ["SQL integration", "ACID compliance"]
    }
}

# Embedding model selection
EMBEDDING_MODELS = {
    "openai/text-embedding-3-small": {
        "dimensions": 1536,
        "cost": "$0.02/1M tokens",
        "quality": "Good for most use cases"
    },
    "openai/text-embedding-3-large": {
        "dimensions": 3072,
        "cost": "$0.13/1M tokens",
        "quality": "Best for complex queries"
    },
    "local/bge-large": {
        "dimensions": 1024,
        "cost": "Free (compute only)",
        "quality": "Comparable to OpenAI small"
    }
}

1.3 检索策略

# Basic semantic search
def semantic_search(query: str, top_k: int = 5):
    query_embedding = embed(query)
    results = vector_db.similarity_search(
        query_embedding,
        top_k=top_k
    )
    return results

# Hybrid search (semantic + keyword)
def hybrid_search(query: str, top_k: int = 5, alpha: float = 0.5):
    """
    alpha=1.0: Pure semantic
    alpha=0.0: Pure keyword (BM25)
    alpha=0.5: Balanced
    """
    semantic_results = vector_db.similarity_search(query)
    keyword_results = bm25_search(query)

    # Reciprocal Rank Fusion
    return rrf_merge(semantic_results, keyword_results, alpha)

# Multi-query retrieval
def multi_query_retrieval(query: str):
    """Generate multiple query variations for better recall"""
    queries = llm.generate_query_variations(query, n=3)
    all_results = []
    for q in queries:
        all_results.extend(semantic_search(q))
    return deduplicate(all_results)

# Contextual compression
def compressed_retrieval(query: str):
    """Retrieve then compress to relevant parts only"""
    docs = semantic_search(query, top_k=10)
    compressed = llm.extract_relevant_parts(docs, query)
    return compressed

1.4 带上下文的生成

RAG_PROMPT_TEMPLATE = """
Answer the user's question based ONLY on the following context.
If the context doesn't contain enough information, say "I don't have enough information to answer that."

Context:
{context}

Question: {question}

Answer:"""

def generate_with_rag(question: str):
    # Retrieve
    context_docs = hybrid_search(question, top_k=5)
    context = "\n\n".join([doc.content for doc in context_docs])

    # Generate
    prompt = RAG_PROMPT_TEMPLATE.format(
        context=context,
        question=question
    )

    response = llm.generate(prompt)

    # Return with citations
    return {
        "answer": response,
        "sources": [doc.metadata for doc in context_docs]
    }

2. Agent 架构

2.1 ReAct 模式（推理 + 行动）

Thought: I need to search for information about X
Action: search("X")
Observation: [search results]
Thought: Based on the results, I should...
Action: calculate(...)
Observation: [calculation result]
Thought: I now have enough information
Action: final_answer("The answer is...")

REACT_PROMPT = """
You are an AI assistant that can use tools to answer questions.

Available tools:
{tools_description}

Use this format:
Thought: [your reasoning about what to do next]
Action: [tool_name(arguments)]
Observation: [tool result - this will be filled in]
... (repeat Thought/Action/Observation as needed)
Thought: I have enough information to answer
Final Answer: [your final response]

Question: {question}
"""

class ReActAgent:
    def __init__(self, tools: list, llm):
        self.tools = {t.name: t for t in tools}
        self.llm = llm
        self.max_iterations = 10

    def run(self, question: str) -> str:
        prompt = REACT_PROMPT.format(
            tools_description=self._format_tools(),
            question=question
        )

        for _ in range(self.max_iterations):
            response = self.llm.generate(prompt)

            if "Final Answer:" in response:
                return self._extract_final_answer(response)

            action = self._parse_action(response)
            observation = self._execute_tool(action)
            prompt += f"\nObservation: {observation}\n"

        return "Max iterations reached"

2.2 函数调用模式

# Define tools as functions with schemas
TOOLS = [
    {
        "name": "search_web",
        "description": "Search the web for current information",
        "parameters": {
            "type": "object",
            "properties": {
                "query": {
                    "type": "string",
                    "description": "Search query"
                }
            },
            "required": ["query"]
        }
    },
    {
        "name": "calculate",
        "description": "Perform mathematical calculations",
        "parameters": {
            "type": "object",
            "properties": {
                "expression": {
                    "type": "string",
                    "description": "Math expression to evaluate"
                }
            },
            "required": ["expression"]
        }
    }
]

class FunctionCallingAgent:
    def run(self, question: str) -> str:
        messages = [{"role": "user", "content": question}]

        while True:
            response = self.llm.chat(
                messages=messages,
                tools=TOOLS,
                tool_choice="auto"
            )

            if response.tool_calls:
                for tool_call in response.tool_calls:
                    result = self._execute_tool(
                        tool_call.name,
                        tool_call.arguments
                    )
                    messages.append({
                        "role": "tool",
                        "tool_call_id": tool_call.id,
                        "content": str(result)
                    })
            else:
                return response.content

2.3 规划-执行模式

class PlanAndExecuteAgent:
    """
    1. Create a plan (list of steps)
    2. Execute each step
    3. Replan if needed
    """

    def run(self, task: str) -> str:
        # Planning phase
        plan = self.planner.create_plan(task)
        # Returns: ["Step 1: ...", "Step 2: ...", ...]

        results = []
        for step in plan:
            # Execute each step
            result = self.executor.execute(step, context=results)
            results.append(result)

            # Check if replan needed
            if self._needs_replan(task, results):
                new_plan = self.planner.replan(
                    task,
                    completed=results,
                    remaining=plan[len(results):]
                )
                plan = new_plan

        # Synthesize final answer
        return self.synthesizer.summarize(task, results)

2.4 多 Agent 协作

class AgentTeam:
    """
    Specialized agents collaborating on complex tasks
    """

    def __init__(self):
        self.agents = {
            "researcher": ResearchAgent(),
            "analyst": AnalystAgent(),
            "writer": WriterAgent(),
            "critic": CriticAgent()
        }
        self.coordinator = CoordinatorAgent()

    def solve(self, task: str) -> str:
        # Coordinator assigns subtasks
        assignments = self.coordinator.decompose(task)

        results = {}
        for assignment in assignments:
            agent = self.agents[assignment.agent]
            result = agent.execute(
                assignment.subtask,
                context=results
            )
            results[assignment.id] = result

        # Critic reviews
        critique = self.agents["critic"].review(results)

        if critique.needs_revision:
            # Iterate with feedback
            return self.solve_with_feedback(task, results, critique)

        return self.coordinator.synthesize(results)

3. Prompt IDE 模式

3.1 带变量的 Prompt 模板

class PromptTemplate:
    def __init__(self, template: str, variables: list[str]):
        self.template = template
        self.variables = variables

    def format(self, **kwargs) -> str:
        # Validate all variables provided
        missing = set(self.variables) - set(kwargs.keys())
        if missing:
            raise ValueError(f"Missing variables: {missing}")

        return self.template.format(**kwargs)

    def with_examples(self, examples: list[dict]) -> str:
        """Add few-shot examples"""
        example_text = "\n\n".join([
            f"Input: {ex['input']}\nOutput: {ex['output']}"
            for ex in examples
        ])
        return f"{example_text}\n\n{self.template}"

# Usage
summarizer = PromptTemplate(
    template="Summarize the following text in {style} style:\n\n{text}",
    variables=["style", "text"]
)

prompt = summarizer.format(
    style="professional",
    text="Long article content..."
)

3.2 Prompt 版本管理与 A/B 测试

class PromptRegistry:
    def __init__(self, db):
        self.db = db

    def register(self, name: str, template: str, version: str):
        """Store prompt with version"""
        self.db.save({
            "name": name,
            "template": template,
            "version": version,
            "created_at": datetime.now(),
            "metrics": {}
        })

    def get(self, name: str, version: str = "latest") -> str:
        """Retrieve specific version"""
        return self.db.get(name, version)

    def ab_test(self, name: str, user_id: str) -> str:
        """Return variant based on user bucket"""
        variants = self.db.get_all_versions(name)
        bucket = hash(user_id) % len(variants)
        return variants[bucket]

    def record_outcome(self, prompt_id: str, outcome: dict):
        """Track prompt performance"""
        self.db.update_metrics(prompt_id, outcome)

3.3 Prompt 链式调用

class PromptChain:
    """
    Chain prompts together, passing output as input to next
    """

    def __init__(self, steps: list[dict]):
        self.steps = steps

    def run(self, initial_input: str) -> dict:
        context = {"input": initial_input}
        results = []

        for step in self.steps:
            prompt = step["prompt"].format(**context)
            output = llm.generate(prompt)

            # Parse output if needed
            if step.get("parser"):
                output = step"parser"

            context[step["output_key"]] = output
            results.append({
                "step": step["name"],
                "output": output
            })

        return {
            "final_output": context[self.steps[-1]["output_key"]],
            "intermediate_results": results
        }

# Example: Research → Analyze → Summarize
chain = PromptChain([
    {
        "name": "research",
        "prompt": "Research the topic: {input}",
        "output_key": "research"
    },
    {
        "name": "analyze",
        "prompt": "Analyze these findings:\n{research}",
        "output_key": "analysis"
    },
    {
        "name": "summarize",
        "prompt": "Summarize this analysis in 3 bullet points:\n{analysis}",
        "output_key": "summary"
    }
])

4. LLMOps 与可观测性

4.1 需要追踪的指标

LLM_METRICS = {
    # Performance
    "latency_p50": "50th percentile response time",
    "latency_p99": "99th percentile response time",
    "tokens_per_second": "Generation speed",

    # Quality
    "user_satisfaction": "Thumbs up/down ratio",
    "task_completion": "% tasks completed successfully",
    "hallucination_rate": "% responses with factual errors",

    # Cost
    "cost_per_request": "Average $ per API call",
    "tokens_per_request": "Average tokens used",
    "cache_hit_rate": "% requests served from cache",

    # Reliability
    "error_rate": "% failed requests",
    "timeout_rate": "% requests that timed out",
    "retry_rate": "% requests needing retry"
}

4.2 日志与链路追踪

import logging
from opentelemetry import trace

tracer = trace.get_tracer(__name__)

class LLMLogger:
    def log_request(self, request_id: str, data: dict):
        """Log LLM request for debugging and analysis"""
        log_entry = {
            "request_id": request_id,
            "timestamp": datetime.now().isoformat(),
            "model": data["model"],
            "prompt": data["prompt"][:500],  # Truncate for storage
            "prompt_tokens": data["prompt_tokens"],
            "temperature": data.get("temperature", 1.0),
            "user_id": data.get("user_id"),
        }
        logging.info(f"LLM_REQUEST: {json.dumps(log_entry)}")

    def log_response(self, request_id: str, data: dict):
        """Log LLM response"""
        log_entry = {
            "request_id": request_id,
            "completion_tokens": data["completion_tokens"],
            "total_tokens": data["total_tokens"],
            "latency_ms": data["latency_ms"],
            "finish_reason": data["finish_reason"],
            "cost_usd": self._calculate_cost(data),
        }
        logging.info(f"LLM_RESPONSE: {json.dumps(log_entry)}")

# Distributed tracing
@tracer.start_as_current_span("llm_call")
def call_llm(prompt: str) -> str:
    span = trace.get_current_span()
    span.set_attribute("prompt.length", len(prompt))

    response = llm.generate(prompt)

    span.set_attribute("response.length", len(response))
    span.set_attribute("tokens.total", response.usage.total_tokens)

    return response.content

4.3 评估框架

class LLMEvaluator:
    """
    Evaluate LLM outputs for quality
    """

    def evaluate_response(self,
                          question: str,
                          response: str,
                          ground_truth: str = None) -> dict:
        scores = {}

        # Relevance: Does it answer the question?
        scores["relevance"] = self._score_relevance(question, response)

        # Coherence: Is it well-structured?
        scores["coherence"] = self._score_coherence(response)

        # Groundedness: Is it based on provided context?
        scores["groundedness"] = self._score_groundedness(response)

        # Accuracy: Does it match ground truth?
        if ground_truth:
            scores["accuracy"] = self._score_accuracy(response, ground_truth)

        # Harmfulness: Is it safe?
        scores["safety"] = self._score_safety(response)

        return scores

    def run_benchmark(self, test_cases: list[dict]) -> dict:
        """Run evaluation on test set"""
        results = []
        for case in test_cases:
            response = llm.generate(case["prompt"])
            scores = self.evaluate_response(
                question=case["prompt"],
                response=response,
                ground_truth=case.get("expected")
            )
            results.append(scores)

        return self._aggregate_scores(results)

5. 生产环境模式

5.1 缓存策略

import hashlib
from functools import lru_cache

class LLMCache:
    def __init__(self, redis_client, ttl_seconds=3600):
        self.redis = redis_client
        self.ttl = ttl_seconds

    def _cache_key(self, prompt: str, model: str, **kwargs) -> str:
        """Generate deterministic cache key"""
        content = f"{model}:{prompt}:{json.dumps(kwargs, sort_keys=True)}"
        return hashlib.sha256(content.encode()).hexdigest()

    def get_or_generate(self, prompt: str, model: str, **kwargs) -> str:
        key = self._cache_key(prompt, model, **kwargs)

        # Check cache
        cached = self.redis.get(key)
        if cached:
            return cached.decode()

        # Generate
        response = llm.generate(prompt, model=model, **kwargs)

        # Cache (only cache deterministic outputs)
        if kwargs.get("temperature", 1.0) == 0:
            self.redis.setex(key, self.ttl, response)

        return response

5.2 限流与重试

import time
from tenacity import retry, wait_exponential, stop_after_attempt

class RateLimiter:
    def __init__(self, requests_per_minute: int):
        self.rpm = requests_per_minute
        self.timestamps = []

    def acquire(self):
        """Wait if rate limit would be exceeded"""
        now = time.time()

        # Remove old timestamps
        self.timestamps = [t for t in self.timestamps if now - t < 60]

        if len(self.timestamps) >= self.rpm:
            sleep_time = 60 - (now - self.timestamps[0])
            time.sleep(sleep_time)

        self.timestamps.append(time.time())

# Retry with exponential backoff
@retry(
    wait=wait_exponential(multiplier=1, min=4, max=60),
    stop=stop_after_attempt(5)
)
def call_llm_with_retry(prompt: str) -> str:
    try:
        return llm.generate(prompt)
    except RateLimitError:
        raise  # Will trigger retry
    except APIError as e:
        if e.status_code >= 500:
            raise  # Retry server errors
        raise  # Don't retry client errors

5.3 降级策略

class LLMWithFallback:
    def __init__(self, primary: str, fallbacks: list[str]):
        self.primary = primary
        self.fallbacks = fallbacks

    def generate(self, prompt: str, **kwargs) -> str:
        models = [self.primary] + self.fallbacks

        for model in models:
            try:
                return llm.generate(prompt, model=model, **kwargs)
            except (RateLimitError, APIError) as e:
                logging.warning(f"Model {model} failed: {e}")
                continue

        raise AllModelsFailedError("All models exhausted")

# Usage
llm_client = LLMWithFallback(
    primary="gpt-4-turbo",
    fallbacks=["gpt-3.5-turbo", "claude-3-sonnet"]
)

架构决策矩阵

资源

• Dify Platform
• LangChain Docs
• LlamaIndex
• Anthropic Cookbook

局限性

• 仅在任务明确匹配上述范围时使用此技能。
• 不要将输出视为环境特定验证、测试或专家评审的替代。
• 如果缺少必要输入、权限、安全边界或成功标准，停下来询问澄清。