🤖 LLM Application Patterns

Production-ready patterns for building LLM applications, inspired by Dify and industry best practices.

When to Use This Skill

Use this skill when:

• Designing LLM-powered applications
• Implementing RAG (Retrieval-Augmented Generation)
• Building AI agents with tools
• Setting up LLMOps monitoring
• Choosing between agent architectures

1. RAG Pipeline Architecture

Overview

RAG (Retrieval-Augmented Generation) grounds LLM responses in your data.

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

1.1 Document Ingestion

# 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 & Storage

# 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 Retrieval Strategies

# 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 Generation with Context

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 Architectures

2.1 ReAct Pattern (Reasoning + Acting)

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 Function Calling Pattern

# 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