AI-Native Development

Overview

AI-Native Development focuses on building applications where AI is a first-class citizen, not an afterthought. This skill provides comprehensive patterns for integrating LLMs, implementing RAG (Retrieval-Augmented Generation), using vector databases, building agentic workflows, and optimizing AI application performance and cost.

Real-World Use Cases:

• AI-Powered Documentation Search: Semantic search over technical docs with hybrid retrieval
• Code Review Assistant: Automated code analysis with security, performance, and style checks
• Customer Support Chatbot: RAG-based Q&A with citation of knowledge base articles
• Content Analysis Pipeline: Multi-agent workflows for analyzing articles, tutorials, and documentation
• Research Assistant: Multi-query RAG with source tracking and synthesis

When to use this skill:

• Building chatbots, Q&A systems, or conversational interfaces
• Implementing semantic search or recommendation engines
• Creating AI agents that can use tools and take actions
• Integrating LLMs (OpenAI, Anthropic, Ollama local models) into applications
• Building RAG systems for knowledge retrieval
• Optimizing AI costs and latency (93% savings with local models)
• Implementing AI observability and monitoring
• Setting up local LLM inference for CI/CD pipelines

---

Why AI-Native Development Matters

Traditional software is deterministic; AI-native applications are probabilistic:

• Context is Everything: LLMs need relevant context to provide accurate answers
• RAG Over Fine-Tuning: Retrieval is cheaper and more flexible than fine-tuning
• Embeddings Enable Semantic Search: Move beyond keyword matching to understanding meaning
• Agentic Workflows: LLMs can reason, plan, and use tools autonomously
• Cost Management: Token usage directly impacts operational costs
• Observability: Debugging probabilistic systems requires new approaches
• Prompt Engineering: How you ask matters as much as what you ask

---

Core Concepts

1. Embeddings & Vector Search

Embeddings are vector representations of text that capture semantic meaning. Similar concepts have similar vectors.

Key Capabilities:

• Convert text to high-dimensional vectors (1536 or 3072 dimensions)
• Measure semantic similarity using cosine similarity
• Find relevant documents through vector search
• Batch process for efficiency

Detailed Implementation: See references/vector-databases.md for:

• OpenAI embeddings setup and batch processing
• Cosine similarity algorithms
• Chunking strategies (500-1000 tokens with 10-20% overlap)

2. Vector Databases

Store and retrieve embeddings efficiently at scale.

Popular Options:

• Pinecone: Serverless, managed service ($0.096/hour)
• Chroma: Open source, self-hosted
• Weaviate: Flexible schema, hybrid search
• Qdrant: Rust-based, high performance

Detailed Implementation: See references/vector-databases.md for:

• Complete setup guides for each database
• Upsert, query, update, delete operations
• Metadata filtering and hybrid search
• Cost comparison and best practices

3. RAG (Retrieval-Augmented Generation)

RAG combines retrieval systems with LLMs to provide accurate, grounded answers.

Core Pattern:

1. Retrieve relevant documents from vector database
2. Construct context from top results
3. Generate answer with LLM using retrieved context

Advanced Patterns:

• RAG with citations and source tracking
• Hybrid search (semantic + keyword)
• Multi-query RAG for better recall
• HyDE (Hypothetical Document Embeddings)
• Contextual compression for relevance

Detailed Implementation: See references/rag-patterns.md for:

• Basic and advanced RAG patterns with full code
• Citation strategies
• Hybrid search with Reciprocal Rank Fusion
• Conversation memory patterns
• Error handling and validation

4. Function Calling & Tool Use

Enable LLMs to use external tools and APIs reliably.

Capabilities:

• Define tools with JSON schemas
• Execute functions based on LLM decisions
• Handle parallel tool calls
• Stream responses with tool use

Detailed Implementation: See references/function-calling.md for:

• Tool definition patterns (OpenAI and Anthropic)
• Function calling loops
• Parallel and streaming tool execution
• Input validation with Zod
• Error handling and fallback strategies

5. Agentic Workflows

Enable LLMs to reason, plan, and take autonomous actions.

Patterns:

• ReAct: Reasoning + Acting loop with observations
• Tree of Thoughts: Explore multiple reasoning paths
• Multi-Agent: Specialized agents collaborating on complex tasks
• Autonomous Agents: Self-directed goal achievement

Detailed Implementation: See references/agentic-workflows.md for:

• Complete ReAct loop implementation
• Tree of Thoughts exploration
• Multi-agent coordinator patterns
• Agent memory management
• Error recovery and safety guards

5.1 Multi-Agent Orchestration (Opus 4.5)

Advanced multi-agent patterns leveraging Opus 4.5's extended thinking capabilities.

When to Use Extended Thinking:

• Coordinating 3+ specialized agents
• Complex dependency resolution between agent outputs
• Dynamic task allocation based on agent capabilities
• Conflict resolution when agents produce contradictory results

Orchestrator Pattern:

interface AgentTask {
  id: string;
  type: 'research' | 'code' | 'review' | 'design';
  input: unknown;
  dependencies: string[]; // Task IDs that must complete first
}

interface AgentResult {
  taskId: string;
  output: unknown;
  confidence: number;
  reasoning: string;
}

async function orchestrateAgents(
  goal: string,
  availableAgents: Agent[]
): Promise<AgentResult[]> {
  // Step 1: Use extended thinking to decompose goal into tasks
  const taskPlan = await planTasks(goal, availableAgents);

  // Step 2: Build dependency graph
  const dependencyGraph = buildDependencyGraph(taskPlan.tasks);

  // Step 3: Execute tasks respecting dependencies
  const results: AgentResult[] = [];
  const completed = new Set<string>();

  while (completed.size < taskPlan.tasks.length) {
    // Find tasks with satisfied dependencies
    const ready = taskPlan.tasks.filter(task =>
      !completed.has(task.id) &&
      task.dependencies.every(dep => completed.has(dep))
    );

    // Execute ready tasks in parallel
    const batchResults = await Promise.all(
      ready.map(task => executeAgentTask(task, availableAgents))
    );

    // Validate results - use extended thinking for conflicts
    const validatedResults = await validateAndResolveConflicts(
      batchResults,
      results
    );

    results.push(...validatedResults);
    ready.forEach(task => completed.add(task.id));
  }

  return results;
}

Task Planning with Extended Thinking:

Based on Anthropic's Extended Thinking documentation:

import Anthropic from '@anthropic-ai/sdk';

const anthropic = new Anthropic();

async function planTasks(
  goal: string,
  agents: Agent[]
): Promise<{ tasks: AgentTask[]; rationale: string }> {
  // Extended thinking requires budget_tokens < max_tokens
  // Minimum budget: 1,024 tokens
  const response = await anthropic.messages.create({
    model: 'claude-opus-4-5-20251101', // Or claude-sonnet-4-5-20250929
    max_tokens: 16000,
    thinking: {
      type: 'enabled',
      budget_tokens: 10000 // Extended thinking for complex planning
    },
    messages: [{
      role: 'user',
      content: `
        Goal: ${goal}

        Available agents and their capabilities:
        ${agents.map(a => `- ${a.name}: ${a.capabilities.join(', ')}`).join('\n')}

        Decompose this goal into tasks. For each task, specify:
        1. Which agent should handle it
        2. What input it needs
        3. Which other tasks it depends on
        4. Expected output format

        Think carefully about:
        - Optimal parallelization opportunities
        - Potential conflicts between agent outputs
        - Information that needs to flow between tasks
      `
    }]
  });

  // Response contains thinking blocks followed by text blocks
  // content: [{ type: 'thinking', thinking: '...' }, { type: 'text', text: '...' }]
  return parseTaskPlan(response);
}

Conflict Resolution:

async function validateAndResolveConflicts(
  newResults: AgentResult[],
  existingResults: AgentResult[]
): Promise<AgentResult[]> {
  // Check for conflicts with existing results
  const conflicts = detectConflicts(newResults, existingResults);

  if (conflicts.length === 0) {
    return newResults;
  }

  // Use extended thinking to resolve conflicts
  const resolution = await anthropic.messages.create({
    model: 'claude-opus-4-5-20251101',
    max_tokens: 8000,
    thinking: {
      type: 'enabled',
      budget_tokens: 5000
    },
    messages: [{
      role: 'user',
      content: `
        The following agent outputs conflict:

        ${conflicts.map(c => `
          Conflict: ${c.description}
          Agent A (${c.agentA.name}): ${JSON.stringify(c.resultA)}
          Agent B (${c.agentB.name}): ${JSON.stringify(c.resultB)}
        `).join('\n\n')}

        Analyze each conflict and determine:
        1. Which output is more likely correct and why
        2. If both have merit, how to synthesize them
        3. What additional verification might be needed
      `
    }]
  });

  return applyResolutions(newResults, resolution);
}

Adaptive Agent Selection:

async function selectOptimalAgent(
  task: AgentTask,
  agents: Agent[],
  context: ExecutionContext
): Promise<Agent> {
  // Score each agent based on:
  // - Capability match
  // - Current load
  // - Historical performance on similar tasks
  // - Cost (model tier)

  const scores = agents.map(agent => ({
    agent,
    score: calculateAgentScore(agent, task, context)
  }));

  // For complex tasks, use Opus; for simple tasks, use Haiku
  const complexity = assessTaskComplexity(task);

  if (complexity > 0.7) {
    // Filter to agents that can use Opus
    const opusCapable = scores.filter(s => s.agent.supportsOpus);
    return opusCapable.sort((a, b) => b.score - a.score)[0].agent;
  }

  return scores.sort((a, b) => b.score - a.score)[0].agent;
}

Agent Communication Protocol:

interface AgentMessage {
  from: string;
  to: string | 'broadcast';
  type: 'request' | 'response' | 'update' | 'conflict';
  payload: unknown;
  timestamp: Date;
}

class AgentCommunicationBus {
  private messages: AgentMessage[] = [];
  private subscribers: Map<string, (msg: AgentMessage) => void> = new Map();

  send(message: AgentMessage): void {
    this.messages.push(message);

    if (message.to === 'broadcast') {
      this.subscribers.forEach(callback => callback(message));
    } else {
      this.subscribers.get(message.to)?.(message);
    }
  }

  subscribe(agentId: string, callback: (msg: AgentMessage) => void): void {
    this.subscribers.set(agentId, callback);
  }

  getHistory(agentId: string): AgentMessage[] {
    return this.messages.filter(
      m => m.from === agentId || m.to === agentId || m.to === 'broadcast'
    );
  }
}

6. Multi-Agent Synthesis & Aggregation

When multiple specialized agents analyze content, synthesize their findings into coherent output.

Real-World Example: Code Review Multi-Agent System

async def multi_agent_code_review(code: str, language: str) -> ReviewReport:
    """Multi-agent code analysis with synthesis."""
    # Fan-out: Run specialized review agents in parallel
    agents = [
        SecurityReviewAgent(),      # OWASP Top 10, injection risks
        PerformanceAnalyzer(),      # Big-O complexity, database queries
        StyleEnforcer(),             # PEP8, ESLint standards
        TestCoverageChecker(),      # Missing test cases
        DocumentationReviewer(),    # Docstrings, comments
        DependencyAuditor()         # Outdated libs, CVEs
    ]

    findings = await asyncio.gather(
        *[agent.analyze(code, language) for agent in agents],
        return_exceptions=True  # Don't fail all if one fails
    )

    # Filter successful results with confidence scores
    valid_findings = [
        f for f in findings
        if not isinstance(f, Exception) and f.confidence > 0.6
    ]

    # Fan-in: Synthesize into coherent review report
    return await synthesize_review_report(valid_findings)

Real-World Example: E-commerce Product Categorization

async def categorize_product(product_data: dict) -> ProductCategories:
    """Multi-agent product classification."""
    agents = [
        ImageClassifierAgent(),     # Visual category from image
        TextAnalysisAgent(),        # Category from title/description
        AttributeExtractor(),       # Structured attributes (color, size)
        SimilarProductFinder(),     # Category from similar products
    ]

    classifications = await asyncio.gather(
        *[agent.predict(product_data) for agent in agents]
    )

    # Resolve conflicts with weighted voting
    return await weighted_category_vote(classifications)

Confidence Score Handling:

def resolve_conflicts(findings: list[AgentFinding]) -> dict:
    """When agents disagree, prioritize by confidence."""
    conflicts = detect_contradictions(findings)

    for conflict in conflicts:
        # Higher confidence wins
        winner = max(conflict.agents, key=lambda a: a.confidence)
        record_resolution(
            conflict=conflict.description,
            resolution=winner.recommendation,
            priority_agent=winner.agent_type,
            reasoning=f"Confidence {winner.confidence:.2f} > others"
        )

7. Streaming Responses

Deliver real-time AI responses for better UX.

Capabilities:

• Stream LLM output token-by-token
• Server-Sent Events (SSE) for web clients
• Streaming with function calls
• Backpressure handling

Detailed Implementation: See ../streaming-api-patterns/SKILL.md for streaming patterns

8. LLM-as-Judge Evaluation (v1.1.0)

Use LLMs to evaluate LLM outputs for quality assurance.

Quality Aspects:

QUALITY_DIMENSIONS = {
    "relevance": "How relevant is the output to the input?",
    "depth": "How thorough and detailed is the analysis?",
    "coherence": "How well-structured and clear is the content?",
    "accuracy": "Are facts and code snippets correct?",
    "completeness": "Are all required sections present?"
}

Evaluator Pattern:

from langchain_community.evaluation import load_evaluator

def create_quality_evaluator(aspect: str):
    """Create LLM-as-judge evaluator for a quality aspect."""
    return load_evaluator(
        evaluator=LabeledScoreStringEvalChain,
        criteria={aspect: QUALITY_DIMENSIONS[aspect]},
        llm=ChatOpenAI(model="gpt-4o-mini"),  # Cost-effective judge
        normalize_by=10  # Output 0.0-1.0 scores
    )

async def evaluate_output_quality(
    input_content: str,
    output_content: str
) -> dict[str, float]:
    """Evaluate output across all quality dimensions."""
    scores = {}
    for aspect in QUALITY_DIMENSIONS:
        evaluator = create_quality_evaluator(aspect)
        result = await evaluator.aevaluate_strings(
            input=input_content,
            prediction=output_content
        )
        scores[aspect] = result["score"]
    return scores

Quality Gate Integration:

QUALITY_THRESHOLD = 0.7  # Minimum acceptable score

async def quality_gate(state: dict) -> dict:
    """Block low-quality outputs with retry capability."""
    scores = await evaluate_output_quality(
        state["input"], state["output"]
    )
    avg_score = sum(scores.values()) / len(scores)

    return {
        "quality_scores": scores,
        "quality_passed": avg_score >= QUALITY_THRESHOLD
    }

9. Cost Optimization

Strategies:

• Use local models for CI/dev: Ollama with DeepSeek R1 70B, Qwen 2.5 Coder 32B (93% savings)
• Use smaller models for simple tasks (GPT-3.5 vs GPT-4, Haiku vs Sonnet)
• Implement prompt caching (Anthropic's ephemeral cache, Claude native cache)
• Batch requests when possible
• Set max_tokens to prevent runaway generation
• Monitor usage with alerts
• Use provider factory for automatic cloud/local switching

Token Counting:

import { encoding_for_model } from 'tiktoken'

function countTokens(text: string, model = 'gpt-4'): number {
  const encoder = encoding_for_model(model)
  const tokens = encoder.encode(text)
  encoder.free()
  return tokens.length
}

Detailed Implementation: See references/observability.md for:

• Cost estimation and budget tracking
• Model selection strategies
• Prompt caching patterns

10. Local LLM Inference with Ollama (v1.2.0)

Run LLMs locally for cost reduction, privacy, and offline development.

When to Use Local Models:

• CI/CD pipelines (93% cost reduction)
• Development and testing (no API costs)
• Privacy-sensitive data (no data leaves your machine)
• Offline development environments
• High-volume batch processing

Recommended Models (Apple Silicon M4 Max 256GB):

Task	Model	Size	Notes
Reasoning	deepseek-r1:70b	~42GB	GPT-4 level reasoning
Coding	qwen2.5-coder:32b	~35GB	73.7% Aider benchmark
Embeddings	nomic-embed-text	~0.5GB	768 dims, fast

LangChain Ollama Provider (v1.0.1):

from langchain_ollama import ChatOllama, OllamaEmbeddings

# Chat completion with keep_alive for CI performance
llm = ChatOllama(
    model="deepseek-r1:70b",
    base_url="http://localhost:11434",
    temperature=0.0,
    num_ctx=32768,          # Context window (Apple Silicon can handle large)
    keep_alive="5m",        # CRITICAL: Keep model loaded between calls
)

# Embeddings
embeddings = OllamaEmbeddings(
    model="nomic-embed-text",
    base_url="http://localhost:11434",
)

# Generate embedding
vector = await embeddings.aembed_query("Hello world")  # Single text
vectors = await embeddings.aembed_documents(["text1", "text2"])  # Batch

Tool Calling with Ollama:

from langchain_core.tools import tool
from pydantic import BaseModel, Field

@tool
def search_documents(query: str) -> str:
    """Search the document database."""
    return f"Found 5 documents for: {query}"

# Bind tools to model
llm_with_tools = llm.bind_tools([search_documents])

# Or with structured output
class CodeAnalysis(BaseModel):
    language: str = Field(description="Programming language")
    complexity: int = Field(ge=1, le=10, description="Complexity score")
    issues: list[str] = Field(description="Identified issues")

structured_llm = llm.with_structured_output(CodeAnalysis)
result = await structured_llm.ainvoke("Analyze this Python code...")

Streaming Responses:

async def stream_response(prompt: str):
    """Stream tokens as they're generated."""
    async for chunk in llm.astream(prompt):
        if hasattr(chunk, "content"):
            yield chunk.content

Provider Factory Pattern:

from app.shared.services.llm import get_llm_provider, get_embedding_provider

# Automatically uses Ollama if OLLAMA_ENABLED=true, else cloud API
llm = get_llm_provider(task_type="reasoning")  # deepseek-r1:70b or Gemini
llm = get_llm_provider(task_type="coding")     # qwen2.5-coder:32b or Claude
embedder = get_embedding_provider()             # nomic-embed-text or OpenAI

# Check availability
from app.shared.services.llm import is_ollama_available, get_available_ollama_models
if is_ollama_available():
    models = get_available_ollama_models()

Environment Configuration:

# Enable Ollama for local inference
export OLLAMA_ENABLED=true
export OLLAMA_HOST=http://localhost:11434
export OLLAMA_MODEL_REASONING=deepseek-r1:70b
export OLLAMA_MODEL_CODING=qwen2.5-coder:32b
export OLLAMA_MODEL_EMBED=nomic-embed-text

# Performance tuning for M4 Max
export OLLAMA_MAX_LOADED_MODELS=3    # Keep 3 models in memory
export OLLAMA_KEEP_ALIVE=5m          # 5 minute keep-alive

CI Integration (GitHub Actions):

# .github/workflows/evaluation.yml
jobs:
  evaluate:
    runs-on: self-hosted  # M4 Max 256GB runner
    env:
      OLLAMA_ENABLED: "true"
      OLLAMA_HOST: "http://localhost:11434"
      OLLAMA_MODEL_REASONING: "deepseek-r1:70b"
      OLLAMA_MODEL_CODING: "qwen2.5-coder:32b"
      OLLAMA_MODEL_EMBED: "nomic-embed-text"
    steps:
      - name: Ensure Ollama models ready
        run: |
          # Pre-warm embedding model for faster first call
          curl -s http://localhost:11434/api/embeddings \
            -d '{"model":"nomic-embed-text","prompt":"warmup"}' > /dev/null

Cost Comparison:

Provider	Monthly Cost	Latency	Privacy
Cloud APIs	~₪675/month	200-500ms	❌
Ollama Local	~₪50/month (electricity)	50-200ms	✅
Savings	93%	2-3x faster	Full control

Best Practices:

• ✅ Use keep_alive="5m" in CI to avoid cold starts
• ✅ Pre-warm models before first call in CI
• ✅ Set num_ctx=32768 on Apple Silicon (plenty of memory)
• ✅ Use factory pattern for cloud/local switching
• ✅ Run 3 models simultaneously on M4 Max 256GB
• ❌ Don't use keep_alive=-1 (keeps model forever, wastes memory)
• ❌ Don't skip pre-warming in CI (cold start adds 30-60s)

Production Implementation:

# app/services/llm_provider.py - Production-ready LLM factory
from app.config import settings

def get_llm_provider(task_type: str = "reasoning"):
    """Get LLM provider based on environment configuration."""
    if settings.OLLAMA_ENABLED:
        return get_ollama_provider(task_type)
    return get_cloud_provider(task_type)

def get_ollama_provider(task_type: str):
    """Configure Ollama for local inference."""
    from langchain_ollama import ChatOllama

    model_map = {
        "reasoning": "deepseek-r1:70b",
        "coding": "qwen2.5-coder:32b",
        "general": "llama3.3:70b"
    }

    return ChatOllama(
        model=model_map.get(task_type, "deepseek-r1:70b"),
        base_url=settings.OLLAMA_HOST,
        temperature=0.0,
        num_ctx=32768,
        keep_alive="5m"
    )

def get_cloud_provider(task_type: str):
    """Configure cloud LLM provider."""
    from langchain.chat_models import init_chat_model

    model_map = {
        "reasoning": "claude-opus-4-20250514",
        "coding": "claude-sonnet-4-20250514",
        "general": "gemini-2.0-flash-001"
    }

    return init_chat_model(
        model_map.get(task_type, "claude-sonnet-4-20250514"),
        temperature=0.0
    )

---

8. Observability & Monitoring

Track LLM performance, costs, and quality in production.

Tools:

• Langfuse: Open-source LLM observability, tracing, evaluation, monitoring
• Custom Logging: Structured logs with metrics

Key Metrics:

• Throughput (requests/minute)
• Latency (P50, P95, P99)
• Token usage and cost
• Error rate
• Quality scores (relevance, coherence, factuality)

Detailed Implementation: See references/observability.md for:

• Langfuse integration (self-hosted LLM observability)
• Custom logger implementation
• Performance monitoring
• Quality evaluation
• Debugging and error analysis

---

Searching References

This skill includes detailed reference material. Use grep to find specific patterns:

# Find RAG patterns
grep -r "RAG" references/

# Search for specific vector database
grep -A 10 "Pinecone Setup" references/vector-databases.md

# Find agentic workflow examples
grep -B 5 "ReAct Pattern" references/agentic-workflows.md

# Locate function calling patterns
grep -n "parallel.*tool" references/function-calling.md

# Search for cost optimization
grep -i "cost\|pricing\|budget" references/observability.md

# Find all code examples for embeddings
grep -A 20 "async function.*embedding" references/

---

Best Practices

Context Management

• ✅ Keep context windows under 75% of model limit
• ✅ Use sliding window for long conversations
• ✅ Summarize old messages before they scroll out
• ✅ Remove redundant or irrelevant context

Embedding Strategy

• ✅ Chunk documents to 500-1000 tokens
• ✅ Overlap chunks by 10-20% for continuity
• ✅ Include metadata (title, source, date) with chunks
• ✅ Re-embed when source data changes

RAG Quality

• ✅ Use hybrid search (semantic + keyword)
• ✅ Re-rank results for relevance
• ✅ Include citation/source in context
• ✅ Set temperature low (0.1-0.3) for factual answers
• ✅ Validate answers against retrieved context

Function Calling

• ✅ Provide clear, concise function descriptions
• ✅ Use strict JSON schema for parameters
• ✅ Handle missing or invalid parameters gracefully
• ✅ Limit to 10-20 tools to avoid confusion
• ✅ Validate function outputs before returning to LLM

Cost Optimization

• ✅ Use smaller models for simple tasks
• ✅ Implement prompt caching for repeated content
• ✅ Batch requests when possible
• ✅ Set max_tokens to prevent runaway generation
• ✅ Monitor usage with alerts for anomalies

Security

• ✅ Validate and sanitize user inputs
• ✅ Never include secrets in prompts
• ✅ Implement rate limiting
• ✅ Filter outputs for harmful content
• ✅ Use separate API keys per environment

---

Templates

Use the provided templates for common AI patterns:

• templates/rag-pipeline.ts - Basic RAG implementation
• templates/agentic-workflow.ts - ReAct agent pattern

---

Examples

Complete RAG Chatbot

See examples/chatbot-with-rag/ for a full-stack implementation:

• Vector database setup with document ingestion
• RAG query with citations
• Streaming chat interface
• Cost tracking and monitoring

---

Checklists

AI Implementation Checklist

See checklists/ai-implementation.md for comprehensive validation covering:

• Vector database setup and configuration
• Embedding generation and chunking strategy
• RAG pipeline with quality validation
• Function calling with error handling
• Streaming response implementation
• Cost monitoring and budget alerts
• Observability and logging
• Security and input validation

---

Common Patterns

Semantic Caching

Reduce costs by caching similar queries:

const cache = new Map<string, { embedding: number[]; response: string }>()

async function cachedRAG(query: string) {
  const queryEmbedding = await createEmbedding(query)

  // Check if similar query exists in cache
  for (const [cachedQuery, cached] of cache.entries()) {
    const similarity = cosineSimilarity(queryEmbedding, cached.embedding)
    if (similarity > 0.95) {
      return cached.response
    }
  }

  // Not cached, perform RAG
  const response = await ragQuery(query)
  cache.set(query, { embedding: queryEmbedding, response })
  return response
}

Conversational Memory

Maintain context across multiple turns:

interface ConversationMemory {
  messages: Message[] // Last 10 messages
  summary?: string // Summary of older messages
}

async function getConversationContext(userId: string): Promise<Message[]> {
  const memory = await db.memory.findUnique({ where: { userId } })

  return [
    { role: 'system', content: `Previous conversation summary: ${memory.summary}` },
    ...memory.messages.slice(-5) // Last 5 messages
  ]
}

---

Prompt Engineering

Few-Shot Learning

Provide examples to guide LLM behavior:

const fewShotExamples = `
Example 1:
Input: "I love this product!"
Sentiment: Positive

Example 2:
Input: "It's okay, nothing special"
Sentiment: Neutral
`

// Include in system prompt

Chain of Thought (CoT)

Ask LLM to show reasoning:

const prompt = `${problem}\n\nLet's think step by step:`

---

Resources

• OpenAI API Documentation
• Anthropic Claude API
• LangChain Documentation
• LangChain-Ollama - ChatOllama, OllamaEmbeddings
• Ollama Documentation - Local LLM inference
• Ollama Model Library - Available models
• Pinecone Documentation
• Chroma Documentation
• Langfuse Observability

---

Next Steps

After mastering AI-Native Development:

1. Explore Streaming API Patterns skill for real-time AI responses
2. Use Type Safety & Validation skill for AI input/output validation
3. Apply Edge Computing Patterns skill for global AI deployment
4. Reference Observability Patterns for production monitoring