Création de systèmes d'agents d'entreprise : conception et optimisation des composants de base

Introduction

La création d'agents IA de niveau entreprise nécessite un examen attentif de la conception des composants, de l'architecture système et des pratiques d'ingénierie. Cet article explore les composants clés et les meilleures pratiques pour créer des systèmes d'agents robustes et évolutifs.

1. Ingénierie de modèles d'invite

1.1 Modèle de conception de modèle

from typing import Protocol, Dict
from jinja2 import Template

class PromptTemplate(Protocol):
    def render(self, **kwargs) -> str:
        pass

class JinjaPromptTemplate:
    def __init__(self, template_string: str):
        self.template = Template(template_string)

    def render(self, **kwargs) -> str:
        return self.template.render(**kwargs)

class PromptLibrary:
    def __init__(self):
        self.templates: Dict[str, PromptTemplate] = {}

    def register_template(self, name: str, template: PromptTemplate):
        self.templates[name] = template

    def get_template(self, name: str) -> PromptTemplate:
        return self.templates[name]

1.2 Contrôle de version et tests

class PromptVersion:
    def __init__(self, version: str, template: str, metadata: dict):
        self.version = version
        self.template = template
        self.metadata = metadata
        self.test_cases = []

    def add_test_case(self, inputs: dict, expected_output: str):
        self.test_cases.append((inputs, expected_output))

    def validate(self) -> bool:
        template = JinjaPromptTemplate(self.template)
        for inputs, expected in self.test_cases:
            result = template.render(**inputs)
            if not self._validate_output(result, expected):
                return False
        return True

2. Système de mémoire hiérarchique

2.1 Architecture de la mémoire

from typing import Any, List
from datetime import datetime

class MemoryEntry:
    def __init__(self, content: Any, importance: float):
        self.content = content
        self.importance = importance
        self.timestamp = datetime.now()
        self.access_count = 0

class MemoryLayer:
    def __init__(self, capacity: int):
        self.capacity = capacity
        self.memories: List[MemoryEntry] = []

    def add(self, entry: MemoryEntry):
        if len(self.memories) >= self.capacity:
            self._evict()
        self.memories.append(entry)

    def _evict(self):
        # Implement memory eviction strategy
        self.memories.sort(key=lambda x: x.importance * x.access_count)
        self.memories.pop(0)

class HierarchicalMemory:
    def __init__(self):
        self.working_memory = MemoryLayer(capacity=5)
        self.short_term = MemoryLayer(capacity=50)
        self.long_term = MemoryLayer(capacity=1000)

    def store(self, content: Any, importance: float):
        entry = MemoryEntry(content, importance)

        if importance > 0.8:
            self.working_memory.add(entry)
        elif importance > 0.5:
            self.short_term.add(entry)
        else:
            self.long_term.add(entry)

2.2 Récupération et indexation de la mémoire

from typing import List, Tuple
import numpy as np
from sklearn.metrics.pairwise import cosine_similarity

class MemoryIndex:
    def __init__(self, embedding_model):
        self.embedding_model = embedding_model
        self.embeddings = []
        self.memories = []

    def add(self, memory: MemoryEntry):
        embedding = self.embedding_model.embed(memory.content)
        self.embeddings.append(embedding)
        self.memories.append(memory)

    def search(self, query: str, k: int = 5) -> List[Tuple[MemoryEntry, float]]:
        query_embedding = self.embedding_model.embed(query)
        similarities = cosine_similarity(
            [query_embedding], 
            self.embeddings
        )[0]

        top_k_indices = np.argsort(similarities)[-k:]

        return [
            (self.memories[i], similarities[i]) 
            for i in top_k_indices
        ]

3. Chaînes de raisonnement observables

3.1 Structure de la chaîne

from typing import List, Optional
from dataclasses import dataclass
import uuid

@dataclass
class ThoughtNode:
    content: str
    confidence: float
    supporting_evidence: List[str]

class ReasoningChain:
    def __init__(self):
        self.chain_id = str(uuid.uuid4())
        self.nodes: List[ThoughtNode] = []
        self.metadata = {}

    def add_thought(self, thought: ThoughtNode):
        self.nodes.append(thought)

    def get_path(self) -> List[str]:
        return [node.content for node in self.nodes]

    def get_confidence(self) -> float:
        if not self.nodes:
            return 0.0
        return sum(n.confidence for n in self.nodes) / len(self.nodes)

3.2 Surveillance et analyse de la chaîne

import logging
from opentelemetry import trace
from prometheus_client import Histogram

reasoning_time = Histogram(
    'reasoning_chain_duration_seconds',
    'Time spent in reasoning chain'
)

class ChainMonitor:
    def __init__(self):
        self.tracer = trace.get_tracer(__name__)

    def monitor_chain(self, chain: ReasoningChain):
        with self.tracer.start_as_current_span("reasoning_chain") as span:
            span.set_attribute("chain_id", chain.chain_id)

            with reasoning_time.time():
                for node in chain.nodes:
                    with self.tracer.start_span("thought") as thought_span:
                        thought_span.set_attribute(
                            "confidence", 
                            node.confidence
                        )
                        logging.info(
                            f"Thought: {node.content} "
                            f"(confidence: {node.confidence})"
                        )

4. Découplage et réutilisation des composants

4.1 Conception de l'interface

from abc import ABC, abstractmethod
from typing import Generic, TypeVar

T = TypeVar('T')

class Component(ABC, Generic[T]):
    @abstractmethod
    def process(self, input_data: T) -> T:
        pass

class Pipeline:
    def __init__(self):
        self.components: List[Component] = []

    def add_component(self, component: Component):
        self.components.append(component)

    def process(self, input_data: Any) -> Any:
        result = input_data
        for component in self.components:
            result = component.process(result)
        return result

4.2 Registre des composants

class ComponentRegistry:
    _instance = None

    def __new__(cls):
        if cls._instance is None:
            cls._instance = super().__new__(cls)
            cls._instance.components = {}
        return cls._instance

    def register(self, name: str, component: Component):
        self.components[name] = component

    def get(self, name: str) -> Optional[Component]:
        return self.components.get(name)

    def create_pipeline(self, component_names: List[str]) -> Pipeline:
        pipeline = Pipeline()
        for name in component_names:
            component = self.get(name)
            if component:
                pipeline.add_component(component)
        return pipeline

5. Surveillance et optimisation des performances

5.1 Mesures de performances

from dataclasses import dataclass
from typing import Dict
import time

@dataclass
class PerformanceMetrics:
    latency: float
    memory_usage: float
    token_count: int
    success_rate: float

class PerformanceMonitor:
    def __init__(self):
        self.metrics: Dict[str, List[PerformanceMetrics]] = {}

    def record_operation(
        self,
        operation_name: str,
        metrics: PerformanceMetrics
    ):
        if operation_name not in self.metrics:
            self.metrics[operation_name] = []
        self.metrics[operation_name].append(metrics)

    def get_average_metrics(
        self,
        operation_name: str
    ) -> Optional[PerformanceMetrics]:
        if operation_name not in self.metrics:
            return None

        metrics_list = self.metrics[operation_name]
        return PerformanceMetrics(
            latency=sum(m.latency for m in metrics_list) / len(metrics_list),
            memory_usage=sum(m.memory_usage for m in metrics_list) / len(metrics_list),
            token_count=sum(m.token_count for m in metrics_list) / len(metrics_list),
            success_rate=sum(m.success_rate for m in metrics_list) / len(metrics_list)
        )

5.2 Stratégies d'optimisation

class PerformanceOptimizer:
    def __init__(self, monitor: PerformanceMonitor):
        self.monitor = monitor
        self.thresholds = {
            'latency': 1.0,  # seconds
            'memory_usage': 512,  # MB
            'token_count': 1000,
            'success_rate': 0.95
        }

    def analyze_performance(self, operation_name: str) -> List[str]:
        metrics = self.monitor.get_average_metrics(operation_name)
        if not metrics:
            return []

        recommendations = []

        if metrics.latency > self.thresholds['latency']:
            recommendations.append(
                "Consider implementing caching or parallel processing"
            )

        if metrics.memory_usage > self.thresholds['memory_usage']:
            recommendations.append(
                "Optimize memory usage through batch processing"
            )

        if metrics.token_count > self.thresholds['token_count']:
            recommendations.append(
                "Implement prompt optimization to reduce token usage"
            )

        if metrics.success_rate < self.thresholds['success_rate']:
            recommendations.append(
                "Review error handling and implement retry mechanisms"
            )

        return recommendations

Conclusion

La création de systèmes d'agents de niveau entreprise nécessite une attention particulière aux points suivants :

• Gestion structurée des invites et contrôle de version
• Systèmes de mémoire efficaces et évolutifs
• Processus de raisonnement observables et traçables
• Conception de composants modulaires et réutilisables
• Surveillance et optimisation complètes des performances

Ce qui précède est le contenu détaillé de. pour plus d'informations, suivez d'autres articles connexes sur le site Web de PHP en chinois!