使用条件链构建智能 LLM 应用程序 - 深入探讨

from langchain.chains import LLMChain
from langchain.prompts import ChatPromptTemplate
from langchain.output_parsers import PydanticOutputParser
from pydantic import BaseModel, Field
from typing import Optional, List
import asyncio

class RouteDecision(BaseModel):
    route: str = Field(description="The selected processing route")
    confidence: float = Field(description="Confidence score of the decision")
    reasoning: str = Field(description="Explanation for the routing decision")

class IntelligentRouter:
    def __init__(self, routes: List[str]):
        self.routes = routes
        self.parser = PydanticOutputParser(pydantic_object=RouteDecision)
        self.route_prompt = ChatPromptTemplate.from_template(
            """Analyze the following input and decide the best processing route.
            Available routes: {routes}
            Input: {input}
            {format_instructions}
            """
        )

    async def decide_route(self, input_text: str) -> RouteDecision:
        prompt = self.route_prompt.format(
            routes=self.routes,
            input=input_text,
            format_instructions=self.parser.get_format_instructions()
        )

        chain = LLMChain(
            llm=self.llm,
            prompt=self.route_prompt
        )

        result = await chain.arun(input=input_text)
        return self.parser.parse(result)

class MultiLangProcessor:
    def __init__(self):
        self.router = IntelligentRouter([
            "translation",
            "summarization",
            "sentiment_analysis",
            "content_moderation"
        ])
        self.processors = {
            "translation": TranslationChain(),
            "summarization": SummaryChain(),
            "sentiment_analysis": SentimentChain(),
            "content_moderation": ModerationChain()
        }

    async def process(self, content: str) -> Dict:
        try:
            route = await self.router.decide_route(content)
            if route.confidence < 0.8:
                return await self.handle_low_confidence(content, route)

            processor = self.processors[route.route]
            result = await processor.run(content)
            return {
                "status": "success",
                "route": route.route,
                "result": result
            }
        except Exception as e:
            return await self.handle_error(e, content)

class ErrorHandler:
    def __init__(self):
        self.fallback_llm = ChatOpenAI(
            model_name="gpt-3.5-turbo",
            temperature=0.3
        )
        self.retry_limit = 3
        self.backoff_factor = 1.5

    async def handle_error(
        self, 
        error: Exception, 
        context: Dict
    ) -> Dict:
        error_type = type(error).__name__

        if error_type in self.error_strategies:
            return await self.error_strategies[error_type](
                error, context
            )

        return await self.default_error_handler(error, context)

    async def retry_with_backoff(
        self, 
        func, 
        *args, 
        **kwargs
    ):
        for attempt in range(self.retry_limit):
            try:
                return await func(*args, **kwargs)
            except Exception as e:
                if attempt == self.retry_limit - 1:
                    raise e
                await asyncio.sleep(
                    self.backoff_factor ** attempt
                )

class ModelFallbackChain:
    def __init__(self):
        self.models = [
            ChatOpenAI(model_name="gpt-4"),
            ChatOpenAI(model_name="gpt-3.5-turbo"),
            ChatOpenAI(model_name="gpt-3.5-turbo-16k")
        ]

    async def run_with_fallback(
        self, 
        prompt: str
    ) -> Optional[str]:
        for model in self.models:
            try:
                return await self.try_model(model, prompt)
            except Exception as e:
                continue

        return await self.final_fallback(prompt)

class ChunkingStrategy:
    def __init__(self, chunk_size: int = 1000):
        self.chunk_size = chunk_size

    def chunk_content(
        self, 
        content: str
    ) -> List[str]:
        # Implement smart content chunking
        return [
            content[i:i + self.chunk_size]
            for i in range(0, len(content), self.chunk_size)
        ]

    async def process_chunks(
        self, 
        chunks: List[str]
    ) -> List[Dict]:
        results = []
        for chunk in chunks:
            try:
                result = await self.process_single_chunk(chunk)
                results.append(result)
            except Exception as e:
                results.append(self.handle_chunk_error(e, chunk))
        return results