世界上最简单最暴力的 RAG 系统

pip install auto-coder

easy_byzerllm deploy deepseek-chat --token _api_key

auto-coder doc serve \--model deepseek_chat \--doc_dir /Users/allwefantasy/projects/llm_friendly_packages/github.com/allwefantasy/byzer-sql

from typing import Any, Dict, List, Optional, Tuple, Generatorfrom autocoder.common import AutoCoderArgsfrom concurrent.futures import ThreadPoolExecutor, as_completedfrom byzerllm import ByzerLLMfrom loguru import loggerimport jsonimport os
import byzerllm

class LongContextRAG:def __init__(self, llm: ByzerLLM, args: AutoCoderArgs, path: str) -> None:self.llm = llmself.args = argsself.path = path
@byzerllm.prompt()def _check_relevance(self, query: str, document: str) -> str:"""请判断以下文档是否能够回答给出的问题。只需回答"是"或"否"。
问题：{{ query }}
文档：{{ document }}
回答："""
@byzerllm.prompt()def _answer_question(self, query: str, relevant_docs: List[str]) -> Generator[str, None, None]:"""使用以下文档来回答问题。如果文档中没有相关信息，请说"我没有足够的信息来回答这个问题"。
文档：{% for doc in relevant_docs %}{{ doc }}{% endfor %}
问题：{{ query }}
回答："""
def _retrieve_documents(self) -> List[str]:documents = []for root, dirs, files in os.walk(self.path):for file in files:if file.endswith(".md"):file_path = os.path.join(root, file)with open(file_path, "r", encoding="utf-8") as f:documents.append(f.read())return documents
def stream_chat_oai(self,conversations,model: Optional[str] = None,role_mapping=None,llm_config: Dict[str, Any] = {},):query = conversations[-1]["content"]documents = self._retrieve_documents()
with ThreadPoolExecutor(max_workers=self.args.index_filter_workers or 5) as executor:future_to_doc = {executor.submit(self._check_relevance.with_llm(self.llm).run, query, doc): docfor doc in documents}relevant_docs = []for future in as_completed(future_to_doc):try:doc = future_to_doc[future]if "是" in future.result().strip().lower():relevant_docs.append(doc)except Exception as exc:logger.error(f"Document processing generated an exception: {exc}")
if not relevant_docs:return ["没有找到相关的文档来回答这个问题。"], []else:chunks = self._answer_question.with_llm(self.llm).run(query, relevant_docs)return chunks, []