RAG开源项目Qanything源码阅读3-在线推理

最新推荐文章于 2025-05-21 01:46:59 发布

去追风，去看海

最新推荐文章于 2025-05-21 01:46:59 发布

阅读量1k

点赞数 18

CC 4.0 BY-SA版权

分类专栏： NLP 文章标签： nlp

本文链接：https://blog.youkuaiyun.com/weixin_40959890/article/details/140147214

NLP 专栏收录该内容

104 篇文章

订阅专栏

原文：前沿重器[47] | RAG开源项目Qanything源码阅读3-在线推理
项目：https://github.com/netease-youdao/QAnything
第一篇：RAG开源项目Qanything源码阅读1-概述+服务
第二篇：RAG开源项目Qanything源码阅读2-离线文件处理

0，推理大概流程

检索&粗排
精排
检索文档后处理
prompt和请求大模型

1，外部服务

回顾一下在“前沿重器[45] RAG开源项目Qanything源码阅读1-概述+服务”中提到的服务核心文件，所有的接口都是在qanything_kernel\qanything_server\sanic_api.py里面启动的：

app.add_route(document, "/api/docs", methods=['GET'])
app.add_route(new_knowledge_base, "/api/local_doc_qa/new_knowledge_base", methods=['POST'])  # tags=["新建知识库"]
app.add_route(upload_weblink, "/api/local_doc_qa/upload_weblink", methods=['POST'])  # tags=["上传网页链接"]
app.add_route(upload_files, "/api/local_doc_qa/upload_files", methods=['POST'])  # tags=["上传文件"] 
app.add_route(local_doc_chat, "/api/local_doc_qa/local_doc_chat", methods=['POST'])  # tags=["问答接口"] 
app.add_route(list_kbs, "/api/local_doc_qa/list_knowledge_base", methods=['POST'])  # tags=["知识库列表"] 
app.add_route(list_docs, "/api/local_doc_qa/list_files", methods=['POST'])  # tags=["文件列表"]
app.add_route(get_total_status, "/api/local_doc_qa/get_total_status", methods=['POST'])  # tags=["获取所有知识库状态"]
app.add_route(clean_files_by_status, "/api/local_doc_qa/clean_files_by_status", methods=['POST'])  # tags=["清理数据库"]
app.add_route(delete_docs, "/api/local_doc_qa/delete_files", methods=['POST'])  # tags=["删除文件"] 
app.add_route(delete_knowledge_base, "/api/local_doc_qa/delete_knowledge_base", methods=['POST'])  # tags=["删除知识库"] 
app.add_route(rename_knowledge_base, "/api/local_doc_qa/rename_knowledge_base", methods=['POST'])  # tags=["重命名知识库"]

而推理，就是这里的local_doc_chat，直接看这个函数，就在qanything_kernel\qanything_server\handler.py里面。

async def local_doc_chat(req: request):
    local_doc_qa: LocalDocQA = req.app.ctx.local_doc_qa
    user_id = safe_get(req, 'user_id')
    if user_id is None:
        return sanic_json({"code": 2002, "msg": f'输入非法！request.json：{req.json}，请检查！'})
    is_valid = validate_user_id(user_id)
    if not is_valid:
        return sanic_json({"code": 2005, "msg": get_invalid_user_id_msg(user_id=user_id)})
    debug_logger.info('local_doc_chat %s', user_id)
    kb_ids = safe_get(req, 'kb_ids')
    question = safe_get(req, 'question')
    rerank = safe_get(req, 'rerank', default=True)
    debug_logger.info('rerank %s', rerank)
    streaming = safe_get(req, 'streaming', False)
    history = safe_get(req, 'history', [])
    debug_logger.info("history: %s ", history)
    debug_logger.info("question: %s", question)
    debug_logger.info("kb_ids: %s", kb_ids)
    debug_logger.info("user_id: %s", user_id)
 
    not_exist_kb_ids = local_doc_qa.milvus_summary.check_kb_exist(user_id, kb_ids)
    if not_exist_kb_ids:
        return sanic_json({"code": 2003, "msg": "fail, knowledge Base {} not found".format(not_exist_kb_ids)})
 
    file_infos = []
    milvus_kb = local_doc_qa.match_milvus_kb(user_id, kb_ids)
    for kb_id in kb_ids:
        file_infos.extend(local_doc_qa.milvus_summary.get_files(user_id, kb_id))
    valid_files = [fi for fi in file_infos if fi[2] == 'green']
    if len(valid_files) == 0:
        return sanic_json({"code": 200, "msg": "当前知识库为空，请上传文件或等待文件解析完毕", "question": question,
                           "response": "All knowledge bases {} are empty or haven't green file, please upload files".format(
                               kb_ids), "history": history, "source_documents": [{}]})
    else:
        debug_logger.info("streaming: %s", streaming)
        if streaming:
            debug_logger.info("start generate answer")
 
            async def generate_answer(response):
                debug_logger.info("start generate...")
                for resp, next_history in local_doc_qa.get_knowledge_based_answer(
                        query=question, milvus_kb=milvus_kb, chat_history=history, streaming=True, rerank=rerank
                ):
                    chunk_data = resp["result"]
                    if not chunk_data:
                        continue
                    chunk_str = chunk_data[6:]
                    if chunk_str.startswith("[DONE]"):
                        source_documents = []
                        for inum, doc in enumerate(resp["source_documents"]):
                            source_info = {'file_id': doc.metadata['file_id'],
                                           'file_name': doc.metadata['file_name'],
                                           'content': doc.page_content,
                                           'retrieval_query': doc.metadata['retrieval_query'],
                                           'score': str(doc.metadata['score'])}
                            source_documents.append(source_info)
 
                        retrieval_documents = format_source_documents(resp["retrieval_documents"])
                        source_documents = format_source_documents(resp["source_documents"])
                        chat_data = {'user_info': user_id, 'kb_ids': kb_ids, 'query': question, 'history': history,
                                     'prompt': resp['prompt'], 'result': next_history[-1][1],
                                     'retrieval_documents': retrieval_documents, 'source_documents': source_documents}
                        qa_logger.info("chat_data: %s", chat_data)
                        debug_logger.info("response: %s", chat_data['result'])
                        stream_res = {
                            "code": 200,
                            "msg": "success",
                            "question": question,
                            # "response":next_history[-1][1],
                            "response": "",
                            "history": next_history,
                            "source_documents": source_documents,
                        }
                    else:
                        chunk_js = json.loads(chunk_str)
                        delta_answer = chunk_js["answer"]
                        stream_res = {
                            "code": 200,
                            "msg": "success",
                            "question": "",
                            "response": delta_answer,
                            "history": [],
                            "source_documents": [],
                        }
                    await response.write(f"data: {json.dumps(stream_res, ensure_ascii=False)}\n\n")
                    if chunk_str.startswith("[DONE]"):
                        await response.eof()
                    await asyncio.sleep(0.001)
 
            response_stream = ResponseStream(generate_answer, content_type='text/event-stream')
            return response_stream
 
        else:
            for resp, history in local_doc_qa.get_knowledge_based_answer(
                    query=question, milvus_kb=milvus_kb, chat_history=history, streaming=False, rerank=rerank
            ):
                pass
            retrieval_documents = format_source_documents(resp["retrieval_documents"])
            source_documents = format_source_documents(resp["source_documents"])
            chat_data = {'user_id': user_id, 'kb_ids': kb_ids, 'query': question, 'history': history,
                         'retrieval_documents': retrieval_documents, 'prompt': resp['prompt'], 'result': resp['result'],
                         '`': source_documents}
            qa_logger.info("chat_data: %s", chat_data)
            debug_logger.info("response: %s", chat_data['result'])
            return sanic_json({"code": 200, "msg": "success chat", "question": question, "response": resp["result"],
                               "history": history, "source_documents": source_documents})

上面代码的重点内容：

首先因为是正式项目，在鉴权、数据库检测上都做了很多健壮性的处理，例如，对user_id的判别、对数据库及其对应用户的权限判别check_kb_exist，再者还有知识库的判空等。
此处有区分是否使用了流式streaming。
最终结果的输出有进行结构化，结构化这事的业务代码专门弄了个函数format_source_documents。
这里区分了retrieval_documents和source_documents，两者有所区别，在后面展开聊关键算法流程的时候会展开讲。
get_knowledge_based_answer是内部获取知识点并进行生成的关键函数，就是上一条所说的关键算法流程。

# qanything_kernel\utils\general_utils.py 

def format_source_documents(ori_source_documents):
    source_documents = []
    for inum, doc in enumerate(ori_source_documents):
        # for inum, doc in enumerate(answer_source_documents):
        # doc_source = doc.metadata['source']
        file_id = doc.metadata['file_id']
        file_name = doc.metadata['file_name']
        # source_str = doc_source if isURL(doc_source) else os.path.split(doc_source)[-1]
        source_info = {'file_id': doc.metadata['file_id'],
                       'file_name': doc.metadata['file_name'],
                       'content': doc.page_content,
                       'retrieval_query': doc.metadata['retrieval_query'],
                       'kernel': doc.metadata['kernel'],
                       'score': str(doc.metadata['score']),
                       'embed_version': doc.metadata['embed_version']}
        source_documents.append(source_info)
    return source_documents

2，RAG推理流程

get_knowledge_based_answer的函数很简单，不过单独拿出来，对可读性是有挺大帮助的。
RAG说白了就是先搜后交给大模型生成，终于讲到这段代码了，流程在这里qanything_kernel\core\local_doc_qa.py。

   
# qanything_kernel\core\local_doc_qa.py
@get_time
    def get_knowledge_based_answer(self, query, milvus_kb, chat_history=None, streaming: bool = STREAMING,
                                   rerank: bool = False):
        if chat_history is None:
            chat_history = []
        retrieval_queries = [query]

        source_documents = self.get_source_documents(retrieval_queries, milvus_kb)

        deduplicated_docs = self.deduplicate_documents(source_documents)
        retrieval_documents = sorted(deduplicated_docs, key=lambda x: x.metadata['score'], reverse=True)
        if rerank and len(retrieval_documents) > 1:
            debug_logger.info(f"use rerank, rerank docs num: {len(retrieval_documents)}")
            retrieval_documents = self.rerank_documents(query, retrieval_documents)

        source_documents = self.reprocess_source_documents(query=query,
                                                           source_docs=retrieval_documents,
                                                           history=chat_history,
                                                           prompt_template=PROMPT_TEMPLATE)
        prompt = self.generate_prompt(query=query,
                                      source_docs=source_documents,
                                      prompt_template=PROMPT_TEMPLATE)
        t1 = time.time()
        for answer_result in self.llm.generatorAnswer(prompt=prompt,
                                                      history=chat_history,
                                                      streaming=streaming):
            resp = answer_result.llm_output["answer"]
            prompt = answer_result.prompt
            history = answer_result.history

            # logging.info(f"[debug] get_knowledge_based_answer history = {history}")
            history[-1][0] = query
            response = {"query": query,
                        "prompt": prompt,
                        "result": resp,
                        "retrieval_documents": retrieval_documents,
                        "source_documents": source_documents}
            yield response, history
        t2 = time.time()
        debug_logger.info(f"LLM time: {t2 - t1}")

首先注意到这里有个装饰器@get_time。可以用来记录执行的时间。

def get_time(func):
    def inner(*arg, **kwargs):
        s_time = time.time()
        res = func(*arg, **kwargs)
        e_time = time.time()
        print('函数 {} 执行耗时: {} 秒'.format(func.__name__, e_time - s_time))
        return res

    return inner

2.1 检索&粗排

get_source_documents是检索的过程，即给定了retrieval_queries和milvus_kb，即query和所需要查的数据库，开始进行查询。这个的返回结果，会放在retrieval_documents里面，即**“检索到的文档”**，下面是源码。

def get_source_documents(self, queries, milvus_kb, cosine_thresh=None, top_k=None):
    milvus_kb: MilvusClient
    if not top_k:
        top_k = self.top_k
    source_documents = []
    embs = self.embeddings._get_len_safe_embeddings(queries)
    t1 = time.time()
    batch_result = milvus_kb.search_emb_async(embs=embs, top_k=top_k, queries=queries)
    t2 = time.time()
    debug_logger.info(f"milvus search time: {t2 - t1}")
    for query, query_docs in zip(queries, batch_result):
        for doc in query_docs:
            doc.metadata['retrieval_query'] = query  # 添加查询到文档的元数据中
            doc.metadata['embed_version'] = self.embeddings.embed_version
            source_documents.append(doc)
    if cosine_thresh:
        source_documents = [item for item in source_documents if float(item.metadata['score']) > cosine_thresh]
 
    return source_documents

_get_len_safe_embeddings给定query获取向量。在上一期RAG开源项目Qanything源码阅读2-离线文件处理有讲过，这个内部是请求一个向量模型的服务，背后的模型是需要和离线文件处理那个模型一致，所以部署同一个就会比较稳当，当然的，接口也是triton，一个grpc接口，有关GRPC，上次忘了放链接，这次放这里心法利器[6] | python grpc实践，非常建议大家详细了解并且学会。
search_emb_async是用于做向量检索的。这个就是pymilvus的核心功能了。
此处，查询出来还要过一个阈值卡控，对相似度达不到阈值的文档，需要过滤，阈值设置在cosine_thresh。

_get_len_safe_embeddings 使用的embedding 代码（可跳过，继续回到 `get_knowledge_based_answer`）

# qanything_kernel\connector\embedding\embedding_for_local.py
"""Wrapper around YouDao embedding models."""
from typing import List

from qanything_kernel.connector.embedding.embedding_client import EmbeddingClient
from qanything_kernel.configs.model_config import LOCAL_EMBED_SERVICE_URL, LOCAL_EMBED_MODEL_NAME, LOCAL_EMBED_MAX_LENGTH, LOCAL_EMBED_BATCH
from qanything_kernel.utils.custom_log import debug_logger
import concurrent.futures
from tqdm import tqdm 

embedding_client = EmbeddingClient(
    server_url=LOCAL_EMBED_SERVICE_URL,
    model_name=LOCAL_EMBED_MODEL_NAME,
    model_version='1',
    resp_wait_s=120,
    tokenizer_path='qanything_kernel/connector/embedding/embedding_model_0630')


class YouDaoLocalEmbeddings:
    def __init__(self):
        pass

    def _get_embedding(self, queries):
        embeddings = embedding_client.get_embedding(queries, max_length=LOCAL_EMBED_MAX_LENGTH)
        return embeddings

    def _get_len_safe_embeddings(self, texts: List[str]) -> List[List[float]]:
        all_embeddings = []
        batch_size = LOCAL_EMBED_BATCH

        with concurrent.futures.ThreadPoolExecutor() as executor:
            futures = []
            for i in range(0, len(texts), batch_size):
                batch = texts[i:i + batch_size]
                future = executor.submit(self._get_embedding, batch)
                futures.append(future)
            debug_logger.info(f'embedding number: {len(futures)}')
            for future in tqdm(futures):
                embeddings = future.result()
                all_embeddings += embeddings
        return all_embeddings

    @property
    def embed_version(self):
        return embedding_client.getModelVersion()

回到 get_knowledge_based_answer
留意到 qanything_kernel\core\local_doc_qa.py 文件里的 get_knowledge_based_answer 里这一串代码：

retrieval_documents = sorted(deduplicated_docs, key=lambda x: x.metadata['score'], reverse=True)
if rerank and len(retrieval_documents) > 1:
    debug_logger.info(f"use rerank, rerank docs num: {len(retrieval_documents)}")
    retrieval_documents = self.rerank_documents(query, retrieval_documents)

此处注意，这里的检索还涉及一个过程“粗排”（上面第一行代码），这个粗排是指查询数据库的时候，需要根据相似度进行排序，只取TOPN，毕竟如果不进行这个TOP的卡控，那数据库里所有的数据都会被查出来，这没什么意义了。这里之所以叫粗排，是因为这种相似度的对比是比较粗略的，只能过滤掉“肯定不是”的那些无关结果。具体“哪个好”，用额外的、更精准的模型来做会更好，达到“优中取优”的目的。

2.2 检索&粗排

继续关注这里的 qanything_kernel\core\local_doc_qa.py 的 get_knowledge_based_answer里调用的 rerank_documents，这个就是精排，或者像这里说的重排。

def rerank_documents(self, query, source_documents):
    return self.rerank_documents_for_local(query, source_documents)
 
def rerank_documents_for_local(self, query, source_documents):
    if len(query) > 300:  # tokens数量超过300时不使用local rerank
        return source_documents
 
    source_documents_reranked = []
    try:
        response = requests.post(f"{self.local_rerank_service_url}/rerank",
                                    json={"passages": [doc.page_content for doc in source_documents], "query": query})
        scores = response.json()
        for idx, score in enumerate(scores):
            source_documents[idx].metadata['score'] = score
            if score < 0.35 and len(source_documents_reranked) > 0:
                continue
            source_documents_reranked.append(source_documents[idx])
 
        source_documents_reranked = sorted(source_documents_reranked, key=lambda x: x.metadata['score'], reverse=True)
    except Exception as e:
        debug_logger.error("rerank error: %s", traceback.format_exc())
        debug_logger.warning("rerank error, use origin retrieval docs")
        source_documents_reranked = sorted(source_documents, key=lambda x: x.metadata['score'], reverse=True)
 
    return source_documents_reranked

简单地，这里就是把所有召回回来的文章请求到重排服务来算分，根据算分来进行过滤和排序，筛选出最优的文章。和向量模型类似，一样是用triton部署的，看模型名像是QAEnsemble_embed_rerank。

2.3 检索文档后处理

更进一步，需要对文档进行后处理，即reprocess_source_documents函数。qanything_kernel\core\local_doc_qa.py

#source_documents = self.reprocess_source_documents(query=query,
#                                                           source_docs=retrieval_documents,
#                                                           history=chat_history,
#                                                           prompt_template=PROMPT_TEMPLATE)

def reprocess_source_documents(self, query: str,
                                source_docs: List[Document],
                                history: List[str],
                                prompt_template: str) -> List[Document]:
    # 组装prompt,根据max_token
    query_token_num = self.llm.num_tokens_from_messages([query])
    history_token_num = self.llm.num_tokens_from_messages([x for sublist in history for x in sublist])
    template_token_num = self.llm.num_tokens_from_messages([prompt_template])
 
    # logging.info(f"<self.llm.token_window, self.llm.max_token, self.llm.offcut_token, query_token_num, history_token_num, template_token_num>, types = {type(self.llm.token_window), type(self.llm.max_token), type(self.llm.offcut_token), type(query_token_num), type(history_token_num), type(template_token_num)}, values = {query_token_num, history_token_num, template_token_num}")
    limited_token_nums = self.llm.token_window - self.llm.max_token - self.llm.offcut_token - query_token_num - history_token_num - template_token_num
    new_source_docs = []
    total_token_num = 0
    for doc in source_docs:
        doc_token_num = self.llm.num_tokens_from_docs([doc])
        if total_token_num + doc_token_num <= limited_token_nums:
            new_source_docs.append(doc)
            total_token_num += doc_token_num
        else:
            remaining_token_num = limited_token_nums - total_token_num
            doc_content = doc.page_content
            doc_content_token_num = self.llm.num_tokens_from_messages([doc_content])
            while doc_content_token_num > remaining_token_num:
                # Truncate the doc content to fit the remaining tokens
                if len(doc_content) > 2 * self.llm.truncate_len:
                    doc_content = doc_content[self.llm.truncate_len: -self.llm.truncate_len]
                else:  # 如果最后不够truncate_len长度的2倍，说明不够切了，直接赋值为空
                    doc_content = ""
                    break
                doc_content_token_num = self.llm.num_tokens_from_messages([doc_content])
            doc.page_content = doc_content
            new_source_docs.append(doc)
            break
 
    debug_logger.info(f"limited token nums: {limited_token_nums}")
    debug_logger.info(f"template token nums: {template_token_num}")
    debug_logger.info(f"query token nums: {query_token_num}")
    debug_logger.info(f"history token nums: {history_token_num}")
    debug_logger.info(f"new_source_docs token nums: {self.llm.num_tokens_from_docs(new_source_docs)}")
    return new_source_docs

这里的llm，是一个自己封装好的大模型工具，具体是在qanything_kernel\connector\llm\llm_for_fastchat.py这个位置。里面支持计算token、请求大模型等通用功能。这个工具可以结合自己场景的需求搬过去直接使用。
计算limited_token_nums主要是方便组装prompt，避免某些文字被吃掉。
这里是需要对文档进行新的拼接和调整，如果查询的文档太多太长，则需要截断，且截断的时候需要注意，要保证截断的位置必须是完整地句子，如果不够长直接不切了。

2.4 prompt和请求大模型

然后就是开始生成prompt，generate_prompt。说白了就是一个简单的拼接。另外，这里的prompt拼接，更多使用replace来完成，之前有看过别的模式，例如用字符串的format应该也可以，不过replace的适用范围会更广一些。

def generate_prompt(self, query, source_docs, prompt_template):
    context = "\n".join([doc.page_content for doc in source_docs])
    prompt = prompt_template.replace("{question}", query).replace("{context}", context)
    return prompt

顺带就看看他们的prompt吧，实际上并不复杂。

PROMPT_TEMPLATE = """参考信息：
{context}
---
我的问题或指令：
{question}
---
请根据上述参考信息回答我的问题或回复我的指令。前面的参考信息可能有用，也可能没用，你需要从我给出的参考信息中选出与我的问题最相关的那些，来为你的回答提供依据。回答一定要忠于原文，简洁但不丢信息，不要胡乱编造。我的问题或指令是什么语种，你就用什么语种回复,
你的回复："""

最后一步就是开始请求大模型了。即generatorAnswer函数。

def generatorAnswer(self, prompt: str,
                    history: List[List[str]] = [],
                    streaming: bool = False) -> AnswerResult:
 
    if history is None or len(history) == 0:
        history = [[]]
    logging.info(f"history_len: {self.history_len}")
    logging.info(f"prompt: {prompt}")
    logging.info(f"prompt tokens: {self.num_tokens_from_messages([{'content': prompt}])}")
    logging.info(f"streaming: {streaming}")
            
    response = self._call(prompt, history[:-1], streaming)
    complete_answer = ""
    for response_text in response:
 
        if response_text:
            chunk_str = response_text[6:]
            if not chunk_str.startswith("[DONE]"):
                chunk_js = json.loads(chunk_str)
                complete_answer += chunk_js["answer"]
                
        history[-1] = [prompt, complete_answer]
        answer_result = AnswerResult()
        answer_result.history = history
        answer_result.llm_output = {"answer": response_text}
        answer_result.prompt = prompt
        yield answer_result

这里就是请求大模型的基本话术了，相对还是比较简单的，一方面是请求大模型，另一方面是解析大模型内的结果。有留意到，这里有对内容做一些校验：

if response_text:
    chunk_str = response_text[6:]
    if not chunk_str.startswith("[DONE]"):
        chunk_js = json.loads(chunk_str)
        complete_answer += chunk_js["answer"]

可以看出应该是有一些泛用性处理，能解决更多复杂的问题吧。

小结

本文把QAnything项目内的重要的推理部分穿讲了一遍，可以看出这个项目已经非常完成，基本具备上线所需的关键部分，同时也有很严格的校验逻辑，严格程度很高也比较稳定，经过这个学习，自己对工程代码和具体实施的理解有了很大的提升，希望大家也有收获。当然有空再复习一遍应该有更大收获。

QAnything在服务的完整性、健壮性，以及文档处理上都有了很多的更新，但都不要指望用上就能达到很高的水准，需要进一步提升还需要更多内里的修炼：

query理解辅助更好地提升检索的准确性
联合训练提升大模型和检索结果的协同
更深入定制的文档处理提升内容的可读性等

补充

`qanything_kernel\connector\llm\llm_for_fastchat.py`

from abc import ABC
import tiktoken
import os
from dotenv import load_dotenv
from openai import OpenAI
from typing import Optional, List
import sys
import json
import requests
import logging
sys.path.append("../../../")
from qanything_kernel.connector.llm.base import (BaseAnswer, AnswerResult)
from qanything_kernel.configs.model_config import LOCAL_LLM_SERVICE_URL, LOCAL_LLM_MODEL_NAME, LOCAL_LLM_MAX_LENGTH

load_dotenv()

logging.basicConfig(level=logging.INFO)

class OpenAICustomLLM(BaseAnswer, ABC):
    model: str = LOCAL_LLM_MODEL_NAME
    token_window: int = LOCAL_LLM_MAX_LENGTH
    max_token: int = 512
    offcut_token: int = 50
    truncate_len: int = 50
    temperature: float = 0
    stop_words: str = None
    history: List[List[str]] = []
    history_len: int = 2

    def __init__(self):
        super().__init__()
        # self.client = OpenAI(base_url="http://localhost:7802/v1", api_key="EMPTY")
        if LOCAL_LLM_SERVICE_URL.startswith("http://"):
            base_url = f"{LOCAL_LLM_SERVICE_URL}/v1" 
        else:
            base_url = f"http://{LOCAL_LLM_SERVICE_URL}/v1" 
        self.client = OpenAI(base_url=base_url, api_key="EMPTY")

    @property
    def _llm_type(self) -> str:
        return "CustomLLM using FastChat w/ huggingface transformers or vllm backend"

    @property
    def _history_len(self) -> int:
        return self.history_len

    def set_history_len(self, history_len: int = 10) -> None:
        self.history_len = history_len

    def token_check(self, query: str) -> int:
        
        if LOCAL_LLM_SERVICE_URL.startswith("http://"):
            base_url = f"{LOCAL_LLM_SERVICE_URL}/api/v1/token_check" 
        else:
            base_url = f"http://{LOCAL_LLM_SERVICE_URL}/api/v1/token_check" 

        headers = {"Content-Type": "application/json"}
        
        response = requests.post(
            base_url, 
            data=json.dumps(
                {'prompts': [{'model': self.model, 'prompt': query, 'max_tokens': self.max_token}]}
            ),
            headers=headers, timeout=60)

        # {'prompts': [{'fits': True, 'tokenCount': 317, 'contextLength': 8192}]}
        result = response.json()
        token_num = 0
        try:
            token_num = result['prompts'][0]['tokenCount']
            return token_num
        except Exception as e:
            logging.error(f"token_check Exception {base_url} w/ {e}")
            return token_num

    def num_tokens_from_messages(self, message_texts):
        num_tokens = 0
        for message in message_texts:
            num_tokens += self.token_check(message)
        return num_tokens

    def num_tokens_from_docs(self, docs):
        num_tokens = 0
        for doc in docs:
            num_tokens += self.token_check(doc.page_content)
        return num_tokens

    def _call(self, prompt: str, history: List[List[str]], streaming: bool=False) -> str:
        messages = []
        for pair in history:
            question, answer = pair
            messages.append({"role": "user", "content": question})
            messages.append({"role": "assistant", "content": answer})
        messages.append({"role": "user", "content": prompt})
        logging.info(messages)

        try:

            if streaming:
                response = self.client.chat.completions.create(
                    model=self.model,
                    messages=messages,
                    stream=True,
                    max_tokens=self.max_token,
                    # temperature=self.temperature,
                    stop=[self.stop_words] if self.stop_words is not None else None,
                )

                for event in response:
                    if not isinstance(event, dict):
                        event = event.model_dump()

                    if event["choices"] is None:
                        event_text = event["text"] + " error_code:" + str(event["error_code"])
                    else:
                        event_text = event["choices"][0]['delta']['content']
                    if isinstance(event_text, str) and event_text != "":
                        # logging.info(f"[debug] event_text = [{event_text}]")
                        delta = {'answer': event_text}
                        yield "data: " + json.dumps(delta, ensure_ascii=False)

            else:
                response = self.client.chat.completions.create(
                    model=self.model,
                    messages=messages,
                    stream=False,
                    max_tokens=self.max_token,
                    # temperature=self.temperature,
                    stop=[self.stop_words] if self.stop_words is not None else None,
                )
                
                # logging.info(f"[debug] response.choices = [{response.choices}]")
                event_text = response.choices[0].message.content if response.choices else ""
                delta = {'answer': event_text}
                yield "data: " + json.dumps(delta, ensure_ascii=False)

        except Exception as e:
            logging.info(f"Error calling API: {e}")
            delta = {'answer': f"{e}"}
            yield "data: " + json.dumps(delta, ensure_ascii=False)

        finally:
            # logging.info("[debug] try-finally")
            yield f"data: [DONE]\n\n"

    def generatorAnswer(self, prompt: str,
                        history: List[List[str]] = [],
                        streaming: bool = False) -> AnswerResult:

        if history is None or len(history) == 0:
            history = [[]]
        logging.info(f"history_len: {self.history_len}")
        logging.info(f"prompt: {prompt}")
        logging.info(f"prompt tokens: {self.num_tokens_from_messages([prompt])}")
        logging.info(f"streaming: {streaming}")
                
        response = self._call(prompt, history[:-1], streaming)
        complete_answer = ""
        for response_text in response:

            if response_text:
                chunk_str = response_text[6:]
                if not chunk_str.startswith("[DONE]"):
                    chunk_js = json.loads(chunk_str)
                    complete_answer += chunk_js["answer"]
                    
            history[-1] = [prompt, complete_answer]
            answer_result = AnswerResult()
            answer_result.history = history
            if streaming:
                answer_result.llm_output = {"answer": response_text}
            else:
                answer_result.llm_output = {"answer": complete_answer}
            answer_result.prompt = prompt
            yield answer_result


if __name__ == "__main__":

    base_url = f"http://{LOCAL_LLM_SERVICE_URL}/api/v1/token_check" 
    headers = {"Content-Type": "application/json"}
    query = "hello"
    response = requests.post(
        base_url, 
        data=json.dumps(
            {'prompts': [{'model': LOCAL_LLM_MODEL_NAME, 'prompt': query, 'max_tokens': 512}]}
        ),
        headers=headers, timeout=60)

    # {'prompts': [{'fits': True, 'tokenCount': 317, 'contextLength': 8192}]}
    result = response.json()
    logging.info(f"[debug] result = {result}")


    llm = OpenAICustomLLM()
    streaming = True
    chat_history = []
    prompt = "你是谁"
    prompt = """参考信息：
中央纪委国家监委网站讯 据山西省纪委监委消息：山西转型综合改革示范区党工委副书记、管委会副主任董良涉嫌严重违纪违法，目前正接受山西省纪委监委纪律审查和监察调查。\\u3000\\u3000董良简历\\u3000\\u3000董良，男，汉族，1964年8月生，河南鹿邑人，在职研究生学历，邮箱random@xxx.com，联系电话131xxxxx909，1984年3月加入中国共产党，1984年8月参加工作\\u3000\\u3000历任太原经济技术开发区管委会副主任、太原武宿综合保税区专职副主任，山西转型综合改革示范区党工委委员、管委会副主任。2021年8月，任山西转型综合改革示范区党工委副书记、管委会副主任。(山西省纪委监委)
---
我的问题或指令：
帮我提取上述人物的中文名，英文名，性别，国籍，现任职位，最高学历，毕业院校，邮箱，电话
---
请根据上述参考信息回答我的问题或回复我的指令。前面的参考信息可能有用，也可能没用，你需要从我给出的参考信息中选出与我的问题最相关的那些，来为你的回答提供依据。回答一定要忠于原文，简洁但不丢信息，不要胡乱编造。我的问题或指令是什么语种，你就用什么语种回复,
你的回复："""
    final_result = ""
    for answer_result in llm.generatorAnswer(prompt=prompt,
                                                      history=chat_history,
                                                      streaming=streaming):
        resp = answer_result.llm_output["answer"]
        if "DONE" not in resp:
            final_result += json.loads(resp[6:])["answer"]
        # logging.info(resp)

    logging.info(f"final_result = {final_result}")