add_function_pmids_retrieval_and_ref_ench_sent

README.md

@@ -14,6 +14,7 @@ pinned: false
 - 整一个帮我写综述的Agent，希望他能完成文献内容的收集，文本分类和总结，科学事实对比，撰写综述等功能
 - 计划用到RAG, function calling等技术
 - 还在不断摸索中，欢迎大佬指导！
+- [huggingface 体验链接](https://huggingface.co/spaces/Yijun-Yang/ReadReview/), zeroGPUs 比较吝啬 我把本地推理给阉割了 不要用本地模型哈 用API 用本地模型会报错
 
 ## 流程图
 基本上就是在上海AIlab的茴香豆上面改的 这里主要讲解使用流程 架构和茴香豆一样 [茴香豆架构](https://github.com/InternLM/HuixiangDou/blob/main/docs/architecture_zh.md)
@@ -81,7 +82,7 @@ git clone https://github.com/jabberwockyang/MedicalReviewAgent.git
 cd MedicalReviewAgent
 pip install -r requirements.txt
    ```
-   huggingface-cli下载模型
+   huggingface-cli下载模型(optional, 第一次调用的时候hf会下载,但是可能有墙)
 
    ```bash
 cd /root && mkdir models
@@ -96,7 +97,8 @@ huggingface-cli download maidalun1020/bce-reranker-base_v1 --local-dir /root/mod
    ```bash
 conda activate ReviewAgent
 cd MedicalReviewAgent
-python3 app.py
+python3 app.py --model_downloaded True # 如果已经在/root/models下载了模型 这个参数会换一个配置文件,里面的modelpath是本地路径不是hf的仓库路径 自己显卡跑跑用这个
+python3 app.py # 如果不打算用本地/root/models储存的模型 这是hf的spaces的构建配置
    ```
    gradio在本地7860端口运行
 
@@ -170,7 +172,7 @@ python3 app.py
      
 ## 感谢
 1. [茴香豆](https://github.com/InternLM/HuixiangDou)
-2. [E-utilities](https://eutils.ncbi.nlm.nih.gov/entrez/eutils/efetch.fcgi?db=pmc&id=PMCID)
+2. [E-utilities](https://www.ncbi.nlm.nih.gov/books/NBK25499/)
 3. [Ragflow](https://github.com/infiniflow/ragflow/blob/main/README_zh.md)
 4. [Advanced RAG pipeline](https://medium.aiplanet.com/evaluating-naive-rag-and-advanced-rag-pipeline-using-langchain-v-0-1-0-and-ragas-17d24e74e5cf)
    

app.py

@@ -42,6 +42,10 @@ def parse_args():
                         action='store_true',
                         default=True,
                         help='Auto deploy required Hybrid LLM Service.')
+    parser.add_argument("--model_downloaded",
+                        type=bool,
+                        default=False,
+                        help="If the model has been downloaded in the root/models folder. Default is False.")
     args = parser.parse_args()
     return args
 
@@ -55,6 +59,8 @@ def update_remote_buttons(remote):
                                                 interactive=True,visible=True),
                 gr.Textbox(label="您的API",lines = 1,
                         interactive=True,visible=True),
+                gr.Textbox(label="base url",lines = 1,
+                        interactive=True,visible=True),
                 gr.Dropdown([],label="选择模型",
                             interactive=True,visible=True)
         ]
@@ -80,7 +86,7 @@ def udate_model_dropdown(remote_company):
     }
     return gr.Dropdown(choices= model_choices[remote_company])
 
-def update_remote_config(remote_ornot,remote_company = None,api = None,model = None):
+def update_remote_config(remote_ornot,remote_company = None,api = None,baseurl = None, model = None):
     with open(CONFIG_PATH, encoding='utf8') as f:
         config = pytoml.load(f)
          
@@ -91,6 +97,7 @@ def update_remote_config(remote_ornot,remote_company = None,api = None,model = N
             config['llm']['enable_remote'] = 1
             config['llm']['server']['remote_type'] = remote_company
             config['llm']['server']['remote_api_key'] = api
+            config['llm']['server']['remote_base_url'] = baseurl
             config['llm']['server']['remote_llm_model'] = model
         else:
             config['llm']['enable_local'] = 1
@@ -188,11 +195,17 @@ def upload_file(files):
 
     return files
 
-def generate_articles_repo(keywords:str,retmax:int):
-    keys= [k.strip() for k in keywords.split('\n')]
+def generate_articles_repo(strings:str,retmax:int):
+    
+    string = [k.strip() for k in strings.split('\n')]
+
+    pmids = [k for k in string if k.isdigit()]
+    keys = [k for k in string if not k.isdigit()]
+    
     repodir, _, _ = get_ready('repo_work')
 
     articelfinder = ArticleRetrieval(keywords = keys,
+                                     pmids = pmids,
                                      repo_dir = repodir,
                                      retmax = retmax)
     articelfinder.initiallize()
@@ -212,7 +225,7 @@ def delete_articles_repo():
 
 def update_repo():
     keys,len,retmax,pdflen = update_repo_info()
-    if keys:
+    if keys or len:
         newinfo = f"搜索得到文献：\n    关键词：{keys}\n    文献数量：{len}\n    获取上限：{retmax}\n\n上传文献：\n    数量：{pdflen}"
     else:
         if pdflen:
@@ -400,8 +413,8 @@ def summarize_text(query,chunksize:int,remote_ornot:bool):
       
     logger.info(f'{code}, {query}, {reply}, {references}')
     urls = getpmcurls(references)
-    mds = '\n'.join([f'[{ref}]({url})' for ref,url in zip(references,urls)])
-    return reply, gr.Markdown(label="参考文献",value = mds) 
+    mds = '\n\n'.join([f'[{ref}]({url})' for ref,url in zip(references,urls)])
+    return gr.Markdown(label="看看",value = reply,line_breaks=True) , gr.Markdown(label="参考文献",value = mds,line_breaks=True) 
 
 def main_interface():   
     with gr.Blocks() as demo:
@@ -436,13 +449,14 @@ def main_interface():
                 remote_company = gr.Dropdown(["kimi", "deepseek", "zhipuai",'gpt'],
                                             label="选择大模型提供商",interactive=False,visible=False)
                 api = gr.Textbox(label="您的API",lines = 1,interactive=False,visible=False)
+                baseurl = gr.Textbox(label="base url",lines = 1,interactive=False,visible=False)
                 model = gr.Dropdown([],label="选择模型",interactive=False,visible=False)
             
             confirm_button = gr.Button("保存配置")
 
-            remote_ornot.change(update_remote_buttons, inputs=[remote_ornot],outputs=[apimd,remote_company,api,model])
+            remote_ornot.change(update_remote_buttons, inputs=[remote_ornot],outputs=[apimd,remote_company,api,baseurl,model])
             remote_company.change(udate_model_dropdown, inputs=[remote_company],outputs=[model])
-            confirm_button.click(update_remote_config, inputs=[remote_ornot,remote_company,api,model],outputs=[confirm_button])
+            confirm_button.click(update_remote_config, inputs=[remote_ornot,remote_company,api,baseurl,model],outputs=[confirm_button])
 
 
         with gr.Tab("文献查找+数据库生成"):
@@ -478,6 +492,7 @@ def main_interface():
             with gr.Row(equal_height=True):
                 with gr.Column(scale=1):
                     input_keys = gr.Textbox(label="感兴趣的关键词",
+                                            value = "输入关键词或者PMID, 换行分隔",
                                                     lines = 5)
                     retmax = gr.Slider(
                             minimum=0,
@@ -593,7 +608,7 @@ def main_interface():
             query = gr.Textbox(label="想写什么")
             
             write_button = gr.Button("写综述")
-            output_text = gr.Textbox(label="看看",lines=10)
+            output_text = gr.Markdown(label="看看")
             output_references = gr.Markdown(label="参考文献")
             
             update_options.click(update_chunksize_dropdown,
@@ -620,8 +635,12 @@ def main_interface():
 # start service
 if __name__ == '__main__':
     args = parse_args()
-    # copy config from config-bak
-    shutil.copy('config-bak.ini', args.config_path) # yyj
+    # copy config from config-bak 
+    if args.model_downloaded:
+        shutil.copy('config-mod_down-bak.ini', args.config_path) # yyj
+    else:
+        shutil.copy('config-bak.ini', args.config_path) # yyj
+
     CONFIG_PATH = args.config_path
 
     if args.standalone is True:

applocal.py

@@ -0,0 +1,663 @@
+import argparse
+import json
+import time
+import os
+import glob
+import random
+import shutil
+from enum import Enum
+from threading import Thread
+from multiprocessing import Process, Value
+
+import gradio as gr
+import pytoml
+from loguru import logger
+# import spaces
+
+from huixiangdou.service import Worker, llm_serve, ArticleRetrieval, CacheRetriever, FeatureStore, FileOperation
+
+class PARAM_CODE(Enum):
+    """Parameter code."""
+    SUCCESS = 0
+    FAILED = 1
+    ERROR = 2
+
+def parse_args():
+    """Parse args."""
+    parser = argparse.ArgumentParser(description='Worker.')
+    parser.add_argument('--work_dir',
+                        type=str,
+                        default='workdir',
+                        help='Working directory.')
+    parser.add_argument('--repo_dir',
+                        type=str,
+                        default='repodir',
+                        help='Repository directory.')
+    parser.add_argument(
+        '--config_path',
+        default='config.ini',
+        type=str,
+        help='Worker configuration path. Default value is config.ini')
+    parser.add_argument('--standalone',
+                        action='store_true',
+                        default=True,
+                        help='Auto deploy required Hybrid LLM Service.')
+    parser.add_argument("--model_downloaded",
+                        type=bool,
+                        default=False,
+                        help="If the model has been downloaded in the root/models folder. Default is False.")
+    args = parser.parse_args()
+    return args
+
+def update_remote_buttons(remote):
+    if remote:
+        return [
+                gr.Markdown("[如何配置API]('https://github.com/jabberwockyang/MedicalReviewAgent/blob/main/README.md')",
+                            visible=True),
+                gr.Dropdown(["kimi", "deepseek", "zhipuai",'gpt'],
+                                                label="选择大模型提供商",
+                                                interactive=True,visible=True),
+                gr.Textbox(label="您的API",lines = 1,
+                        interactive=True,visible=True),
+                gr.Textbox(label="base url",lines = 1,
+                        interactive=True,visible=True),
+                gr.Dropdown([],label="选择模型",
+                            interactive=True,visible=True)
+        ]
+    else:
+        return [
+                gr.Markdown("[如何配置API]('https://github.com/jabberwockyang/MedicalReviewAgent/blob/main/README.md')",
+                            visible=False),
+                gr.Dropdown(["kimi", "deepseek", "zhipuai",'gpt'],
+                                                label="选择大模型提供商",
+                                                interactive=False,visible=False),
+                gr.Textbox(label="您的API",lines = 1,
+                        interactive=False,visible=False),
+                gr.Dropdown([],label="选择模型",
+                            interactive=False,visible=False)
+        ]
+
+def udate_model_dropdown(remote_company):
+    model_choices = {
+        'kimi': ['moonshot-v1-128k'],
+        'deepseek': ['deepseek-chat'],
+        'zhipuai': ['glm-4'],
+        'gpt': ['gpt-4-32k-0613','gpt-3.5-turbo']
+    }
+    return gr.Dropdown(choices= model_choices[remote_company])
+
+def update_remote_config(remote_ornot,remote_company = None,api = None,baseurl = None, model = None):
+    with open(CONFIG_PATH, encoding='utf8') as f:
+        config = pytoml.load(f)
+         
+        if remote_ornot:
+            if remote_company == None or api == None or model == None:
+                raise ValueError('remote_company, api, model not provided')
+            config['llm']['enable_local'] = 0
+            config['llm']['enable_remote'] = 1
+            config['llm']['server']['remote_type'] = remote_company
+            config['llm']['server']['remote_api_key'] = api
+            config['llm']['server']['remote_base_url'] = baseurl
+            config['llm']['server']['remote_llm_model'] = model
+        else:
+            config['llm']['enable_local'] = 1
+            config['llm']['enable_remote'] = 0
+    with open(CONFIG_PATH, 'w') as f:
+        pytoml.dump(config, f)
+    return gr.Button("配置已保存")
+
+# @spaces.GPU(duration=360)
+def get_ready(query:str,chunksize=None,k=None):
+    
+    with open(CONFIG_PATH, encoding='utf8') as f:
+        config = pytoml.load(f)
+    workdir = config['feature_store']['work_dir']
+    repodir = config['feature_store']['repo_dir']
+
+    if query == 'repo_work': # no need to return assistant
+        return repodir, workdir, config
+    theme = ''
+    try:
+        with open(os.path.join(config['feature_store']['repo_dir'],'config.json'), 'r') as f:
+            repo_config = json.load(f)
+        theme = ' '.join(repo_config['keywords'])
+    except:
+        pass
+
+    if query == 'annotation':
+        if not chunksize or not k:
+            raise ValueError('chunksize or k not provided')
+        chunkdir = os.path.join(workdir, f'chunksize_{chunksize}')
+        clusterdir = os.path.join(chunkdir, 'cluster_features', f'cluster_features_{k}')
+        assistant = Worker(work_dir=chunkdir, config_path=CONFIG_PATH,language='en')
+        samples_json = os.path.join(clusterdir,'samples.json')
+        with open(samples_json, 'r') as f:
+            samples = json.load(f)
+            f.close()
+        return clusterdir, samples, assistant, theme
+    
+    elif query == 'inspiration':
+        if not chunksize or not k:
+            raise ValueError('chunksize or k not provided')
+        
+        chunkdir = os.path.join(workdir, f'chunksize_{chunksize}')
+        clusterdir = os.path.join(chunkdir, 'cluster_features', f'cluster_features_{k}')
+        assistant = Worker(work_dir=chunkdir, config_path=CONFIG_PATH,language='en')
+        annofile = os.path.join(clusterdir,'annotation.jsonl')
+        with open(annofile, 'r') as f:
+            annoresult = f.readlines()
+
+            f.close()
+        annoresult = [json.loads(obj) for obj in annoresult]
+        return clusterdir, annoresult, assistant, theme
+    elif query == 'summarize': # no need for params k
+        if not chunksize:
+            raise ValueError('chunksize not provided')
+        chunkdir = os.path.join(workdir, f'chunksize_{chunksize}')
+        assistant = Worker(work_dir=chunkdir, config_path=CONFIG_PATH,language='en')
+        return assistant,theme
+
+    else:
+        raise ValueError('query not recognized')
+            
+def update_repo_info():
+    with open(CONFIG_PATH, encoding='utf8') as f:
+        config = pytoml.load(f)
+    repodir = config['feature_store']['repo_dir']
+    if os.path.exists(repodir):
+        pdffiles = glob.glob(os.path.join(repodir, '*.pdf'))
+        number_of_pdf = len(pdffiles)
+        if os.path.exists(os.path.join(repodir,'config.json')):
+                
+            with open(os.path.join(repodir,'config.json'), 'r') as f:
+                repo_config = json.load(f)
+
+            keywords = repo_config['keywords']
+            length = repo_config['len']
+            retmax = repo_config['retmax']
+
+            return keywords,length,retmax,number_of_pdf
+        else:
+            return None,None,None,number_of_pdf
+    else:
+        return None,None,None,None
+               
+def upload_file(files):
+    repodir, workdir, _ = get_ready('repo_work')
+    if not os.path.exists(repodir):
+        os.makedirs(repodir)
+
+    for file in files:
+        destination_path = os.path.join(repodir, os.path.basename(file.name))
+
+        shutil.copy(file.name, destination_path)
+    
+
+    return files
+
+def generate_articles_repo(strings:str,retmax:int):
+    
+    string = [k.strip() for k in strings.split('\n')]
+
+    pmids = [k for k in string if k.isdigit()]
+    keys = [k for k in string if not k.isdigit()]
+    
+    repodir, _, _ = get_ready('repo_work')
+
+    articelfinder = ArticleRetrieval(keywords = keys,
+                                     pmids = pmids,
+                                     repo_dir = repodir,
+                                     retmax = retmax)
+    articelfinder.initiallize()
+    return update_repo()
+
+def delete_articles_repo():
+    # 在这里运行生成数据库的函数
+    repodir, workdir, _ = get_ready('repo_work')
+    if os.path.exists(repodir):
+        shutil.rmtree(repodir)
+    if os.path.exists(workdir):
+        shutil.rmtree(workdir)
+
+    return gr.Textbox(label="文献库概况",lines =3,
+                      value = '文献库和相关数据库已删除',
+                      visible = True)
+
+def update_repo():
+    keys,len,retmax,pdflen = update_repo_info()
+    if keys or len:
+        newinfo = f"搜索得到文献：\n    关键词：{keys}\n    文献数量：{len}\n    获取上限：{retmax}\n\n上传文献：\n    数量：{pdflen}"
+    else:
+        if pdflen:
+            newinfo = f'搜索得到文献：无\n上传文献：\n    数量：{pdflen}'
+        else:
+            newinfo = '目前还没有文献库'
+
+    return gr.Textbox(label="文献库概况",lines =1,
+                      value = newinfo,
+                      visible = True)
+
+def update_database_info():
+    with open(CONFIG_PATH, encoding='utf8') as f:
+        config = pytoml.load(f)
+    workdir = config['feature_store']['work_dir']
+    chunkdirs = glob.glob(os.path.join(workdir, 'chunksize_*'))
+    chunkdirs.sort()
+    list_of_chunksize = [int(chunkdir.split('_')[-1]) for chunkdir in chunkdirs]
+    # print(list_of_chunksize)
+    jsonobj = {}
+    for chunkdir in chunkdirs:
+        k_dir = glob.glob(os.path.join(chunkdir, 'cluster_features','cluster_features_*'))
+        k_dir.sort()
+        list_of_k = [int(k.split('_')[-1]) for k in k_dir]
+        jsonobj[int(chunkdir.split('_')[-1])] = list_of_k
+        
+
+    new_options = [f"chunksize:{chunksize}, k:{k}" for chunksize in list_of_chunksize for k in jsonobj[chunksize]]
+    
+    return new_options, jsonobj
+
+# @spaces.GPU(duration=360)
+def generate_database(chunksize:int,nclusters:str|list[str]):
+    # 在这里运行生成数据库的函数
+    repodir, workdir, _ = get_ready('repo_work')
+    if not os.path.exists(repodir):
+        return gr.Textbox(label="数据库已生成",value = '请先生成文献库',visible = True)
+    nclusters = [int(i) for i in nclusters]
+    # 文献库和数据库的覆盖删除逻辑待定 
+    # 理论上 文献库只能生成一次 所以每次生成文献库都要删除之前的文献库和数据库
+    # 数据库可以根据文献库多次生成 暂不做删除 目前没有节省算力的逻辑 重复计算后覆盖 以后优化 
+    # 不同的chunksize和nclusters会放在不同的文件夹下 不会互相覆盖
+    # if os.path.exists(workdir):
+    #     shutil.rmtree(workdir)
+
+    cache = CacheRetriever(config_path=CONFIG_PATH)
+    fs_init = FeatureStore(embeddings=cache.embeddings,
+                           reranker=cache.reranker,
+                            chunk_size=chunksize,
+                            n_clusters=nclusters,
+                           config_path=CONFIG_PATH)
+
+    # walk all files in repo dir
+    file_opr = FileOperation()
+    files = file_opr.scan_dir(repo_dir=repodir)
+    fs_init.initialize(files=files, work_dir=workdir,file_opr=file_opr)
+    file_opr.summarize(files)
+    del fs_init
+    cache.pop('default')
+    texts, _ = update_database_info()
+    return gr.Textbox(label="数据库概况",value = '\n'.join(texts) ,visible = True)
+
+def delete_database():
+    _, workdir, _ = get_ready('repo_work')
+    if os.path.exists(workdir):
+        shutil.rmtree(workdir)
+    return  gr.Textbox(label="数据库概况",lines =3,value = '数据库已删除',visible = True)
+
+def update_database_textbox():
+    texts, _ = update_database_info()
+    if texts == []:
+        return gr.Textbox(label="数据库概况",value = '目前还没有数据库',visible = True)
+    else:
+        return gr.Textbox(label="数据库概况",value = '\n'.join(texts),visible = True)
+
+def update_chunksize_dropdown():
+    _, jsonobj = update_database_info()
+    return gr.Dropdown(choices= jsonobj.keys())
+
+def update_ncluster_dropdown(chunksize:int):
+    _, jsonobj = update_database_info()
+    nclusters = jsonobj[chunksize]
+    return gr.Dropdown(choices= nclusters)
+
+# @spaces.GPU(duration=360)
+def annotation(n,chunksize:int,nclusters:int,remote_ornot:bool):
+    '''
+    use llm to annotate cluster
+    n: percentage of clusters to annotate
+    '''
+    query = 'annotation'
+    if remote_ornot:
+        backend = 'remote'
+    else:
+        backend = 'local'
+
+    clusterdir, samples, assistant, theme = get_ready('annotation',chunksize,nclusters)
+    new_obj_list = []
+    n = round(n * len(samples.keys()))
+    for cluster_no in random.sample(samples.keys(), n):
+        chunk = '\n'.join(samples[cluster_no]['samples'][:10])
+
+        code, reply, cluster_no = assistant.annotate_cluster(
+                                                theme = theme,
+                                                cluster_no=cluster_no,
+                                                chunk=chunk,
+                                                history=[],
+                                                groupname='',
+                                                backend=backend)
+        references = f"cluster_no: {cluster_no}"
+        new_obj = {
+            'cluster_no': cluster_no,
+            'chunk': chunk,
+            'annotation': reply
+        }
+        new_obj_list.append(new_obj)
+        logger.info(f'{code}, {query}, {reply}, {references}')
+
+        with open(os.path.join(clusterdir, 'annotation.jsonl'), 'a') as f:
+            json.dump(new_obj, f, ensure_ascii=False)
+            f.write('\n')
+            
+    return '\n\n'.join([obj['annotation'] for obj in new_obj_list])
+
+# @spaces.GPU(duration=360)
+def inspiration(annotation:str,chunksize:int,nclusters:int,remote_ornot:bool):
+    query = 'inspiration'
+    if remote_ornot:
+        backend = 'remote'
+    else:
+        backend = 'local'
+        
+    clusterdir, annoresult, assistant, theme = get_ready('inspiration',chunksize,nclusters)
+    new_obj_list = []
+
+    if annotation is not None: # if the user wants to get inspiration from specific clusters only  
+        annoresult = [obj for obj in annoresult if obj['annotation'] in [txt.strip() for txt in annotation.split('\n')]]
+    
+    for index in random.sample(range(len(annoresult)), min(5, len(annoresult))):
+        cluster_no = annoresult[index]['cluster_no']
+        chunks = annoresult[index]['annotation']
+        
+        code, reply = assistant.getinspiration(
+                                                theme = theme,
+                                                annotations = chunks,
+                                                history=[], 
+                                                groupname='',backend=backend)
+        new_obj = {
+            'inspiration': reply,
+            'cluster_no': cluster_no
+        }
+        new_obj_list.append(new_obj)
+        logger.info(f'{code}, {query}, {cluster_no},{reply}')
+
+        with open(os.path.join(clusterdir, 'inspiration.jsonl'), 'a') as f:
+            json.dump(new_obj, f, ensure_ascii=False)
+        with open(os.path.join(clusterdir, 'inspiration.txt'), 'a') as f:
+            f.write(f'{reply}\n')
+            
+    return '\n\n'.join(list(set([obj['inspiration'] for obj in new_obj_list])))
+
+
+def getpmcurls(references):
+    urls = []
+    for ref in references:
+        if ref.startswith('PMC'):
+            
+            refid = ref.replace('.txt','')
+            urls.append(f'https://www.ncbi.nlm.nih.gov/pmc/articles/{refid}/')
+        else:
+            urls.append(ref)
+    return urls
+    
+# @spaces.GPU(duration=360)
+def summarize_text(query,chunksize:int,remote_ornot:bool):
+    if remote_ornot:
+        backend = 'remote'
+    else:
+        backend = 'local'
+        
+    assistant,_ = get_ready('summarize',chunksize=chunksize,k=None)
+    code, reply, references = assistant.generate(query=query,
+                                                history=[],
+                                                groupname='',backend = backend)
+      
+    logger.info(f'{code}, {query}, {reply}, {references}')
+    urls = getpmcurls(references)
+    mds = '\n\n'.join([f'[{ref}]({url})' for ref,url in zip(references,urls)])
+    return gr.Markdown(label="看看",value = reply,line_breaks=True) , gr.Markdown(label="参考文献",value = mds,line_breaks=True) 
+
+def main_interface():   
+    with gr.Blocks() as demo:
+        with gr.Row():
+            gr.Markdown(
+                """
+                # 医学文献综述助手 (又名 不想看文献)
+                """
+            )
+
+        with gr.Tab("模型服务配置"):
+            gr.Markdown("""
+            #### 配置模型服务 🛠️
+
+            1. **是否使用远程大模型**
+            - 勾选此项，如果你想使用远程的大模型服务。
+            - 如果不勾选，将默认使用本地模型服务。
+
+            2. **API配置**
+            - 配置大模型提供商和API，确保模型服务能够正常运行。
+            - 提供商选择：kimi、deepseek、zhipuai、gpt。
+            - 输入您的API密钥和选择对应模型。
+            - 点击“保存配置”按钮以保存您的设置。
+
+            📝 **备注**：请参考[如何使用]('https://github.com/jabberwockyang/MedicalReviewAgent/blob/main/README.md')获取更多信息。
+
+            """)
+
+            remote_ornot = gr.Checkbox(label="是否使用远程大模型")
+            with gr.Accordion("API配置", open=True):
+                apimd = gr.Markdown("[如何配置API]('https://github.com/jabberwockyang/MedicalReviewAgent/blob/main/README.md')",visible=False)
+                remote_company = gr.Dropdown(["kimi", "deepseek", "zhipuai",'gpt'],
+                                            label="选择大模型提供商",interactive=False,visible=False)
+                api = gr.Textbox(label="您的API",lines = 1,interactive=False,visible=False)
+                baseurl = gr.Textbox(label="base url",lines = 1,interactive=False,visible=False)
+                model = gr.Dropdown([],label="选择模型",interactive=False,visible=False)
+            
+            confirm_button = gr.Button("保存配置")
+
+            remote_ornot.change(update_remote_buttons, inputs=[remote_ornot],outputs=[apimd,remote_company,api,baseurl,model])
+            remote_company.change(udate_model_dropdown, inputs=[remote_company],outputs=[model])
+            confirm_button.click(update_remote_config, inputs=[remote_ornot,remote_company,api,baseurl,model],outputs=[confirm_button])
+
+
+        with gr.Tab("文献查找+数据库生成"):
+            gr.Markdown("""
+#### 查找文献 📚
+
+1. **输入关键词批量PubMed PMC文献**
+   - 在“感兴趣的关键词”框中输入您感兴趣的关键词，每行一个。
+   - 设置查找数量（0-1000）。
+   - 点击“搜索PubMed PMC”按钮进行文献查找。
+
+2. **上传PDF**
+   - 通过“上传PDF”按钮上传您已有的PDF文献文件。
+
+3. **更新文献库情况 删除文献库**
+   - 点击“更新文献库情况”按钮，查看当前文献库的概况。
+   - 如果需要重置或删除现有文献库，点击“删除文献库”按钮。
+
+
+#### 生成数据库 🗂️
+
+1. **设置数据库构建参数 生成数据库**
+   - 选择块大小（Chunk Size）和聚类数（Number of Clusters）。
+   - 提供选项用于选择合适的块大小和聚类数。
+   - 点击“生成数据库”按钮开始数据库生成过程。
+
+2. **更新数据库情况 删除数据库**
+   - 点击“更新数据库情况”按钮，查看当前数据库的概况。
+   - 点击“删除数据库”按钮移除现有数据库。
+
+📝 **备注**：请参考[如何选择数据库构建参数]('https://github.com/jabberwockyang/MedicalReviewAgent/tree/main')获取更多信息。
+""")
+            with gr.Row(equal_height=True):
+                with gr.Column(scale=1):
+                    input_keys = gr.Textbox(label="感兴趣的关键词",
+                                            value = "输入关键词或者PMID, 换行分隔",
+                                                    lines = 5)
+                    retmax = gr.Slider(
+                            minimum=0,
+                            maximum=1000,
+                            value=500,
+                            interactive=True,
+                            label="查多少",
+                        )
+                    generate_repo_button = gr.Button("搜索PubMed PMC")
+                with gr.Column(scale=2):
+                    file_output = gr.File(scale=2)
+                    upload_button = gr.UploadButton("上传PDF", 
+                                    file_types=[".pdf",".csv",".doc"], 
+                                    file_count="multiple",scale=0)
+                    
+            with gr.Row(equal_height=True):
+                with gr.Column(scale=0):
+                    delete_repo_button = gr.Button("删除文献库")
+                    update_repo_button = gr.Button("更新文献库情况")
+                with gr.Column(scale=2):
+
+                    repo_summary =gr.Textbox(label= '文献库概况', value="目前还没有文献库")
+
+            generate_repo_button.click(generate_articles_repo, 
+                                inputs=[input_keys,retmax],
+                                outputs = [repo_summary])
+            
+            
+            delete_repo_button.click(delete_articles_repo, inputs=None,
+                                outputs = repo_summary)
+            update_repo_button.click(update_repo, inputs=None,
+                                outputs = repo_summary)
+            upload_button.upload(upload_file, upload_button, file_output)
+            
+            with gr.Accordion("数据库构建参数", open=True):
+                gr.Markdown("[如何选择数据库构建参数]('https://github.com/jabberwockyang/MedicalReviewAgent/tree/main')")
+                chunksize = gr.Slider(label="Chunk Size",
+                                      info= 'How long you want the chunk to be?',
+                                        minimum=128, maximum=4096,value=1024,step=1,
+                                        interactive=True)
+                ncluster = gr.CheckboxGroup(["10", "20", "50", '100','200','500','1000'], 
+                                            # default=["20", "50", '100'],
+                                            label="Number of Clusters", 
+                                            info="How many Clusters you want to generate")
+
+            with gr.Row():
+                gene_database_button = gr.Button("生成数据库")
+                delete_database_button = gr.Button("删除数据库")
+                update_database_button = gr.Button("更新数据库情况")
+
+            database_summary = gr.Textbox(label="数据库概况",lines = 1,value="目前还没有数据库")
+            
+
+            gene_database_button.click(generate_database, inputs=[chunksize,ncluster],
+                                outputs = database_summary)
+            
+            update_database_button.click(update_database_textbox,inputs=None,
+                                outputs = [database_summary])
+                                         
+            delete_database_button.click(delete_database, inputs=None,
+                                outputs = database_summary)
+        with gr.Tab("写综述"):
+            gr.Markdown("""
+#### 写综述 ✍️
+
+1. **更新数据库情况**
+   - 点击“更新数据库情况”按钮，确保使用最新的数据库信息。
+
+2. **选择块大小和聚类数**
+   - 从下拉菜单中选择合适的块大小和聚类数。
+
+3. **抽样标注文章聚类**
+   - 设置抽样标注比例（0-1）。
+   - 点击“抽样标注文章聚类”按钮开始标注过程。
+
+4. **获取灵感**
+   - 如果不知道写什么，点击“获取灵感”按钮。
+   - 系统将基于标注的文章聚类提供相应的综述子问题。
+
+5. **写综述**
+   - 输入您想写的内容或主题。
+   - 点击“写综述”按钮，生成综述文本。
+
+6. **查看生成结果**
+   - 生成的综述文本将显示在“看看”文本框中。
+   - 参考文献将显示在“参考文献”框中。
+
+📝 **备注**：可以尝试不同的参数进行标注和灵感获取，有助于提高综述的质量和相关性。
+            """)
+
+            with gr.Accordion("聚类标注相关参数", open=True):
+                with gr.Row():
+                    update_options = gr.Button("更新数据库情况", scale=0)
+                    chunksize = gr.Dropdown([], label="选择块大小", scale=0)
+                    nclusters = gr.Dropdown([], label="选择聚类数", scale=0)
+                    ntoread = gr.Slider(
+                            minimum=0,maximum=1,value=0.5,
+                            interactive=True,
+                            label="抽样标注比例",
+                        )
+
+            annotation_button = gr.Button("抽样标注文章聚类")
+            annotation_output =  gr.Textbox(label="文章聚类标注/片段摘要",
+                                            lines = 5, 
+                                            interactive= True,
+                                            show_copy_button=True)
+            inspiration_button = gr.Button("获取灵感")
+            inspiration_output = gr.Textbox(label="灵光一现",
+                                            lines = 5,
+                                            show_copy_button=True)
+
+
+            query = gr.Textbox(label="想写什么")
+            
+            write_button = gr.Button("写综述")
+            output_text = gr.Markdown(label="看看")
+            output_references = gr.Markdown(label="参考文献")
+            
+            update_options.click(update_chunksize_dropdown,
+                                outputs=[chunksize])
+            
+            chunksize.change(update_ncluster_dropdown, 
+                             inputs=[chunksize],
+                             outputs= [nclusters])
+            
+            annotation_button.click(annotation, 
+                                    inputs = [ntoread, chunksize, nclusters,remote_ornot],
+                                    outputs=[annotation_output])
+            
+            inspiration_button.click(inspiration, 
+                                     inputs= [annotation_output, chunksize, nclusters,remote_ornot],
+                                     outputs=[inspiration_output])
+            
+            write_button.click(summarize_text,
+                                inputs=[query, chunksize,remote_ornot],
+                                outputs =[output_text,output_references])
+
+    demo.launch(share=False, server_name='0.0.0.0', debug=True,show_error=True,allowed_paths=['img_0.jpg']) 
+ 
+# start service
+if __name__ == '__main__':
+    args = parse_args()
+    # copy config from config-bak 
+    if args.model_downloaded:
+        shutil.copy('config-mod_down-bak.ini', args.config_path) # yyj
+    else:
+        shutil.copy('config-bak.ini', args.config_path) # yyj
+
+    CONFIG_PATH = args.config_path
+
+    if args.standalone is True:
+        # hybrid llm serve
+        server_ready = Value('i', 0)
+        server_process = Process(target=llm_serve,
+                                args=(args.config_path, server_ready))
+        server_process.start()
+        while True:
+            if server_ready.value == 0:
+                logger.info('waiting for server to be ready..')
+                time.sleep(3)
+            elif server_ready.value == 1:
+                break
+            else:
+                logger.error('start local LLM server failed, quit.')
+                raise Exception('local LLM path')
+        logger.info('Hybrid LLM Server start.')
+
+    main_interface()

config-bak.ini

@@ -23,6 +23,7 @@ local_llm_max_text_length = 32000
 local_llm_bind_port = 8888
 remote_type = ""
 remote_api_key = ""
+remote_base_url = ""
 remote_llm_max_text_length = 32000
 remote_llm_model = ""
 rpm = 500

config-mod_down-bak.ini

@@ -0,0 +1,64 @@
+[feature_store]
+reject_throttle = 0
+embedding_model_path = "/root/models/bce-embedding-base_v1"
+reranker_model_path = "/root/models/bce-reranker-base_v1"
+repo_dir = "repodir"
+work_dir = "workdir"
+n_clusters = [20, 50]
+chunk_size = 1024
+
+[web_search]
+x_api_key = "${YOUR-API-KEY}"
+domain_partial_order = ["openai.com", "pytorch.org", "readthedocs.io", "nvidia.com", "stackoverflow.com", "juejin.cn", "zhuanlan.zhihu.com", "www.cnblogs.com"]
+save_dir = "logs/web_search_result"
+
+[llm]
+enable_local = 1
+enable_remote = 1
+client_url = "http://127.0.0.1:8888/inference"
+
+[llm.server]
+local_llm_path = "/root/models/Qwen1.5-7B-Chat"
+local_llm_max_text_length = 32000
+local_llm_bind_port = 8888
+remote_type = ""
+remote_api_key = ""
+remote_base_url = ""
+remote_llm_max_text_length = 32000
+remote_llm_model = ""
+rpm = 500
+
+[worker]
+enable_sg_search = 0
+save_path = "logs/work.txt"
+
+[worker.time]
+start = "00:00:00"
+end = "23:59:59"
+has_weekday = 1
+
+[sg_search]
+binary_src_path = "/usr/local/bin/src"
+src_access_token = "${YOUR-SRC-ACCESS-TOKEN}"
+
+[sg_search.opencompass]
+github_repo_id = "open-compass/opencompass"
+introduction = "用于评测大型语言模型（LLM）. 它提供了完整的开源可复现的评测框架，支持大语言模型、多模态模型的一站式评测，基于分布式技术，对大参数量模型亦能实现高效评测。评测方向汇总为知识、语言、理解、推理、考试五大能力维度，整合集纳了超过70个评测数据集，合计提供了超过40万个模型评测问题，并提供长文本、安全、代码3类大模型特色技术能力评测。"
+
+[sg_search.lmdeploy]
+github_repo_id = "internlm/lmdeploy"
+introduction = "lmdeploy 是一个用于压缩、部署和服务 LLM（Large Language Model）的工具包。是一个服务端场景下，transformer 结构 LLM 部署工具，支持 GPU 服务端部署，速度有保障，支持 Tensor Parallel，多并发优化，功能全面，包括模型转换、缓存历史会话的 cache feature 等. 它还提供了 WebUI、命令行和 gRPC 客户端接入。"
+
+[frontend]
+type = "none"
+webhook_url = "https://open.feishu.cn/open-apis/bot/v2/hook/xxxxxxxxxxxxxxx"
+message_process_policy = "immediate"
+
+[frontend.lark_group]
+app_id = "cli_a53a34dcb778500e"
+app_secret = "2ajhg1ixSvlNm1bJkH4tJhPfTCsGGHT1"
+encrypt_key = "abc"
+verification_token = "def"
+
+[frontend.wechat_personal]
+bind_port = 9527

huixiangdou/service/findarticles.py

@@ -10,28 +10,61 @@ from lxml import etree
 
 class ArticleRetrieval:
     def __init__(self,
-                    keywords: list,
+                    keywords: list = [],
+                    pmids: list = [],
                     repo_dir = 'repodir',
                     retmax = 500):
+        if keywords is [] and pmids is []:
+            raise ValueError("Either keywords or pmids must be provided.")
+        
         self.keywords = keywords
+        self.pmids = pmids
         self.repo_dir = repo_dir
         self.retmax = retmax
-        
-    ## 通过PMC数据库检索文章
-    def search_pmc(self):
+        self.pmc_ids = []
+
+
+    def esummary_pmc(self):
+        base_url = "https://eutils.ncbi.nlm.nih.gov/entrez/eutils/esummary.fcgi?"
+        params = {
+            "db": "pubmed",
+            "id": ','.join(self.pmids),
+            # "retmax": self.retmax
+        }
+        response = requests.get(base_url, params=params)
+        root = ET.fromstring(response.content)
+        results = []
+        for docsum in root.findall('DocSum'):
+            pmcid = None
+            id_value = docsum.find('Id').text
+            for item in docsum.findall('Item'):
+                if item.attrib.get('Name') == 'ArticleIds':
+                    for id_item in item.findall('Item'):
+                        if id_item.attrib.get('Name') == 'pmc':
+                            pmcid = id_item.text
+                            break
+
+            if pmcid:
+                results.append((id_value, pmcid))
+        self.esummary = results
+        self.pmc_ids = [r[1] for r in results]
+       
+    ## 通过Pubmed数据库检索文章
+    def esearch_pmc(self):
 
         base_url = "https://eutils.ncbi.nlm.nih.gov/entrez/eutils/esearch.fcgi"
         params = {
-            "db": "pmc",
+            "db": "pubmed",
             "term": '+'.join(self.keywords),
             "retmax": self.retmax
         }
         response = requests.get(base_url, params=params)
         root = ET.fromstring(response.content)
-        pmc_ids = [id_element.text for id_element in root.findall('.//Id')]
-        print(f"Found {len(pmc_ids)} articles.")
-        self.pmc_ids = pmc_ids
-        return pmc_ids
+        idlist = root.find('.//IdList') 
+        pmids = [id_element.text for id_element in idlist.findall('.//Id')]
+        print(f"Found {len(pmids)} articles for keywords {self.keywords}.")
+        self.pmids.extend(pmids)
+        
 
     # 解析XML文件
     def _get_all_text(self, element):
@@ -74,8 +107,8 @@ class ArticleRetrieval:
             if full_text.strip() == '':
                 continue
             else:
-                logger.info(full_text[:1000])
-                with open(os.path.join(self.repo_dir,f'PMC{id}.txt'), 'w') as f:
+                logger.info(full_text[:500])
+                with open(os.path.join(self.repo_dir,f'{id}.txt'), 'w') as f:
                     f.write(full_text)
                 self.success += 1
 
@@ -83,7 +116,12 @@ class ArticleRetrieval:
         config = {
             'keywords': self.keywords,
             'repo_dir': self.repo_dir,
-            'pmc_ids': self.pmc_ids,
+            'result': [
+                {
+                    'pmid': r[0],
+                    'pmcid': r[1]
+                } for r in self.esummary
+            ],
             'len': self.success,
             'retmax': self.retmax
         }
@@ -91,12 +129,33 @@ class ArticleRetrieval:
             json.dump(config, f, indent=4, ensure_ascii=False)
 
     def initiallize(self):
-        self.search_pmc()
-        self.fetch_full_text()
-        self.save_config()
+        if self.keywords !=[]:
+            print(self.keywords)
+            self.esearch_pmc() # get pmids from pubmed database using keywords
+
+        self.esummary_pmc() # get pmc ids from pubmed database using pmids
+        self.fetch_full_text() # get full text from pmc database using pmc ids
+        self.save_config() # save config file
 
 if __name__ == '__main__':
     if os.path.exists('repodir'):
         shutil.rmtree('repodir')
-    articelfinder = ArticleRetrieval(keywords = ['covid-19'],repo_dir = 'repodir',retmax = 5)
+    
+    strings = """
+36944324
+38453907
+38300432
+38651453
+38398096
+38255885
+38035547
+38734498"""
+    string = [k.strip() for k in strings.split('\n')]
+
+    pmids = [k for k in string if k.isdigit()]
+    print(pmids)
+    keys = [k for k in string if not k.isdigit() and k != '']
+    print(keys)
+    articelfinder = ArticleRetrieval(keywords = keys,pmids = pmids,
+                                     repo_dir = 'repodir',retmax = 5)
     articelfinder.initiallize()

huixiangdou/service/llm_server_hybrid.py

@@ -202,6 +202,7 @@ class HybridLLMServer:
         self.remote_model = self.server_config['remote_llm_model']
         self.remote_max_length = self.server_config['remote_llm_max_text_length']
         self.api_key = self.server_config['remote_api_key']
+        self.base_url = self.server_config['remote_base_url']
 
     def call_puyu(self, prompt, history):
         self.reload_config()
@@ -345,6 +346,9 @@ class HybridLLMServer:
         if base_url is not None:
             client = OpenAI(api_key=self.server_config['remote_api_key'],
                             base_url=base_url)
+        elif self.base_url != '':
+            client = OpenAI(api_key=self.server_config['remote_api_key'],
+                            base_url=self.base_url)
         else:
             client = OpenAI(api_key=self.server_config['remote_api_key'])
 

huixiangdou/service/worker.py

@@ -17,6 +17,14 @@ from .sg_search import SourceGraphProxy
 from .web_search import WebSearch
 
 
+def convertid2url(text):
+    # Regular expression to find all PMC references
+    pattern = r"\[PMC(\d+)\]"
+    # Function to replace each match with a URL link
+    replacement = lambda match: f"[PMC{match.group(1)}](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC{match.group(1)}/)"
+    # Replace all occurrences in the text
+    formatted_text = re.sub(pattern, replacement, text)
+    return formatted_text
 class Worker:
     """The Worker class orchestrates the logic of handling user queries,
     generating responses and managing several aspects of a chat assistant. It
@@ -85,7 +93,7 @@ class Worker:
             self.SECURITY_TEMAPLTE = 'Evaluate whether the following sentence involves prohibited content such as politics, insult, pornography, terror, religion, cyber violence, racial discrimination, etc., rate it on a scale of 0-10, do not explain, just give the score. The scoring standard is as follows: any violation directly gets 10 points; completely unrelated gets 0 points. Give the score, no explanation: "{}"'  # noqa E501
             self.PERPLESITY_TEMPLATE = 'Question: {} Answer: {}\nRead the dialogue above, does the answer express that they don\'t know? The more comprehensive the answer, the lower the score. Rate it on a scale of 0-10, no explanation, just give the score.\nThe scoring standard is as follows: an accurate answer to the question gets 0 points; a detailed answer gets 1 point; knowing some answers but having uncertain information gets 8 points; knowing a small part of the answer but recommends seeking help from others gets 9 points; not knowing any of the answers and directly recommending asking others for help gets 10 points. Just give the score, no explanation.'  # noqa E501
             self.SUMMARIZE_TEMPLATE = '"{}" \n Read the content above carefully, summarize it in a short and powerful way.'  # noqa E501
-            self.GENERATE_TEMPLATE = 'Background Information: "{}"\n Question: "{}"\n Please read the reference material carefully and answer the question.'  # noqa E501
+            self.GENERATE_TEMPLATE = 'Background Information: "{}"\n Question: "{}"\n Please read the reference material carefully and answer the question.  with reference id at the end of the corresponding content for example:  Primary determinants of the therapeutic approach are age, comorbidities, and diagnostic molecular profile [PMC9958584]'  # noqa E501
             self.ANNOTATE_CLUSTER = 'these are chunklized sentences from different papers about{}, they are clustered by similarity, the following is 10 samples from one of the cluster: "{}"\n Please tag the cluster in one breif sentence.' 
             self.INSPIRATION_TEMPLATE = 'Given the following summary of the articles content about {0} {1}, give some idea or sub-questions of the review about {0}, one question is sufficient.'  # noqa E501
     
@@ -205,7 +213,7 @@ class Worker:
                                 tracker=tracker,
                                 throttle=5,
                                 default=10,backend=backend):
-                context += chunk
+                context += f"reference: {ref} content: {chunk}"
                 context += '\n\n'
                 refs.append(ref)
         refs = list(set(refs))
@@ -219,8 +227,13 @@ class Worker:
                                                     backend=backend,
                                                     history=history)
             tracker.log('feature store doc', [chunk, response])
+            response = convertid2url(response)
+
             return ErrorCode.SUCCESS, response, refs
+        else:
+            return ErrorCode.NO_SEARCH_RESULT, response, references
 
+            
         # try:
         #     references = []
         #     web_context = ''
@@ -266,13 +279,13 @@ class Worker:
         # except Exception as e:
         #     logger.error(e)
 
-        if response is not None and len(response) > 0:
-            prompt = self.PERPLESITY_TEMPLATE.format(query, response)
-            if self.single_judge(prompt=prompt,
-                                 tracker=tracker,
-                                 throttle=10,
-                                 default=0,backend = backend):
-                reborn_code = ErrorCode.BAD_ANSWER
+        # if response is not None and len(response) > 0:
+        #     prompt = self.PERPLESITY_TEMPLATE.format(query, response)
+        #     if self.single_judge(prompt=prompt,
+        #                          tracker=tracker,
+        #                          throttle=10,
+        #                          default=0,backend = backend):
+        #         reborn_code = ErrorCode.BAD_ANSWER
 
         # if self.config['worker']['enable_sg_search']:
         #     if reborn_code == ErrorCode.BAD_ANSWER or reborn_code == ErrorCode.NO_SEARCH_RESULT:  # noqa E501
@@ -301,11 +314,11 @@ class Worker:
         #                                  default=0):
         #                 return ErrorCode.BAD_ANSWER, response, references
 
-        if response is not None and len(response) >= 800:
-            # reply too long, summarize it
-            response = self.llm.generate_response(
-                                                prompt=self.SUMMARIZE_TEMPLATE.format(response),
-                                                backend=backend)
+        # if response is not None and len(response) >= 800:
+        #     # reply too long, summarize it
+        #     response = self.llm.generate_response(
+        #                                         prompt=self.SUMMARIZE_TEMPLATE.format(response),
+        #                                         backend=backend)
 
         # if len(response) > 0 and self.single_judge(
         #         self.SECURITY_TEMAPLTE.format(response),
@@ -314,10 +327,9 @@ class Worker:
         #         default=0):
         #     return ErrorCode.SECURITY, response, references
 
-        if reborn_code != ErrorCode.SUCCESS:
-            return reborn_code, response, references
-
-        return ErrorCode.SUCCESS, response, references
+        # if reborn_code != ErrorCode.SUCCESS:
+        #     return reborn_code, response, references
+        # return ErrorCode.SUCCESS, response, references
 
     def annotate_cluster(self,theme, cluster_no, chunk, history, groupname,backend):
         """Annotates a cluster of questions based on the user query and

run.sh

@@ -1,7 +1,7 @@
 # 云端运行
 conda activate ReviewAgent
 cd MedicalReviewAgent
-python3 app.py
+python3 applocal.py --model_downloaded True
 
 # 本地映射端口
-ssh -CNg -L 7860:127.0.0.1:7860 root@43.133.72.216
+ssh -CNg -L 7860:127.0.0.1:7860 root@youripaddress