diff --git a/README.md b/README.md index ea147e0..02de9da 100644 --- a/README.md +++ b/README.md @@ -1,6 +1,6 @@ -English | [简体中文](https://github.com/RapidAI/Knowledge-QA-LLM/blob/main/docs/README_zh.md) +[简体中文](https://github.com/RapidAI/Knowledge-QA-LLM/blob/main/docs/README_zh.md) | English -## Knowledge QA LLM +## 🧐 Knowledge QA LLM

@@ -74,8 +74,9 @@ English | [简体中文](https://github.com/RapidAI/Knowledge-QA-LLM/blob/main/d │ └── raw_upload_files ├── knowledge_qa_llm │ ├── __init__.py -│ ├── config.yaml # configuration file +│ ├── config.yaml # configuration file │ ├── file_loader # Handle documents in various formats +│ ├── encoder # Extract embeddings │ ├── llm # Large model interface, the large model needs to be deployed separately and called by interface │ ├── utils │ └── vector_utils # embedding access and search @@ -87,7 +88,12 @@ English | [简体中文](https://github.com/RapidAI/Knowledge-QA-LLM/blob/main/d └── webui.py # UI implementation based on streamlit ``` -#### Update Log +#### Change Log +- 2023-08-11 v0.0.7 update: + - Optimize layout, remove the plugin option, and put the extract vector model option on the home page. + - The tips are translated into English for easy communication. + - Add project logo:🧐 + - Update CLI module code. - 2023-08-05 v0.0.6 update: - Adapt more llm_api, include online llm api, such ad ERNIE-Bot-Turbo. - Add the status of extracting embeddings. diff --git a/cli.py b/cli.py index 5aa654e..44a62f1 100644 --- a/cli.py +++ b/cli.py @@ -1,10 +1,11 @@ # -*- encoding: utf-8 -*- # @Author: SWHL # @Contact: liekkaskono@163.com +from knowledge_qa_llm.encoder import EncodeText from knowledge_qa_llm.file_loader import FileLoader -from knowledge_qa_llm.llm import ChatGLM26B +from knowledge_qa_llm.llm import Qwen7B_Chat from knowledge_qa_llm.utils import make_prompt, read_yaml -from knowledge_qa_llm.vector_utils import DBUtils, EncodeText +from knowledge_qa_llm.vector_utils import DBUtils config = read_yaml("knowledge_qa_llm/config.yaml") @@ -16,18 +17,20 @@ # sentences = text[0][1] # 提取特征 -embedding_model = EncodeText(config.get("encoder_model_path")) +embedding_model = EncodeText(config.get("Encoder")["m3e-small"]) # embeddings = embedding_model(sentences) # 插入数据到数据库中 db_tools = DBUtils(config.get("vector_db_path")) # db_tools.insert(file_path, embeddings, sentences) -llm_engine = ChatGLM26B(api_url=config.get("llm_api_url")) +llm_engine = Qwen7B_Chat(api_url=config.get("LLM_API")["Qwen7B_Chat"]) -print("欢迎使用 Knowledge QA LLM，输入内容即可进行对话，stop 终止程序") +print( + "Welcom to 🧐 Knowledge QA LLM，enter the content to start the conversation, enter stop to terminate the program." +) while True: - query = input("\n用户：") + query = input("\n😀 User：") if query.strip() == "stop": break @@ -37,4 +40,4 @@ prompt = make_prompt(query, context, custom_prompt=config.get("DEFAULT_PROMPT")) response = llm_engine(prompt, history=None) - print(response) + print(f"🤖 LLM:{response}") diff --git a/docs/README_zh.md b/docs/README_zh.md index 16995b2..742b24c 100644 --- a/docs/README_zh.md +++ b/docs/README_zh.md @@ -1,6 +1,6 @@ -[English](https://github.com/RapidAI/Knowledge-QA-LLM) | 简体中文 +简体中文 | [English](https://github.com/RapidAI/Knowledge-QA-LLM) -## Knowledge QA LLM +## 🧐 Knowledge QA LLM

@@ -79,6 +79,7 @@ │   ├── __init__.py │   ├── config.yaml # 配置文件 │   ├── file_loader # 处理各种格式的文档 +│   ├── encoder # 提取特征向量 │   ├── llm # 大模型接口，大模型需要单独部署，以接口方式调用 │   ├── utils │   └── vector_utils # embedding的存取和搜索 @@ -91,6 +92,11 @@ ``` #### 更新日志 +- 2023-08-11 v0.0.7 update: + - 优化布局，去掉插件选项，将提取向量模型选项放到主页部分 + - 将提示语英语化，便于交流使用。 + - 添加项目logo: 🧐 + - 更新CLI使用代码 - 2023-08-05 v0.0.6 update: - 适配更多模型接口，包括在线大模型接口，例如文心一言 - 添加提取特征向量的状态提示 diff --git a/knowledge_qa_llm/config.yaml b/knowledge_qa_llm/config.yaml index 5c9486a..7541a5e 100644 --- a/knowledge_qa_llm/config.yaml +++ b/knowledge_qa_llm/config.yaml @@ -1,5 +1,5 @@ -title: Knowledge QA LLM -version: 0.0.6 +title: 🧐 Knowledge QA LLM +version: 0.0.7 LLM_API: Qwen7B_Chat: your_api @@ -34,16 +34,16 @@ Parameter: max_value: 4096 default: 1024 step: 1 - tip: 输入input_ids的最大长度 + tip: The longest length of input_ids top_p: min_value: 0.0 max_value: 1.0 default: 0.7 step: 0.01 - tip: 限制模型为仅考虑最可能的前p个标记 + tip: Limit the model to only consider the most likely first p markers. temperature: min_value: 0.01 max_value: 1.0 default: 0.01 step: 0.01 - tip: 控制模型输出的随机性，温度越低将导致输出更加可预测和重复，越高将更富创意和自发的输出。 + tip: Control the randomness of the model output. The smaller the value, the more standardized it is, and vice versa, the more creative it is. diff --git a/webui.py b/webui.py index 9247304..e5edbcd 100644 --- a/webui.py +++ b/webui.py @@ -23,7 +23,7 @@ def init_sidebar(): - st.sidebar.markdown("### 🛶 参数设置") + st.sidebar.markdown("### 🛶 Parameter Settings") param = config.get("Parameter") param_max_length = param.get("max_length") @@ -61,29 +61,22 @@ def init_sidebar(): ) st.session_state["params"]["temperature"] = temperature - st.sidebar.markdown("### 🧻 知识库") + st.sidebar.markdown("### 🧻 Knowledge Base") uploaded_files = st.sidebar.file_uploader( "default", accept_multiple_files=True, label_visibility="hidden", - help="支持多选", + help="Support for multiple selections", ) upload_dir = config.get("upload_dir") - - ENCODER_OPTIONS = config.get("Encoder") - select_encoder = st.sidebar.selectbox("🧬提取向量模型：", ENCODER_OPTIONS.keys()) - tips(f"初始化{select_encoder}...") - embedding_extract = init_encoder(ENCODER_OPTIONS[select_encoder]) - tips("初始化完成！") - - btn_upload = st.sidebar.button("上传文档并加载数据库", use_container_width=True) + btn_upload = st.sidebar.button("Upload and load database", use_container_width=True) if btn_upload: time_stamp = get_timestamp() save_dir = Path(upload_dir) / time_stamp st.session_state["upload_dir"] = save_dir - tips("正在上传文件到平台...", icon="⏳") + tips("Uploading files to platform...", icon="⏳") for file in uploaded_files: bytes_data = file.getvalue() @@ -91,12 +84,12 @@ def init_sidebar(): save_path = save_dir / file.name with open(save_path, "wb") as f: f.write(bytes_data) - tips("上传完毕！") + tips("Upload completed！") doc_dir = st.session_state["upload_dir"] all_doc_contents = file_loader(doc_dir) - pro_text = "正在提取特征向量..." + pro_text = "Extracting embeddings..." batch_size = config.get("encoder_batch_size", 32) for file_path, one_doc_contents in all_doc_contents.items(): my_bar = st.sidebar.progress(0, text=pro_text) @@ -111,7 +104,7 @@ def init_sidebar(): my_bar.progress( end_idx / content_nums, - f"提取{file_path}数据: [{end_idx}/{content_nums}]", + f"Extract {file_path} datas: [{end_idx}/{content_nums}]", ) my_bar.empty() all_embeddings = np.vstack(all_embeddings) @@ -119,15 +112,13 @@ def init_sidebar(): my_bar.empty() shutil.rmtree(doc_dir.resolve()) - tips("已经加载并存入数据库中，可以提问了！") + tips("You can now ask a question!") had_files = db_tools.get_files() if had_files: - st.sidebar.markdown("仓库已有文档：") + st.sidebar.markdown("Existing documents:") st.sidebar.markdown("\n".join([f" - {v}" for v in had_files])) - return embedding_extract - def init_state(): if "history" not in st.session_state: @@ -153,26 +144,26 @@ def predict( logger.info(f"Using {type(model).__name__}") query_embedding = embedding_extract(text) - with st.spinner("从文档中搜索相关内容"): + with st.spinner("Search for relevant contents from docs..."): search_res, search_elapse = db_tools.search_local( query_embedding, top_k=config.get("top_k") ) if search_res is None: - bot_print("从文档中搜索相关内容为空，暂不能回答该问题") + bot_print("The results of searching from docs is empty.") else: context = "\n".join(sum(search_res.values(), [])) - res_cxt = f"**从文档中检索到的相关内容Top5\n(相关性从高到低，耗时:{search_elapse:.5f}s):** \n" + res_cxt = f"**Find Top{search_top}\n(Scores from high to low，cost:{search_elapse:.5f}s):** \n" bot_print(res_cxt) for file, content in search_res.items(): content = "\n".join(content) - one_context = f"**来自文档：《{file}》** \n{content}" + one_context = f"**From：《{file}》** \n{content}" bot_print(one_context) - logger.info(f"上下文：\n{one_context}\n") + logger.info(f"Context：\n{one_context}\n") response, elapse = get_model_response(text, context, custom_prompt, model) - print_res = f"**使用模型：{select_model}**\n**模型推理耗时：{elapse:.5f}s**" + print_res = f"**Use：{select_model}**\n**Infer model cost：{elapse:.5f}s**" bot_print(print_res) bot_print(response) @@ -193,14 +184,14 @@ def get_model_response(text, context, custom_prompt, model): s_model = time.perf_counter() prompt_msg = make_prompt(text, context, custom_prompt) - logger.info(f"最终拼接后的文本：\n{prompt_msg}\n") + logger.info(f"Final prompt: \n{prompt_msg}\n") response = model(prompt_msg, history=None, **params_dict) elapse = time.perf_counter() - s_model - logger.info(f"模型回答: \n{response}\n") + logger.info(f"Reponse of LLM: \n{response}\n") if not response: - response = "抱歉，未能正确回答该问题" + response = "Sorry, I didn't answer the question correctly" return response, elapse @@ -222,7 +213,7 @@ def tips(txt: str, wait_time: int = 2, icon: str = "🎉"): db_path = config.get("vector_db_path") db_tools = DBUtils(db_path) - embedding_extract = init_sidebar() + init_sidebar() init_state() llm_module = importlib.import_module("knowledge_qa_llm.llm") @@ -240,22 +231,24 @@ def tips(txt: str, wait_time: int = 2, icon: str = "🎉"): } ) - PLUGINS_OPTIONS = { - "文档": 0, - } TOP_OPTIONS = [5, 10, 15] + ENCODER_OPTIONS = config.get("Encoder") menu_col1, menu_col2, menu_col3 = st.columns([1, 1, 1]) - select_model = menu_col1.selectbox("🎨基础模型：", MODEL_OPTIONS.keys()) - select_plugin = menu_col2.selectbox("🛠Plugin：", PLUGINS_OPTIONS.keys()) - search_top = menu_col3.selectbox("🔍查找Top_K", TOP_OPTIONS) + select_model = menu_col1.selectbox("🎨Base model：", MODEL_OPTIONS.keys()) + select_encoder = menu_col2.selectbox( + "🧬Extract Embedding Model：", ENCODER_OPTIONS.keys() + ) + search_top = menu_col3.selectbox("🔍Search Top_K", TOP_OPTIONS) + + embedding_extract = init_encoder(ENCODER_OPTIONS[select_encoder]) input_prompt_container = st.container() with input_prompt_container: with st.expander("💡Prompt", expanded=False): text_area = st.empty() input_prompt = text_area.text_area( - label="输入", + label="Input", max_chars=500, height=200, label_visibility="hidden", @@ -268,15 +261,13 @@ def tips(txt: str, wait_time: int = 2, icon: str = "🎉"): with st.chat_message("user", avatar="😀"): st.markdown(input_txt) - plugin_id = PLUGINS_OPTIONS[select_plugin] llm = MODEL_OPTIONS[select_model] if not input_prompt: input_prompt = config.get("DEFAULT_PROMPT") - if plugin_id == 0: - predict( - input_txt, - llm, - input_prompt, - ) + predict( + input_txt, + llm, + input_prompt, + )