[Request] 未来新知识库功能请求 #6054
Comments
|
👀 @BryceWG Thank you for raising an issue. We will investigate into the matter and get back to you as soon as possible. |
🥰 Requirement descriptionIt is generally believed that the knowledge base mainly uses rag technology, which corresponds to directly using files as context. 🧐 SolutionWhen selecting an attached file in a conversation, you can choose to directly use the 'file' already in the knowledge base as the context 📝 Supplementary informationNo response |
|
其实现在的交互就是支持的，但是由于之前是 RAG 的方式没法做全文注入所以效果不理想，这次2.0会把全文注入的能力加上，应该在一些需要全文引用的场景，效果会大大提升的 |
In fact, the current interaction is supported, but since the previous RAG method could not be used for full text injection, the effect is not ideal. This time, 2.0 will add the ability to inject the full text, and it should be greatly improved in some scenarios that require full text citation. of |
|
说到引用,能不能实现引用对话的内容?类似豆包的这个功能,选中对话中的一段文字,作为上下文:
|
Speaking of quotations, can the content of the quotation dialogue be implemented? Similar to this function of Doubao, select a paragraph of text in the dialogue as the context:
|
same request |
same request |
|
补充一下，目前lobechat上传文件后还需要等待向量化才送给大模型。有些PPT或PDF文件本身有比较复杂的图表时会报错向量化失败。但是其他类似的平台例如cherry studio似乎是将整个文件发送给大模型，速度快而且读的内容更精准。 |
To add, lobechat currently needs to wait for vectorization to be given to the big model after uploading the file. Some PPT or PDF files themselves have relatively complex charts and will report error vectorization failure. But other similar platforms such as cherry studio seem to send the entire file to the big model, which is fast and read more accurately. |
我也用过cherry,只有少数服务商的api支持直接接收文件,按作者说的没有单独适配的api其实都是本地解析出文件内容再发送 |
I have also used cherry. Only a few APIs of service providers support direct reception of files. According to the author's statement, APIs that do not have separate adaptations are actually parsed locally and then sent |
|
1.建议接入Doc2X等一线文档解析API，提高知识库文档解析精度。 |
|
|
希望后面版本中，知识库在设置里增加一个选择量化的模型的入口，方便用户选择自己想用的在线或本地模型构建知识库，帮助文档里，更改环境变量的方式，换了一下，有太麻烦了，有些模型又报错，后面知识库希望可以直接添加网页地址，自动量化网页数据到知识库 |
I hope that in the later version, the knowledge base will add an entry to select the quantized model in the settings, so that users can choose the online or local model they want to use to build the knowledge base. In the help document, the way to change the environment variables, it has changed, it is too I'm in trouble. Some models reported an error again. I hope that the knowledge base can directly add the web address and automatically quantify the web page data to the knowledge base. |
|
期望可以在线创建文档,markdown格式在线编辑内容,保存后的数据可以手动向量化,在对话时可以勾选某个文档或者某个文档的目录,等同于AI笔记+对话 |
I hope to create documents online, edit content online in markdown format, and save data can be manually vectorized. During conversation, you can check a document or a directory of a document, which is equivalent to AI notes + dialogue. |
对于 Gemini 这样的服务商，提供了文件上传的接口，希望在全文注入时可选直接调用这种接口以获得更好的性能 |
For service providers like Gemini, they provide an interface for file upload, hoping to directly call this interface when injecting the full text to obtain better performance. |
|
个人建议可以在聊天界面的侧边栏增加一个 panel，里面有当前对话的文件/知识库列表及复选框，每次可选发送部分文件/在部分文件中进行 RAG 检索（或许是 NotebookLM 类似的交互体验） |
Personally, I suggest you add a panel to the sidebar of the chat interface, which contains the current conversation file/knowledge base list and check boxes. You can select some files/retrieve RAG in some files (perhaps similar to NotebookLM. Interactive experience) |
|
建议知识库加上这些功能 |
Recommended knowledge base to add these features
|
|
现在对没有做过OCR处理的PDF分块会报错，希望对于这种文档可以内置一个ocrmypdf做下预处理 |
Now there will be an error for PDF chunking that has not been processed by OCR. I hope that for this kind of document, you can build a ocrmypdf to pre-process it. |
If the knowledge base contains pictures, can the pictures be output when the answer is output? |
|
知识库里面，批量新增几百个文件，然后批量向量化，会存在某些文件向量化失败的情况。 能否在文件列表做个筛选，区分已经向量化和未向量化的文件，这样可以快速筛选失败的再次批量向量化 |
In the knowledge base, if hundreds of new files are added in batches and then batch vectorization is used, some file vectorization will fail. Can you filter the file list to distinguish between vectored and unvectored files, so that you can quickly filter failed batch vectorization again |
感谢，看到你这个评论才知道为啥一直分块不成功 |
Thank you, I only know why I have been failing to divide the pieces after seeing this comment |
At present, our knowledge base still relies on traditional similarity search methods, which have certain limitations in handling complex queries and semantic understanding. In order to further improve the performance of the knowledge base and the accuracy of content replies, I suggest considering the introduction of knowledge graph technology. Specifically, consider integrating the following projects into our knowledge base features: LightRAG: Expected results: Improve accuracy: Through the introduction of knowledge graphs, the system can more accurately understand the semantics of user queries and provide more accurate responses. Implementation Suggestions: Technical evaluation: A technical evaluation of LightRAG and KAG is carried out to determine their compatibility with existing systems. I hope the team can consider this suggestion and jointly promote the optimization and upgrading of knowledge base functions. |
|
目前，我们的知识库仍然依赖于传统的相似度检索方法，这在处理复杂查询和语义理解方面存在一定的局限性。为了进一步提升知识库的性能和内容回复的精准度，我建议考虑引入知识图谱技术。 具体而言，可以考虑将以下项目集成到我们的知识库功能中： LightRAG： 预期效果： 提升精准度：通过知识图谱的引入，系统能更准确地理解用户查询的语义，提供更精准的回复。 实施建议： 技术评估：对LightRAG和KAG进行技术评估，确定其与现有系统的兼容性。 希望团队能考虑这一建议，共同推动知识库功能的优化和升级 |
At present, our knowledge base still relies on traditional similarity search methods, which have certain limitations in handling complex queries and semantic understanding. In order to further improve the performance of the knowledge base and the accuracy of content replies, I suggest considering the introduction of knowledge graph technology. Specifically, consider integrating the following projects into our knowledge base features: LightRAG: Expected results: Improve accuracy: Through the introduction of knowledge graphs, the system can more accurately understand the semantics of user queries and provide more accurate responses. Implementation Suggestions: Technical evaluation: A technical evaluation of LightRAG and KAG is carried out to determine their compatibility with existing systems. I hope the team can consider this suggestion and jointly promote the optimization and upgrading of knowledge base functions. |
|
I hope that we can support direct file submission for capable multimodal models. The files contain PDFs or videos and other stuff. Better if there's a tickbox "Send files without chunking" in the settings for respective models. 如果可以的话希望支持直接上传相关文件，特别是pdf文件。如果可以在模型设置里面对每个模型单独设置这个就更好了。 |
|
希望有一个快捷功能，类似腾讯 IMA 的，快速把对话存储为笔记文件，然后可以选择手动或自动转为知识库~~另外要是能编辑笔记就更好了 |
I hope there is a shortcut function similar to Tencent IMA, which can quickly store the conversation as a note file, and then you can choose to manually or automatically convert it to a knowledge base~~ In addition, it would be better if you can edit the notes. |
|
多个文件一起提问，识别不到所有文件啊，只能识别部分文件的分块来回答。分开单个文件提问是没问题的 |
Ask multiple files together, but all files cannot be identified. Only some files can be identified to answer in chunks. It's OK to ask questions separately in a single file |
场景描述环境配置
当前痛点
功能建议
附加信息当前文件管理界面示例： 优先级说明若该需求当前优先级较低，我会在知识库重构后，尝试自己实现 最后：感谢owner开发了lobe，对我们AI重度用户很有帮助！！ |
Scene DescriptionEnvironment Configuration
Current pain points
Functional Suggestions
Additional InformationExample of the current file management interface: Priority DescriptionIf the current priority of this requirement is low, I will try to implement it myself after refactoring the knowledge base. Finally: Thanks to owner for developing lobe, which is very helpful to our heavy AI users! ! |
赞同，如果上传图片时可以自动按对话的助手名称加日期进行命名就更好了，不然全是image头有点大，还不能缩略图查看 |
Agree, it would be better if you can automatically name the assistant name of the conversation and the date when uploading the image. Otherwise, the image header is a bit big and you can't view the thumbnail image. |
|
建议支持自定义向量化工具 |
赞同，现在有一堆和RAG相关的BUG，希望在2.的时候批量解决一下 |
Agree, there are now a bunch of RAG-related bugs, I hope it will be solved in batches in 2. |
|
🥰 需求描述 🧐 解决方案 📝 补充信息 |
🥰 Requirement description 🧐 Solution 📝 Supplementary information |
|
希望可以支持索引 web 数据到知识库中 原因：Cursor 有一个很好的功能是 @docs 用于在文档中搜寻，而 docs 可以通过输入一个文档的入口页面来自动索引，希望 lobechat 也可以支持这个功能 |
Hopefully, it can support indexing web data into the knowledge base Reason: A good function of Cursor is that @docs is used to search in documents, and docs can be automatically indexed by entering the entry page of a document. Hopefully lobechat can also support this function |
|
希望作者支持对数据库分类，支持二次重排. 
|
|
请问知识库2.0的开发现在是什么进度了？有时间表吗 |
What is the current progress in the development of Knowledge Base 2.0? Is there a timetable |
|
最近几天很火的gpt-4o Visualize ideas and concepts功能希望能接入进来 |
|
建议可以手动指定嵌入模型和从排序模型，配置文件方式不方便，并且覆盖模型提供商有限 |
It is recommended that you can manually specify the embedded model and sort the model, the configuration file method is inconvenient, and the coverage model provider is limited. |
and others
🥰 需求描述
一般认为知识库是主要利用rag技术，与直接把文件作为上下文相对应。
我的设想是为知识库新增一个功能：在对话中选择附加文件时，可以选择直接把已经在知识库里的‘文件’作为上下文，当然也保留把‘知识库’作为上下文。相当于为知识库增加一个云盘的功能，让里面的文件增加一个快速调用的方式。
🧐 解决方案
在对话中选择附加文件时，可以选择直接把已经在知识库里的‘文件’作为上下文
📝 补充信息
No response
@arvinxx : 借该 issue 召集下大家的诉求，如果有对知识库目前不满意的地方，欢迎提出来，3月份开始做知识库 2.0 改造，你的每个诉求我都会看到
The text was updated successfully, but these errors were encountered: