How to enable analysis of Excel data

[jasongilman]

I'd like to enable users to upload spreadsheets that have multiple tabs and information and be able to ask questions that could include analysis of the underlying data. Right now uploading an excel file and asking questions about it doesn't produce good answers. I'm assuming this is because the RAG API is generating vector embeddings for chunks of the excel file converted to text which probably wouldn't produce good results if you see part of the spreadsheet but miss the column headers.

I'm looking for different options to enable this including custom development an MCP server if that might work. I'm not sure if Agents can send complete files to MCP servers though.

I saw some other discussions on things like this but nothing that got completely at what I wanted.

• Generate Excel file? #5171
• [Enhancement]: full context controllable per session #6605
• [Question]: csv and excel filetypes support #1965 - This seemed the most promising. Would this provide the functionality?

[dvejsada]

Why not upload as text rather than for a file search?

[jasongilman]

Can you expand on what you mean?

• If you mean copy and paste the excel contents into the prompt I think that's going to lose a lot of the layout information and probably isn't the best way for an LLM to interrogate tabular data. It also won't scale to very large numbers of rows efficiently
• If you mean convert to CSV and attach it, my impression of how that works is it chunks the CSV and indexes it in a vector db. It's going. Vector similarity to find only some of the rows of the database will be missing headers and can't do things like aggregate data across columns.

I think what I really need is an MCP server which can handle these files but it requires sending the full file from Librechat to the server which isn't supported yet.

[jasongilman]

This is the issue that requests allowing a full file to be sent to an MCP server #8060

[usnavy13]

@jasongilman is there a reason your not using code interpreter for this? We use it for this exact usecase.

[jasongilman]

Wow, that's great. I didn't know it would work for that. Can you give me an example of how that works? We don't have that enabled yet but it's on our roadmap so perhaps we'll prioritize that.