AgentAI — Chat With Your PDF, With Sources
View completed workgithub.com
The problem
Long documents — reports, contracts, manuals — are slow to search by hand, and a generic chatbot will happily make up an answer that isn't in the file.
What I built
A document assistant you can put in front of users:
Answers grounded in your document. The PDF is split into passages and indexed; each question retrieves the most relevant passages and the model answers from them, streaming the answer as it is written.
A second opinion from the web. A web-search tool running over the Model Context Protocol (MCP) runs at the same time and gives a separate answer from the web. If it fails, the document answer still arrives.
Private per session. Each browser session keeps its own uploaded files and conversation; idle sessions and their files are removed after an hour. Uploads are PDF-only with a size limit.
Voice mode. Ask out loud and hear the answer.
AgentAI home screen
What this means for your project
The same pattern works for a help center, internal docs, a product catalogue or a database: your content in, grounded answers out, with a clear separation between what came from your data and what came from elsewhere.
Stack: React, Node.js / Express, LangChain, OpenAI, Server-Sent Events, MCP
Personal project.
0 likes · 3 views · Sep 26, 2026
Upload a PDF and ask questions about it: answers stream in grounded in the document, while a web-search tool over MCP adds a second answer from the web.
