AI Automation
Talk to PDF Agent (Knowledge Tool)
Task decomposition
- User uploads a PDF through a file-upload form.
- PDF text is extracted and joined into a single document.
- Extracted text is embedded and inserted into an in-memory vector store, tagged with the source filename and upload timestamp.
- User asks a question via chat.
- AI Agent uses a knowledge tool to search that same vector store for relevant chunks.
- Agent answers grounded in what it found; if the answer isn’t in the document, it says so instead of guessing.
Objective of the workflow
- Lets you ask questions about a specific document instead of getting a generic answer from the model’s training data.
- The agent is instructed to search the uploaded file first and admit when something isn’t in there, rather than filling the gap with an unrelated answer.
Architecture
