A document assistant built around retrieval, not guessing.
A conversational document tool that parses a file, retrieves relevant passages, and carries context across follow-up questions.
Core stack
RAG
retrieval to feedback
01 / Where it started
The problem I was trying to solve.
A useful document assistant has to do more than accept a prompt. It needs to handle different file types, find the right context quickly, remember the conversation, and make it possible to learn from weak answers.
02 / What I built
My approach.
The application extracts text from PDF, DOCX, and TXT files, breaks it into manageable chunks, caches embeddings, and stores them in ChromaDB. Queries retrieve the closest passages before Google PaLM generates a response. Firebase keeps conversation history and feedback.
03 / Architecture
How the pieces fit together.
Read
Extract PDF, DOCX, or TXT content
04 / What changed
What came out of it.
- 01
Handled PDF, DOCX, and plain-text documents
- 02
Reused cached embeddings instead of recomputing them
- 03
Stored conversations and positive or negative feedback in Firebase
What I learned
Building the full loop changed how I thought about RAG. Parsing, chunk size, persistence, and feedback can matter just as much as the generation model.