Multiple sources, one index
Document loaders ingest local PDFs, website URLs, and Google Drive files. Recursive splitting uses 1,000-character chunks with 200-character overlap.
Document retrieval assistant.
Answers grounded in your documents.
A document-aware assistant connecting local PDFs, websites, and Google Drive to a retrieval pipeline with conversation history.
01 / THE PROBLEM
The system challenge.
A language model does not know your documents by default. This project brings heterogeneous document sources into a searchable index and adds retrieved context to a conversation.
02 / SYSTEM FLOW
Trace the architecture.
Select a stage to see its responsibility.Conceptual flow · based on the documented implementation
STAGE 01
Load local PDFs, website URLs, and Google Drive documents into a common retrieval pipeline.
03 / IMPLEMENTATION
Decisions that matter.
The details that make this an engineering system.
Avoid rebuilding unchanged content
A document hash detects changes before rebuilding the FAISS vector store. An unchanged corpus loads the saved index.
Retrieval with conversation context
The backend retrieves relevant chunks and combines them with conversation memory. Per-chat JSON files persist history across interactions.
A separated interface and API
FastAPI handles chat creation, history, and queries. Streamlit supplies the user interface. This is a public prototype, without published production benchmarks.
04 / OUTCOME & EVIDENCE
What the work demonstrates.
A working retrieval-oriented implementation is available for code review. No accuracy, latency, or scale claims are made.
Review the implementation on GitHubCONTINUE EXPLORINGMulti Agent System
All selected work