Document Q&A application
RAG Knowledge System
Ask the documents. Follow the evidence.
LangChain · FAISS · Vision models
Edited walkthrough of recorded interface states. Retrieval results do not establish general answer accuracy.
Answers with a way back.
Page and slide references connect the answer to the retrieved source material.
View detail ↗
A useful boundary.
The captured refusal shows a case where the documents did not provide the requested answer.
View detail ↗My contribution, evidence and current limitations +
The ingestion pipeline reads text, tables, and visual content, then indexes passages with FAISS. Retrieved evidence is passed to an answering model with page and slide references. Incremental indexing avoids rebuilding unchanged documents.
A saved controlled comparison increased top-five retrieval from 81.8% to 95.5% on 22 answerable questions when visual content was included. The later corrected evaluation reports hit@1 of 68.2% and MRR of 0.787.
The evaluation measures retrieval on a small authored test set, not answer accuracy or citation faithfulness. The corrected top-five result also reflects an answer-key change. This application uses external AI services.
The next evaluation should measure generated answers, citation support, and refusals on unfamiliar documents.
