Smart Document Search
A retrieval study: short answers that point back to the original file instead of free invention.
Project overview
The project is about finding the paragraph that already exists. The file set is mock. No private store is attached.
Problem being explored
People ask questions that already live in PDFs and guides, but keyword search misses them.
Example user scenario
Someone asks “Where is onboarding step 4?” The system should return a short answer plus a link to handbook.pdf and the original passage — not a new policy.
Proposed workflow
- FILES
- CHUNKING
- EMBEDDINGS
- VECTOR SEARCH
- ANSWER
- SOURCE DOCUMENT
Architecture explanation
FILES → chunking → embeddings → vector search → answer → source document. Document updates mean re-chunking changed files so the index does not drift.
Key features
- Ingestion and chunking of files
- Embeddings and vector retrieval
- Short answers tied to a passage
- A visible link to the source document
- A path to re-index when files change
Demo interface
The demo UI sketches a query and a sourced reply. It does not search your drive.
FAQ
Why attach the source document?
A visible link lets a person verify the passage and see when retrieval found nothing useful. Fluent text without a source can hide a miss. Sources also make evaluation possible before a real rollout.
How is this different from Ctrl+F?
Ctrl+F needs the same words. Vector search can match meaning when the question is phrased differently. Exact IDs, codes, and titles still work well as keywords, so a hybrid approach is often best.
What happens when a PDF is updated?
A working system can re-chunk that file, replace stale vectors, and keep answers on the current version. An index that never refreshes will keep citing last month’s PDF. Update jobs are part of the production design.
Can this run over confidential files?
Confidential corpora can be indexed with authentication and role-based access so people only retrieve what they may see. This DEMO / CONCEPT does not connect to confidential stores. Access control belongs in the real implementation, not as an afterthought.
Limitations
The file set is mock. No private document store is attached.
What production would require
Ingestion pipeline, embedding store, access control, and citation UX.