AI & Automation
AI Document Assistant
A demonstration of retrieval-augmented generation over a document library: questions answered from your own files, with a citation to the exact source passage.
- Type
- Demo build
- Category
- AI & Automation
- Features
- 9 documented
- Stack
- 6 technologies
What the business was dealing with
An organisation holds hundreds of policy documents, product manuals and contracts across shared drives and inboxes. Staff either interrupt a senior colleague for an answer or guess. Keyword search returns a list of files rather than an answer, and nobody can tell whether a document they found is the current version.
How it was solved
A retrieval-augmented assistant indexes the document library with semantic chunking that keeps tables and clauses intact. A question retrieves the most relevant passages using hybrid keyword and vector search, and the model answers strictly from those passages — citing the document and page. When nothing relevant is found it says so rather than improvising, and questions it could not answer are reported as content gaps.
Interface
Illustrative screens
These are abstract representations of the interface layout, drawn in CSS — not screenshots of a client system. They show structure rather than data.
Main features
- Ingestion of PDF, Word, Excel and scanned documents with OCR
- Semantic chunking that preserves tables, clauses and section structure
- Hybrid retrieval combining keyword and vector similarity search
- Answers with citations to source document and page
- Explicit refusal to answer when retrieval finds nothing relevant
- Permission-aware retrieval honouring per-document access rules
- Conversational follow-up with retained context
- Admin view: ingestion status, unanswered questions, content gap report
- Scheduled re-indexing as documents change
What this demonstrates
The engineering problems this project exists to show we can handle.
- Grounding answers in retrieved content to reduce fabrication
- Citation handling so answers can be verified rather than trusted
- Permission-aware retrieval in a multi-user environment
- Honest uncertainty behaviour instead of a confident guess
Technology stack
- Claude API
- pgvector
- Python with FastAPI
- PostgreSQL
- OCR preprocessing
- Next.js
Services involved in this project
Want something like this for your business?
Your operational detail will differ — that is normally where the real work is. Tell us how your process runs and we will scope it honestly.