Case Study · Artificial Intelligence
AI Document Q&A Assistant
A retrieval-augmented generation assistant that ingests an organisation's documents and answers plain-English questions with citations back to the source.
- Industry
- Artificial Intelligence
- Timeline
- 2025
- What we did
- AI & RAG engineering · Vector search · Backend architecture
The Challenge
What the client was up against
The client's institutional knowledge was locked inside hundreds of PDFs. Staff either couldn't find answers or asked a colleague, which didn't scale. A general-purpose chatbot was the obvious idea and the wrong one: it knew nothing about their documents and would confidently invent answers, which in their domain was worse than no answer at all.
Our Approach
How we solved it
Retrieval first, generation second
Documents are chunked with overlap, embedded, and stored in a vector database. A question retrieves the most semantically relevant passages, and only those passages are handed to the model as context. The model's job is to answer from the retrieved text, not from its own memory.
Cite, or say you don't know
Every answer carries citations back to the source passages, so a user can verify a claim in one click. When retrieval turns up nothing relevant, the assistant says so rather than filling the silence — the single most important behaviour in a system people are meant to trust.
Make ingestion boring
Uploading a document kicks off a background pipeline: parse, chunk, embed, index. It is observable and re-runnable, so when the chunking strategy improves, the whole corpus can be reprocessed without anyone touching the database by hand.
What We Built
Inside the solution
- PDF and document ingestion with an automated chunk-and-embed pipeline
- Vector similarity search over a private corpus
- Context-aware, conversational natural-language answers
- Citations linking every answer back to its source passage
- Explicit 'no relevant source found' handling instead of guessing
- Re-runnable ingestion so the corpus can be reprocessed on demand
Technology
- NestJS
- React
- PostgreSQL
- Qdrant
- LangChain
- LLM APIs
The Outcome
Grounded, citation-backed answers over a private document corpus — with no invented sources.
- Cited
- Every answer traceable to a source passage
- Private
- Corpus stays inside the client's own infrastructure
- Minutes
- From document upload to being queryable
The hard part of a RAG system is not the model. It is retrieval quality, and knowing when to say 'I don't know'.
More Work
Other case studies
Different industries, same approach: understand the problem first, then build only what solves it.
Facing a Similar Problem?
Tell us what you're working on and we'll tell you how we'd approach it — no obligation.