01
Plain-language questions
Problem
Keyword search fails when you don't know the document's exact wording.
How we solved it
Questions are matched by meaning against embedded passages, not by keyword.
Case study · 17AI Knowledge Retrieval
Retrievabase.ai
Upload contracts, reports and manuals, then ask questions about them and get answers that point to the exact page.
Web app
- Client
- Retrievabase
- Industry
- AI · Knowledge Retrieval
- Platform
- Web
- Outcome
- Answers from your own files, each one cited to its source
01
Overview
Retrievabase.ai lets a team question its own documents the way they'd ask a colleague. Drop in PDFs, Word documents or plain text, ask in normal language, and each reply points to the spot in the file it was drawn from. We delivered the workspace, the document chat, file management, sign-in and subscription billing.
02
The challenge
Finding one clause in a 60-page contract, or one figure in a quarterly report, meant reading until you found it. Experienced people were spending their time searching rather than judging, and anything skimmed under deadline pressure was easy to miss.
03
Our approach
We wouldn't show an answer we couldn't trace. Retrieval comes first, every response carries its citation, and summaries always link back to the original pages.
04
How we built it
4 steps, from architecture to launch.
- 01
Ingest without cleanup
Uploaded PDF, DOCX and TXT files go into Firebase storage as they are and are indexed per workspace.
- 02
Retrieval, then generation
Passages are embedded as vectors, and the OpenAI API answers using only the most relevant ones.
- 03
Citations on every reply
Each answer links to the page and section it relied on.
- 04
Team-ready accounts
OTP sign-in, roles inside each workspace and Stripe plans with clear usage limits.
05
Problem → How we solved itTechnical challenges
02
Short summaries with a source link
Problem
Nobody wants to read 50 pages to decide whether a file matters.
How we solved it
Summaries reduce a long file to a few points, each traceable to the original.
03
One question across many files
Problem
Answers were often spread over several documents.
How we solved it
A single query searches the whole workspace and combines what it finds into one cited answer.
06
6 screensScreens
07
5 featuresWhat we shipped
- 01
Ask your documents
Chat with uploaded files and get cited answers in plain language.
- 02
Library
Bring in PDF, DOCX and TXT files, then filter, sort and organize them by workspace.
- 03
Conversation history
Start new sessions, reopen old ones or delete them.
- 04
Accounts and security
Sign-up with OTP, password reset, profile controls and one-step account deletion.
- 05
Plans and roles
Tiered Stripe billing and per-workspace permissions for uploading, asking and managing.
08
Tech stack
- React + Next.js
- Workspace front end
- TypeScript + Tailwind CSS
- Type-safe UI and styling
- OpenAI API + vector embeddings
- Finding passages and writing answers
- Stripe
- Plans, limits and upgrades
- Firebase
- Sign-in, file storage and data
Tell us what you are shipping.
Start a Project →