Cited answers from clinical textbooks

Ask a clinical question and get an answer drawn only from your own textbooks, with the book and page for every claim.

Project facts

Who it’s for
Clinicians who need a quick, trustworthy reference between patients, physiotherapists especially.
Stage

Stage 2 of 5: Working prototype

Lloyd’s role
Sole designer and developer
Year
2026
Built with
Next.js, React, Python, FastAPI, LangChain, Chroma, OpenAI

In short

The problem
Clinicians need quick, trustworthy reference answers mid-shift, and general AI tools answer confidently without showing their source.
What changes
Answers drawn only from the clinic’s own textbooks, with the book and page for every claim, and an honest refusal when the books don’t cover it.
Proof so far
A working prototype with 358 automated tests. Not yet validated with clinicians.
Clinical Reference presentationA short narrated presentation: the problem, the four steps behind each answer, how it’s built and what changes in a working day.2:30 · No captions yet

Evidence so far

  • 358Automated testsAcross the answer pipeline and the interface
  • Every claimSourcesBook and page, with the passage one tap away
  • RefusesWhen unsureA relevance threshold stops answers the books don’t support
  • Zero temperatureConsistencyThe model is set to give the same answer to the same question

The problem

Mid-shift, clinicians have quick reference questions and often no senior colleague free to ask. The answer is usually on the shelf, but finding it takes longer than there is.

General AI chat tools answer confidently whether or not they’re right, and don’t show where an answer came from. In clinical work, an unsourced answer can’t be trusted.

Who it’s for

  • Clinician

    Needs to
    Get a quick, trustworthy answer between patients, with a textbook page to check.
    Today
    A senior colleague is rarely free, and internet answers can be confident guesses with no source.
  • Library admin

    Needs to
    Add reference books that everyone’s answers draw on.
    Today
    Adding a book meant editing scripts and re-running a process by hand.
  • Invited user

    Needs to
    Sign in and keep their own question history across devices.
    Today
    Early on, history lived only in one browser and access relied on a shared key.

How it works

  1. Clean up the question

    Shorthand such as “rtr after acl??” is rewritten into a clear search query.

  2. Search the library

    It finds the passages closest in meaning, drawing on several books rather than one.

  3. Write from those passages only

    The answer uses the retrieved text and nothing else.

  4. Cite every claim

    Numbered references give the book and page. Tap one to see the passage, highlighted.

Screens

These images are slides from the project presentation. Screenshots of the running app are still to come.

  • Every answer shows its working
  • One question, four steps
  • When the books don’t cover it, it says so
  • How it’s built
  • From the desk to the bedside

Main features

  • An honest “I don’t know”

    If the books don’t cover the question, it says so instead of guessing.

  • Citations you can check

    Medical-style references with the book and page, and the source passage one tap away.

  • Private history

    Google sign-in, with each user’s conversations kept separately.

  • Diagrams in answers

    Flowcharts render inside the answer when a process is easier to see than read.

  • Works on a phone

    Built for a question asked at the bedside, not just at a desk.

  • A clear limit

    Every answer ends with a reminder that it’s reference material, not a substitute for clinical judgement.

Product decisions

  1. Numbered citations with a tidy reference list

    Instead of letting the AI write book titles and pages into the text.

    Cleaner to read, exact about where each claim came from, and formatted by code rather than trusted to the model.

  2. Draw on several books per answer

    Instead of plain similarity search, which clusters in one book.

    With four books and over 11,000 passages, an answer should reflect the whole relevant shelf.

  3. Invite-only Google sign-in

    Instead of a shared key visible in the browser.

    Anyone could have read the key and spent the AI budget; sign-in keeps the server as the single point of control and stores no passwords.

Left out on purpose

  • Open sign-up and password reset. It’s for invited, named users only.
  • Filtering by book category. Kept out to keep the multi-book change focused; the data to support it is already in place.

What changed along the way

  • Off-topic questions are now refused. Questions outside the library were getting answers built from loosely related text, which broke the “never guesses” promise.
  • Tappable citations that show the exact passage. Clinicians needed to check a source without leaving the answer.

How it saves time

  • Questions answered between patients rather than after hours.
  • Evidence with every answer, ready to check.
  • Less time searching through books for a page you half remember.

These are intended outcomes. No time savings have been measured.

How it’s built

Lloyd designed and built the assistant: the retrieval pipeline, the refusal rules, the citation interface and the test suites for both the server and the app.

Lloyd Dela Cruz, Founder, DYOL Digital Labs. About Lloyd

Interface
Next.js, React, TypeScript, shadcn/ui, streamed answers
Server
Python, FastAPI, LangChain, Google sign-in
Knowledge
Chroma vector database, OpenAI embeddings and GPT-4o, run at zero temperature
Quality
Automated tests for the server and the interface

Where it is now

  1. Concept
  2. Working prototypeThis project
  3. Pilot-ready
  4. Public beta
  5. In use

A working prototype that runs locally. It needs your own textbook library to answer from, and there is no public demo.

Still to do before real-world use

  • Validation with practising clinicians
  • A properly licensed textbook library
  • Secure hosting and sign-in for a clinic

What’s next

  1. Category filters for the library
  2. Answer-length caps to control cost
  3. Better handling of large diagrams

A free 20-minute conversation, then an honest assessment. You’ll know what it costs and how long it takes, in writing, at a fixed price.

Or email lloyd.delacruz@outlook.com. We reply within 24 hours.