ServiceFinder

The problem

The facts a student needs before going to a sports facility are spread over several Centre for Sports and Exercise (CSE) pages and PDFs: opening hours, who may use a facility, how to book, and when it is closed. Many of them depend on conditions, such as a check-in deadline, a student ID requirement, a seasonal timetable or a weekly maintenance break.

A general chatbot may answer fluently but drop those conditions or invent them. A wrong answer here means a wasted trip. ServiceFinder answers only from a small set of records that were checked against the official sources, keeps the conditions, and shows where each fact comes from.

What the system does

The student selects one facility, types a question and clicks Ask. The facilities are tennis, table tennis, swimming, the running track and the multipurpose outdoor court.

For the questions it cannot answer, it says so and links the official booking and closures pages.

How it works

1. The student selects a facility and asks a question.
2. Safety checks refuse password, personal-data and rule-bypass requests. Live or date-specific questions are referred to the official pages.
3. Facility aliases and keyword matching select up to four records from one JSON file. Hours questions always include the closure reminder.
4. A conflict check stops the answer if the selected records disagree. A length check keeps the prompt within a fixed budget.
5. One model call turns the records into a short answer with record IDs.
6. The app rejects any answer that cites an ID it did not retrieve, then shows the answer with its supporting records and official links. If the API fails, it shows the records as a labelled non-AI fallback.

The records are 20 short facts written in our own words from nine public CSE pages and documents, each with its source link and a check date of 2026-10-22. The app is one Streamlit page and plain Python functions. It uses keyword and alias matching, not embeddings, and no training. See the data documentation for sources, reuse basis and known gaps.

What we built and what we reused

The domain customisation is in the team-built parts: the checked records, the sports-facility aliases and keywords, and the rules that keep conditions and closure reminders in every answer.

Demonstration

A real answer from the system

Facility Swimming, question "When is the swimming pool open?". This is evaluation case N03 from a run on 2026-10-23 with deepseek-flash. The text is shortened.

Safeguards in action

These are real outputs from the same run. None of them made a model call.

Screenshots

Taken from the running app on 2026-11-10 with deepseek-v4.1-flash and prompt p1. The wording of model answers changes between runs.

Run it locally

You need Python 3.12 (the tested version), a web browser, and an API key for DeepSeek or any OpenAI-compatible chat endpoint. The page does not run the app itself: the app is a Streamlit program that you start on your own machine.

1. Get the code and set up

Clone the repository, open a terminal in its folder, then:

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env

py -m venv .venv
.venv\Scripts\Activate.ps1
pip install -r requirements.txt
copy .env.example .env

2. Add your key

Open .env and fill in:

.env is git-ignored. Never commit it or show it on screen.

3. Start the app

streamlit run src/app.py

Streamlit opens the page in your browser, usually at http://localhost:8501. Select a facility, type a question and click Ask.

What a successful run looks like

Select Swimming and ask "When is the swimming pool open?". A correct answer:

• is a short English answer based only on the records;
• cites record IDs in square brackets, for example [SWM-HOURS-01];
• keeps the closure reminder [ALL-CLOSURE-01] and shows the official closures link;
• lists the supporting records with their official CSE links under Supporting records;
• shows the status label "Answer from records" and the model ID in the footer.

The model words the answer differently on each run. The facts, record IDs, closure reminder and link should stay the same.

Check the safeguards too

Run the evaluation

python evaluation/run_eval.py                  # all 16 cases, once
python evaluation/run_eval.py --cases N01,MC02 # rerun selected cases

Each run is saved in evaluation/results/run_<timestamp>/ with the questions, retrieved and cited record IDs, raw answers, a score sheet and the model, prompt and data versions.

Evidence

We wrote 16 synthetic evaluation cases before running them: 8 normal, 4 with missing or conflicting information, 2 prohibited requests and 2 harder retrieval or ambiguity cases. The target from our proposal was at least 7 of 8 normal cases and all 8 safeguard cases.

How to read this result:

• It is one run of 16 synthetic questions, not a benchmark. It shows the pipeline works on these cases, not that it works on every phrasing.
• An earlier run on the first versions of the case files scored 12 of 16. We then fixed defects in the case files and runner, for example a missing record ID, an impossible expected status and fixtures that turned the wrong-source case into a conflict. The second run is on the corrected cases. Both runs are kept.
• Each claim in the eight normal answers was compared with its cited record. The judging was prepared with AI assistance and then reviewed and confirmed by a team member.

Full results: evaluation README and the saved run folder.

Limitations and a failure case

• No live availability, date-specific opening, bookings or account access. These questions are referred to the official pages.
• Covers five facilities at one centre, in English only.
• The records were checked on 2026-10-22 and can go out of date.
• Keyword lookup can miss unusual phrasings. Usually the app then asks for clarification instead of guessing, but it can also retrieve the wrong record, as in the failure case below.
• Two flaws seen in the second run that did not fail a case: hours answers often repeat the closure information that the app then appends as the closure reminder, and the answer to "When does the pool open?" gave opening times only, although the cited record also has the closing times.

A real failure: the floodlight fee

Facility Multipurpose outdoor court, question "How much is the floodlight fee for the basketball court?". Run on 2026-11-10 with deepseek-v4.1-flash.

• What happened. The app retrieved only OUT-ACCESS-01. The answer said "No fee information is available in the evidence, so I cannot give an amount", under the label "Answer from records".
• What is in the data. OUT-HOURS-01 states HK$30 per hour for the multi-purpose court at night (from the CSE FAQ), and OUT-BOOKING-01 says a floodlight fee may apply from 18:30.
• Why. The topic keywords map "fee" to the access topic only, while fee facts sit in hours and booking records. Retrieval filters by topic before matching words, so the right records were never considered.
• What it shows. The safeguards worked as designed: the model did not invent a number and cited a record it was given. The failure is in retrieval, and the answer still misleads a student into thinking there is no fee. A similar question for tennis ("How much do the floodlights cost at night?") matched no topic at all and was sent to clarification. A fix would add money words to the keyword map and let fee questions search booking and hours records. It has not been made.

Both runs and their screenshots are saved in the probe folder. They are not part of the 16 scored cases.

Ethics and AI use

• Sources. All records are short summaries in our own words from public CSE pages. No pages, tables, images or documents were copied. Each record keeps its official link.
• Privacy. The records contain no personal data. The app refuses passwords and personal details, and all test questions are synthetic.
• Harm from a wrong answer. A wrong time or condition wastes a trip. The app answers only from retrieved records, keeps conditions and closure reminders, rejects uncited IDs, declines when records conflict, and sends date-specific and live questions to the official pages.
• Honest labelling. A non-AI fallback is labelled as such and is never counted as a model result. The page states what was run, with which model, on which date.
• AI in the app. One call per question to a DeepSeek model (or any OpenAI-compatible model you configure), with prompt version p1.
• AI in development. AI coding assistants helped to draft code, review it, fix evaluation cases and prepare judging. Each use, with the human check that followed, is listed in the AI use record.

Team

AILT9019 Team 53, The University of Hong Kong.

• LI, Zhuofei
• KAU, Yat Long
• MAHMUD, Nafis Bin
• SUN, Mingrui