PakDigit Markets · Rehnuma, the markets assistant
Check every number before it is shown
Visitors ask about today's dollar, gold and petrol rates, Shariah status and zakat. A pipeline answers from live data first, then from our guides, and only then asks a model, whose every number is checked against our data.
The business need
Every number must be exact and sourced. The assistant must never advise, predict prices or give its own religious rulings, and it must work in Urdu and English, by voice too.
~1,100
guide sections searched by meaning
4
ways an answer is labelled: live data, guides, AI (checked), policy
0
AI-written numbers shown that are not in our data
How one request flows
From the visitor's words to what they see, step by step.
-
1
Policy guard
Advice, predictions and off-policy topics get a fixed answer. No AI, no cost.
-
2
Instant data answer
A simple rate or status question is answered straight from the database.
-
3
Search our guides
About 1,100 embedded guide sections are searched by meaning.
-
4
Model
The model explains, using today's facts and the guide notes it was given.
-
5
Check
Every number in the reply must exist in our data, under the right label.
-
6
Show the source
A badge says how it was answered: live data, our guides, AI (checked) or policy.
The key ideas
A pipeline, not one prompt
Most questions never reach the model. Rates and statuses come straight from data; only questions that need explaining do.
Code writes the sensitive sentence
A company's Shariah status is written by code from the official list, with its date. The model can never soften or invent a verdict.
Number grounding
After the model replies, every figure is compared with today's facts. An unsupported number means the answer is not shown, and the next provider is tried.
Right number, right label
A number check proves a figure exists in our data, not that it belongs to the right thing. So the check also looks at the label next to each number.
"How I answered" badge
Visitors see whether a reply came from live data, our guides, the AI (checked) or a fixed policy answer, so an AI sentence is never mistaken for an official figure.
Voice in two languages
The browser records the question, converts it to 16 kHz audio and our own speech-to-text model transcribes Urdu or English. Audio is not stored.
Cheap checks first
A short, cached health check means a sleeping model costs a visitor at most a second and a half, once, not a 20-second wait on every question.
Privacy by default
Question text is not stored. The database keeps only daily counts: asked, answered, refused, helpful.
What testing taught us
Found by running the real model, not by guessing. This is the most useful part.
Both numbers were real; the sentence was false
The AI wrote "interbank Rs 276.51, no change from Rs 278.35", mixing two different markets. Each number existed in our data, so a plain number check passed it. The daily Market Brief is now written from templates, and the check also matches labels.
Know when a template beats a model
For a text that is mostly numbers and comparisons, a template is always correct and the AI adds little. Use the model where words matter, and code where facts matter.
A guard can be too strict
"Zakat tips" was once refused as financial advice, and "why does gold rise?" as a prediction. The guard now looks at what is being asked about, not just a single word.
A rejected answer deserves a second opinion
When one model's reply failed the check, the visitor used to get the offline answer. Now the next provider gets a chance first.
What you learn at this level
- Combining data look-ups with RAG
- Checking model output against facts
- Citing sources
- Speech-to-text
- When not to use AI
Where it stops
Very small numbers (one or two digits) and years are not checked yet, because they appear too often in ordinary text.
AI Playground
Change chunk size, similarity threshold, model and prompt on our case studies, and see what was retrieved, the exact prompt, the answer, the speed and the cost.
Open the Playground →Live workshops
Hands-on sessions that build each level step by step, hosted by our founder and by experts on SellYourLogic.