Skip to main content
SupaCortex
Sign in Book a Demo
Feature

An answer you can check, or an honest "I don't know"

A confident wrong answer costs more than no answer. SupaCortex answers only from your documents, cites every claim, checks the citations before anyone reads them, and says so when the documents do not cover the question.

Book a demo See pricing
  • Hybrid search
  • A confidence gate
  • Citations checked and repaired

How does SupaCortex keep its answers grounded?

Every question runs a meaning-based search and a keyword search together, merges the results and weights them by how trusted each source is. If nothing relevant enough comes back, a confidence gate stops there and the reply says the knowledge does not cover it - the model is never asked. Otherwise the model writes from the numbered passages it was given. Its citations are repaired, then validated: each must point to real evidence, and every number and link must appear in what it cites. An answer that fails is regenerated once, and withheld if it fails again.

  • Meaning and keyword search merged, weighted by trust tier.
  • Past-ticket content is dropped when enough official evidence exists.
  • Three outcomes: answer found, needs information, insufficient evidence.
  • Numbered citations with document, section, passage and link.
  • Numbers and links in the answer must appear in the cited evidence.
  • Customer-tagged content is only searched for that customer.

What an answer looks like

Can I carry forward unused leave into next year?

Yes, up to 15 days of earned leave can be carried forward into the next calendar year 1. Anything above 15 days lapses on 31 December unless your manager approves an encashment request before then 2.

1 Leave Policy 2026 › Carry-forward of earned leave
2 Leave Policy 2026 › Encashment

An illustration. Every number in the answer must appear in the passage it cites, or the answer is not shown.

How it works

Four checks between a question and an answer

  1. 01

    Search two ways

    By meaning and by exact words, then merged. Official documents get extra weight; past tickets get less.

  2. 02

    Gate on confidence

    Too little relevant evidence? The reply says so, and no model is called.

  3. 03

    Write from evidence

    The model sees only the numbered passages and must cite them. Personal details in the question are redacted first.

  4. 04

    Validate, or withhold

    Citations repaired and checked, numbers and links verified. One retry; then an honest fallback.

In detail

What the checks actually check

Citations that point somewhere real

A citation is only useful if the passage it names says what the sentence claims.

  • Every citation number must exist in the evidence the model was given.
  • An answer that claims to have found something must cite something.
  • Every number in a sentence must appear in the passage that sentence cites, or in the question.
  • Every link must come from the evidence - the model cannot invent a URL.
  • Loose citations are tidied: odd formats fixed, a weakly related extra citation dropped, an uncited sentence matched to its passage.

Every answer leaves a record

When somebody asks why it said that, the evidence is there to look at.

  • Each answer is logged with the passages used, their scores and the citation each became.
  • Declined questions are logged too, so missing documents show up as gaps you can fill.
  • Questions are stored with emails, phone numbers and ID numbers redacted.
  • Thumbs up and down on answers in the widget and team chat.

Related

What Is Retrieval-Augmented Generation?

How RAG works, why it beats fine-tuning for company knowledge, and where it goes wrong.

Measuring AI Answer Quality

Golden question sets, groundedness, deflection and the numbers that actually predict trust.

Answer Quality & Usage

Thumbs up and down on every answer, conversation logs, and AI credits you can see being spent.

Website & In-App Chat Widget

One script tag for your site, a WebView URL for your mobile app - public or signed-in users.

SupaCortex vs ChatGPT for Teams

A brilliant general assistant, against one that only answers from what your company wrote.

Banking & Financial Services

Answers that cite the approved product document, and say so when there is none.
Frequently asked

Questions people ask before they start

Yes - no system that writes text can promise otherwise. What SupaCortex does is make a wrong answer visible and rarer: every statement cites the passage it relies on, numbers and links are checked against those passages, and when the evidence is weak it declines rather than filling the gap.

That it could not find this in the knowledge it has, in plain words. No model is asked to improvise. If the question is ambiguous, it may instead ask for the one detail it needs.

Two searches at once: one by meaning, which finds "how do I claim travel" in a page titled "Expense reimbursement", and one by exact words, which finds "error E101" when meaning alone would not. The two lists are merged, and official documents are weighted above community content.

Open models, run on infrastructure we operate. Your documents and questions are not sent to a third-party AI API such as OpenAI, and nothing is trained on your data.

Because it is checked before it is shown. Streaming words as they are written would mean showing text that might later fail validation, so the interface shows each stage - searching, writing, checking - and then the whole, checked answer.

A search on its own uses none. An answer uses one credit; a chat turn that also calls your systems uses two.

Ask it the question your current chatbot gets wrong.

Book a 30-minute demo on your own documents. We set up a 14-day trial workspace for you - no credit card, no setup fee.

Book a demo See pricing

Questions first? Email info@greatdevelopers.com or call +91 8585953085.

Still curious? Ask your favorite AI about SupaCortex.

ChatGPT Claude Perplexity Grok