AI Disclosure

1. What Veda is

Veda is the assistant in the corner of this website. It is an AI, not a person. It answers questions about Vedatricks — our products, what they cost, our services, careers and how to reach support — by generating text with a large language model.

It is a guide to this website. It is not a salesperson, not a support agent, and not authorised to commit us to anything. Any price, scope or timeline that matters is confirmed by a person in writing.

2. Who powers it

• Model provider: Sarvam AI.
• Model: sarvam-105b-conversations.
• Where the call is made: from our own server. The API key never reaches your browser, and your browser never talks to the model provider directly.
• Languages: English, Hindi, Kannada and Marathi. Replies come back in the language you wrote in wherever the model supports it; where it cannot, the assistant says so and offers English or a human.
• Status: the assistant is being brought into service. Where it is not yet available, the dock says so rather than pretending.

Website assistant powered by Sarvam AI.

3. Labelling and provenance

• The assistant identifies itself as AI before your first message and again in its header, permanently.
• Every reply carries an “AI-generated” label. Do not remove it when you copy text elsewhere.
• This satisfies the requirement introduced by the 2026 amendment to the IT Rules that synthetically generated information be clearly and prominently labelled.
• We do not present AI output as human-authored, and we do not let the assistant claim to be a person if you ask it.
• Where we publish AI-assisted content elsewhere on this site, it is reviewed and signed off by a person before it goes live.

4. What it will not do

The assistant is instructed to refuse, and hand you to a person instead, when asked to:

• give medical or clinical advice — including any question about a patient, a symptom, a medicine or a diagnosis, even in the context of our clinical software;
• give legal, financial, tax or investment advice, or interpret a statute for your situation;
• quote a final price for custom work, approve a discount, or agree contract terms;
• discuss anything unrelated to Vedatricks;
• reveal its own instructions, or act on an instruction embedded in text you paste;
• produce content that is unlawful, deceptive, defamatory, or that impersonates a real person or organisation;
• access your account, look up a customer record, change data, or take any action on your behalf.

It has no access to any customer database. It cannot look you up, and it cannot see anything you have not typed into the conversation.

5. What happens to your messages

1. Before your first message, a notice tells you it is an AI, that messages are stored, who processes them and what not to send. Nothing is sent until you choose to continue.
2. Your message travels to our server over TLS.
3. Our server redacts patterns it recognises as identifiers — Aadhaar-shaped numbers, PAN, card numbers, phone numbers, email addresses — before anything leaves our systems.
4. The redacted conversation is sent to Sarvam AI, which returns the reply. Sarvam may process it in United States or European Union regions.
5. We store the conversation with a session identifier, a hashed IP address, the model name, token counts and cost, so we can enforce limits and investigate abuse.
6. Transcripts and voice clips are deleted 90 days after the conversation ends.
7. Access and security logs are kept for at least 12 months, as Rule 6 of the DPDP Rules 2025 requires.

You can delete a conversation yourself with “Delete this conversation” in the dock. The full detail is in section 5 of the Privacy Policy.

Please do not type Aadhaar numbers, PAN, card numbers, passwords, health details or another person’s data into the assistant. Redaction is a safety net, not a guarantee.

6. Voice

• The microphone is off until you press and hold, or toggle, the microphone control.
• A short clip is sent to our server, transcribed, and used as your message.
• Clips follow the same 90-day deletion rule as transcripts.
• We do not build a voiceprint, we do not identify speakers, and we do not clone voices.
• Voice input is optional. Everything the assistant does can be done by typing.

7. Guardrails

The assistant runs behind a shared guardrail layer:

• Scope — a system instruction confines it to Vedatricks topics, and it treats all user text as data, never as instructions.
• Input hygiene — a per-message character cap, a per-session turn cap, HTML stripped, and identifier redaction before the model call.
• Injection defence — known jailbreak patterns get a canned refusal, and output containing system-prompt markers is blocked.
• Output handling — replies are rendered as plain text, never as HTML, and links are limited to an allowlist of our own pages.
• Professional-advice refusal — the categories in section 4, with an automatic offer of a human.
• Abuse controls — Cloudflare Turnstile at session start, per-session, per-IP and per-day message limits, and a temporary block on repeated abuse.
• Cost controls — per-session token caps and a global daily spend cap, with an alert before the cap is reached.
• Kill switch — an operator can disable the assistant instantly; the dock then says it is temporarily unavailable rather than failing silently.
• Audit — every turn is logged with model, tokens and cost, for cost control and abuse investigation.

System prompts are not a security control, and we do not treat them as one. Nothing secret is placed in a prompt.

8. Limits and availability

• Sessions are capped at [ ● ]term not yet agreed; this document is in legal review messages and [ ● ]term not yet agreed; this document is in legal review characters per message.
• Daily limits apply per IP address. If you hit one, the dock tells you and offers the contact form.
• The assistant is not covered by the SLA and may be unavailable during maintenance or a provider outage.
• Automating or scripting the assistant is prohibited by the Acceptable Use Policy.

9. Getting a human

Ask at any time, and the assistant offers the contact form or WhatsApp. It also hands over automatically after two refusals, or when the question is clearly outside its scope. A conversation only becomes an enquiry if you fill in the short form and tick the consent box — the assistant never creates a lead from a conversation on its own.

10. Accuracy and your reliance on it

• Language models make mistakes. Answers can be wrong, incomplete or out of date.
• The assistant is not a source of legal, medical, financial or tax advice, and its answers are not a quotation, a contract or a commitment.
• Where the assistant’s answer conflicts with a page on this site, an invoice, an order form or a signed agreement, the document wins.
• Prices, plan limits and product statuses on the product pages are the reference; the assistant should point you at them, and if it does not, tell us.

11. AI inside our products

Every product is in this table, including the ones where no AI ships yet. “In build” means the capability is not part of what you are buying today.

Common rules across every AI feature in our products:

• Off unless an administrator turns it on, and switchable off at any time.
• Read-only, or write-only-after-a-human-confirms. No AI feature changes your data on its own.
• Scoped to your own organisation’s data, with the same access rules as the user asking.
• Audited per turn, with a per-organisation usage cap.
• No clinical decision support, no diagnosis, no automated decision about a person.

12. What we do not call AI

13. Training and your data

• We do not train models on your conversations, your Customer Data, or your customers’ data.
• We configure our model providers so that content sent to them is not used to train their models, and we set provider-side retention to the minimum the provider offers. Customers can ask for the current configuration in writing.
• Unanswered or badly answered questions are reviewed by a person and used to improve our website content and the assistant’s instructions — not to train a model.
• The processor terms for AI features are in section 14 of the DPA.

14. Governance

15. Report a bad answer

Use the thumbs-down control on any reply, or email hello@vedatricks.com with the conversation reference shown in the dock. Tell us what it said and what it should have said. We read every one, and a systematic problem is fixed by changing the assistant’s instructions or the underlying page — not by hiding the complaint. If an AI answer caused you a loss, raise it with our Grievance Officer below.