Skip to content
AskVault Docs
Try Free →

AI safety commitments

Written by Joseph Russo, Principal Architect, Networks and Tech, AskVault · Reviewed by Allison Chan
Last updated: · 5 min read

The five commitments

  1. Source-cited answers. Answers show the sources they were drawn from by default; a workspace can turn the source list off. RAG architecture grounds responses in your content.
  2. No customer-data training. AskVault does not train models on your content. Each model provider's terms decide whether it uses API traffic.
  3. Audience-tag enforcement. Sensitive content scoped to verified visitors only.
  4. Kill switches per skill. Hard caps the LLM cannot override.
  5. Mandatory escalation paths. Every conversation has a path to a human.

Source-citation as anti-hallucination

Our primary defense against hallucination:

  • Retrieval before generation. Bot retrieves relevant chunks first.
  • Generation grounded in those chunks via system prompt.
  • Sources shown with each answer by default, with a link where the page has a URL. This can be turned off per workspace.
  • Strict mode prompts to refuse if no chunks match.

Our design goal is to keep hallucinations low. Most "wrong" answers reflect knowledge gaps in your content, not fabrication. Verify high-stakes answers via the source citations.

No customer-data training

A written commitment:

  • Your content is not used by AskVault to train models.
  • Your conversations never used.
  • No training by AskVault. AskVault does not use your content or conversations to train models. Whether an LLM provider may use API traffic for its own training is set by that provider's terms.

No opt-in alternative exists. If we ever introduced one, it would be explicit, off-by-default, revocable.

Audience-tag enforcement

To prevent leakage:

  • Every retrieval filters by audience.
  • The LLM never sees unauthorized content.
  • The filter runs before the LLM, so text in a prompt cannot change which content is retrieved.

We treat any cross-audience leak as a top-priority bug.

Skill kill switches

Some skills have hard limits beyond LLM control:

  • discount_negotiator. 15% discount hard cap, enforced per negotiation.
  • subscription_manager. ₹10,000/day refund cap per workspace.
  • collect_lead. One lead per email address per workspace.

Enforced at the policy layer, not in the prompt. LLM "decides" to offer 50% off; policy rejects.

Mandatory escalation

Every conversation surface has a path to a human:

  • "Talk to a human" trigger in widget and channel UIs.
  • escalate_to_human skill hands the conversation to a person when it is used.
  • sentiment_router skill sends an alert to Slack when it detects a frustrated customer. It alerts your team; it does not hand off by itself.
  • Escalate after a failed action: when a skill fails, the agent offers a person.

No bot-only mode where customers feel trapped.

Bias and fairness

Best-effort:

  • Bot inherits LLM provider's bias in worst case.
  • Audience tags prevent serving sensitive content to wrong segments.

No bot is perfectly unbiased. We aim for fairness, monitor for issues.

Privacy

See data handling:

  • Encrypted at rest with AES-256 by our database provider.
  • TLS 1.2 or higher in transit.
  • Workspace isolation: every query is filtered to your workspace.
  • Deletion. Delete a whole workspace or your account from Settings. Conversations past your plan's retention window are deleted automatically. Per-person deletion is rolling out.

Isolation testing

Cross-workspace isolation is covered by automated tests.

What we don't promise

Honesty:

  • Zero hallucination. Not possible, and we do not publish a measured rate.
  • Perfect fairness. Best-effort; LLM bias residual.
  • Catching every prompt injection. Defense-in-depth; not invincible.
  • AGI safety. We deploy current models with current guardrails.

Be skeptical of vendors who promise more.

Limits

  • Hallucination rate. Not measured or published. Verify high-stakes answers against their sources.
  • Escalation latency. Not an SLA: the handoff trigger fires immediately, but time to a human response depends on agent availability and office hours.
  • Audit retention. 365 days, flat across all plans.

Common pitfalls

Trusting bot uncritically. Verify high-stakes answers via citations.

Assuming all safety problems solved. Bot is one layer; humans review.

Removing human-escalation path. Required; don't disable.

FAQ

Will my bot get safer over time?

Yes. We continuously improve guardrails. Your bot inherits.

Can I see the system prompts?

Yes for your workspace under AI Config. Provider-side prompts not exposed (industry norm).

What if my bot gives a wrong answer?

Customers can thumbs-down an answer. Fix it by adding or correcting a Q&A pair, or by updating the source page, and then re-ask the question.

Was this page helpful?