Skip to content
threada Docs

Grounded AI

Threada is a governed work automation platform. Grounded AI is the evidence-backed answer capability inside that platform.

This page explains what “grounded” means in practice: what Threada does to make drafts trustworthy, and what it refuses to do.

The pipeline

Every draft Threada produces — whether for a WorkItem reply, a Travel itinerary, or an Invest summary — flows through the same four stages.

ingest  →  index   →  retrieve  →  answer
(sync)     (chunk +   (semantic +    (LLM +
            embed)     keyword)       citations)
  1. Ingest. Threada pulls knowledge from your connected sources: websites, document repositories, integration APIs (Salesforce, Zendesk, Confluence, Shopify, and more — see integrations). Allowed-domain allowlists prevent unauthorized fetches.
  2. Index. Content is chunked, embedded, and stored in a tenant-scoped index. Updates are scheduled per source so the index stays fresh.
  3. Retrieve. When a question or WorkItem arrives, the retriever finds the most relevant chunks using a mix of semantic similarity and keyword matching, scoped to the asking tenant.
  4. Answer. A draft is generated using only the retrieved evidence as context. Every claim in the draft links back to the chunk it came from.

What “grounded” guarantees

Citations. Every factual claim links to the chunk it came from. The operator can hover an evidence pill in the canvas to see the exact passage.

Explicit no-answer. If the retriever can’t find evidence to support an answer, Threada says so. It will not invent citations or fabricate facts to fill the gap.

Tenant scoping. Retrieval is scoped to the asking tenant’s knowledge sources. A public website indexed by two tenants may be reused at the storage layer (to avoid duplicate crawling), but only public content qualifies — customer-uploaded content is never shared.

Brand alignment. Pack policies define the voice, format, and length of generated drafts. Threada doesn’t use a generic model voice; it follows the style and policy constraints configured for your workflow.

What’s reviewable

Threada is designed so operators can verify and override anything:

  • Evidence. The supporting chunks for every claim, with links back to the original document.
  • Rationale. A short trace of why the draft says what it says.
  • History. Every revision, approval, and edit is captured on the WorkItem timeline.
  • No-answer events. Refusals are first-class — they show up in reports as content gaps you can fix.

What’s configurable

In Studio, you control:

  • Knowledge sources — what gets indexed and how often.
  • Prompt and policy — tone, format, refusal threshold, escalation rules.
  • Approval gates — which generated actions require a human review before execution.
  • Evals — golden questions that Threada answers on every release to catch regressions in grounding quality.

See the getting started guide for the end-to-end flow.