Use case · Knowledge

A knowledge assistant that shows its receipts

Ask in the group chat; get an answer with the source linked. Retrieval is scoped to what the asker may see — and what she cannot find, she says she cannot find. Answers must cite a verbatim source passage or they are downgraded.

Hedy's knowledge base assistant answers from your company's documents with every claim cited line-by-line to its source. Retrieval is filtered server-side by the asker's ACL, so people only see what they're cleared to see. When the evidence gate can't find a supporting passage, Hedy says it doesn't know instead of guessing. It runs self-hosted on your infrastructure, keeps knowledge private, and replaces a whole answer-desk team with one accountable AI employee.

Last updated: 20 July 2026

How Hedy runs it

01

Connect the sources

Git repositories, Notion, Confluence and meeting notes — synced and chunked with citations preserved.

02

Ask like a colleague

Questions in Feishu, Lark or Slack. Hybrid retrieval finds the passage; the answer cites it.

03

ACL per asker

Server-side filtering by the asker’s tags. Documents above a person’s clearance are not summarized, hinted at, or counted into answers.

04

Honest misses

No hit means "not found, here is the closest document and the gap has been logged" — never an invented answer.

Governed by design

  • Evidence-gated evals verify answers cite real passages before every release
  • Conduct tests probe for leaks such as salary and credentials as release blockers
  • Every answer is logged with its sources in the audit trail

In depth

Most cloud AI assistants retrieve first and check permissions never. Hedy inverts that. Every query runs against a per-asker ACL enforced on the server, before retrieval, not in the prompt. If a passage sits in a space the person can't read, it never enters the context window. The same question from a sales rep and a finance lead returns different answers because they're cleared for different sources. Access is a property of the retrieval layer, not a suggestion to the model.

The evidence gate is the core of the design. Hedy's answers must quote the source passages that support them, verbatim, with a citation to the document and location. If the gate finds no passage that grounds a claim, that claim doesn't ship. When nothing in the corpus answers the question, Hedy says so plainly rather than filling the gap with a plausible-sounding hallucination. "I don't know, here's what I searched" is a valid, expected output, not a failure.

This is one AI employee doing the work of an answer desk, not a chat toy bolted onto Slack. Hedy is a first-class citizen in Feishu/Lark and Slack, runs on your own model keys including local vLLM or Ollama, and deploys via Docker compose in about 30 minutes. It's fully self-hosted with no sub-processors, so your knowledge base and every question asked against it stay inside your infrastructure.

Governance is the product, not a settings page. Every answer and its citations land in an append-only audit log you can replay. Retrieval respects ACLs by construction. Red-line behavior tests check that Hedy refuses out-of-scope or unauthorized questions, and licensing is offline via Ed25519, so an air-gapped deployment needs no phone-home. You get a knowledge assistant you can actually put in front of an auditor.

Questions

How does Hedy make sure people only get answers from documents they're allowed to see?
Hedy enforces the asker's ACL server-side, before retrieval runs. Documents outside a user's permissions never enter the model's context, so they can't leak into an answer or citation. The same question from two people with different clearance returns different results. Access control lives in the retrieval layer, not in a prompt instruction that a model could ignore.
What happens when the knowledge base doesn't contain the answer?
Hedy says it doesn't know. The evidence gate requires every claim to be backed by a verbatim passage from a real source document. If no supporting passage exists, that claim is not produced, and if nothing answers the question Hedy states that plainly and shows what it searched. It won't invent a plausible answer to avoid an empty response.
Are the answers actually traceable to sources?
Yes. Every answer quotes the source passages that ground it, with citations to the document and location, so you can verify each claim against the original. The full question-and-answer trail, including citations, is written to an append-only audit log you can replay later for review or compliance.
Does the knowledge base data leave our infrastructure?
No. Hedy is self-hosted on your own servers via Docker or Kubernetes, including air-gapped setups, with no sub-processors. You bring your own model keys, including local vLLM or Ollama, so questions and documents never go to a third-party cloud. Licensing is offline through Ed25519, so no phone-home is required.

Related

hedy.one

See it on your own infrastructure.

Book a demo