This page hasn't been translated yet. Showing the English version.

kovac
NewChunking auto-tune, measured on your own documents

The forge for AI agents.

Build agents that answer from your documents. Every answer cites a passage you can open, agents remember users across conversations, and you can measure answer quality before you change anything. We run the infrastructure.

Playground · hr-policy-assistant

How many days of parental leave do new fathers get, and does it apply to contractors?

search("hr-policies", …) · 6 passagesreranked · filter: employment_type

Employees are entitled to 16 weeks of paid parental leave, regardless of gender1. Contractors are not covered by this policy; their entitlements follow their service agreement2.

How it works

From a folder of documents to a deployed agent in an afternoon.

01 — Connect

Bring your knowledge

Upload 59 file formats, including PDF, Office, email, spreadsheets and code, or sync from S3, Google Drive, SharePoint, Confluence, SQL databases and websites. Permissions come with the documents.

02 — Define

Describe the agent

Pick a model, write its instructions, attach knowledge, skills and tools. An agent is configuration rather than code, so a change goes live without a deploy.

03 — Ship & measure

Put it in front of people

Call it over the API, or paste one script tag to add the chat widget to your site. Every run is traced, and answer quality is scored so you can see when it drops.

Retrieval quality you can measure and improve.

All features

Hybrid search with reranking

Each search combines meaning-based and keyword matching, then a reranker reorders the results. Search across languages, and filter on metadata the agent works out from the question.

Citations you can check

Each claim links to the exact passage it came from, highlighted on the original page, tables included.

Document-level permissions

Each user sees answers only from documents they're allowed to read. Permissions come from your connected sources and are enforced in the database.

Memory that cleans itself up

Agents remember users across conversations. In the background, kovac merges duplicates, corrects outdated facts and expires stale ones, and you can inspect or delete any memory.

Test sets built from your documents

kovac writes test questions from your own documents, scores sampled live traffic with an AI judge, and checks a new agent version against your test set before it goes live.

Chunking tuned on your own data

kovac tries different ways of splitting each collection into chunks, measures retrieval for each, and recommends the best only when the improvement holds up statistically.

What teams build on kovac

Security & trust

Your data stays separate from every other customer's.

The database itself enforces tenant isolation on every row, so a bug in our application code cannot expose one customer's data to another. Your security team can audit it.

Read our security overview
  • Tenant isolationEnforced on every row by the database
  • SSO & rolesOIDC single sign-on, role-based access, departments
  • Audit logEvery change recorded, exportable to your SIEM
  • Sandboxed codeSkill scripts run in isolated gVisor containers
  • GuardrailsJailbreak and content checks on input and output

Start small. Scale when it works.

Compare plans
StarterFor one team proving a use case.
TeamMost popularFor agents in production with real users.
EnterpriseCustomDedicated capacity, a DPA and SLAs.

Build your first agent today.

Upload a few documents and ask your first question within minutes.