AI governed by you

Your organization decides how a pack answers, what it knows, who sees the answers, which models do the work, and where it all runs. edisyl never takes those decisions away from you.

1. 01How it answersYour terms, your rules, your targets
2. 02What it knowsOnly the sources you choose
3. 03Who sees the answersOnly the people you add
4. 04Which modelsYour providers, your credentials
5. 05Where it runsIn your own cluster

How it answers

You or your agent asks

"What was on-time delivery for tier 1 suppliers last quarter?"

1. In your words

   Lexicon
   Reads each term the way your team means it

   • "tier 1 suppliers"→ supplier tier = 1
   • "on-time delivery"→ your on-time delivery measure
2. By your rules

   PACK.md
   Applies your source and style rules

   • "last quarter"→ Q3 of your fiscal year
   • Delivery figures→ from the ERP warehouse
3. To your standard

   Calibration
   Runs a plan the pack already verified

   • This kind of question→ a plan tested against your accuracy, speed and cost targets
4. With its evidence

   Every answer
   Answer

   94.2% of tier 1 deliveries arrived on time in Q3.

   Sources:▤ deliveries table, ERP warehouse✎ On-time delivery definition

   Or it declines, and names what is missing, when no source holds the answer.

One example question. Your team sets the Lexicon, PACK.md and the targets; the pack applies them to every question, from any agent.

• In your words. The Lexicon maps your acronyms and house terms to the right data and passages, so the pack reads a question the way your team means it.
• By your rules. PACK.md tells the pack which source to trust for each kind of question, how answers should read, and which requests to decline.
• To your standard. Objectives are the questions your team needs right, and the pack is held to accuracy, speed and cost targets on them. See Calibration.
• With its evidence. Every answer cites the passage or table behind it, and the pack declines rather than guesses when its sources don’t hold the answer.

What it knows

• Only the sources you choose. You pick the databases, schemas and tables a pack can use, and you can hold new tables for approval before they come in. See Adding tables to a connected source.
• Read access you control. edisyl queries your warehouse with the login you connect, so a read-only login keeps it read-only. See Connect a data source.
• Documents you add, and take back. You teach the documents a pack holds. Forgetting one removes it and everything the pack learned from it. See edisyl teach and forget.
• Kept to its pack. What a pack holds stays private to that pack and the people with access to it.

Who sees the answers

• People you add, with the roles you give them. Each person has an organization role, and each pack has its own members with a pack role. A Consumer can only ask, and an Operator needs Contributor on a pack to change it. See Roles and access.
• Your sign-in. People sign in through your identity provider with SAML. See Authentication.
• One organization per connection. An AI tool connected to edisyl reaches only the packs of the organization it signed in to, and only those its user can read. See Connect your agent.
• A record of every answer. Every ask is kept as an Action with its sources and full history, so you can see what was asked, by whom, and what came back. See Ask, Teach, Task.

Where it runs and which models it uses

Your environment

Your Kubernetes cluster and the systems beside it

What it reads

Your warehouse

Snowflake or Postgres, through a read-only login you control

Your documents

Only the ones you add, and you can remove any of them

→↓

edisyl

Your Knowledge Packs

Installed in your cluster with the Helm chart

↔↓

What it trusts

Your sign-in

Your identity provider, over SAML

LLM gateway

Holds your provider credentials and routes every model call

↑↓ MCP, signed in to your org

Your agents and AI tools

Claude, ChatGPT, Copilot, or your own. You choose the model they run on

↓ From the gateway only

Your model providers

Your team picks the providers and models

The platform and your data stay inside your environment. The only ways out are the agents you connect and the model providers you choose.

• In your own cluster. edisyl deploys in your Kubernetes cluster with the Helm chart, so the platform and your data stay in your environment. See Self-hosted.
• Your providers. Every model call the pack makes goes through an LLM gateway in your cluster. It holds your provider credentials, and you choose which providers it uses. See The LLM gateway.
• Your agent’s model is yours. A pack serves any agent over MCP, so you choose the AI tool and the model your agents run on.

Who does what

A pack usually involves three groups of people. On a small team, one person can be all three.

• Pack builders
  Know the job
  • Choose the pack's sources
  • Shape its Lexicon and PACK.md
  • Set its objectives and decide when it is ready
  Usually Operators, with Contributor or Admin on their packs
• Your platform team
  Runs edisyl
  • Deploy it in your cluster
  • Connect sign-in and model providers
  • Apply each upgrade from its release notes
  Usually org Admins
• People and agents who ask
  Use the answers
  • Ask from an AI tool over MCP, or the CLI
  • Reach only the packs they can read
  • Every ask kept as an Action
  Any role. Consumers can only ask

Organization and pack roles decide what each person can do. See Roles and access.

The pack informs, your agent acts

A pack never takes actions in your systems. Your agent does the job with its own tools, such as updating a CRM or opening a ticket, and you decide what those tools can do. The pack supplies the knowledge the agent decides with, and every answer cites its sources.