Knowledge Packs

A Knowledge Pack is what a customer has. It is not a model and not a database. It holds your organization’s own knowledge: its documents, data, and definitions. The platform keeps it current and combines it into one unit. Every ask, teach, and task targets one pack, in the CLI or over MCP.

What a pack is built from

A pack reads four kinds of source:

• Knowledge: documents, definitions, and notes you add with teach.
• Connections: live state from the applications your team works in, such as GitHub and Linear.
• Data sources: warehouse tables in Snowflake or Postgres, modeled into a semantic layer the pack can query.
• Pack references: published understanding from another pack that this one builds on.

Everything a pack holds is private to that pack and the people with access to it. You use a pack the same way whichever source answers. One ask can use several sources in the same response. Each claim is attributed to the source it came from. The sources are not blended into one answer. See Sources for how attribution works.

How a pack stays current

A pack calibrates automatically after every teach or import, and when you load objectives. Calibration compiles new documents into what the pack can answer from. It studies the pack’s models. It writes test questions from what the pack holds, binds each one to a verified answer key, and runs a test. Objectives are the questions your team writes from its work. Calibration tests them first.

Run edisyl calibrate status <pack> to see the state of a pack’s calibration, or ask the assistant in chat (calibration_status).

Delivered where you already work

You don’t open a Knowledge Pack in a separate app. The pack is served over MCP and through the CLI into the AI tool or agent your team already uses. You get answers without leaving that tool.

Why a pack

A general model answers from what it already knows. That is an average of the open web, not how your organization works. A pack gives you:

• Grounded answers. Every answer cites the documents and data it came from. When the evidence isn’t there, the pack declines instead of guessing.
• Your definitions. “Net revenue” or “active customer” means what your team means by it, not a generic meaning.
• Portability. The pack works with any model and any AI tool that supports MCP. Your knowledge is not tied to one vendor.

Finding what’s in your pack

Run edisyl pack list to see the packs you have access to, or ask the assistant in chat (list_packs). Then use edisyl pack to look at one pack, including what it is built on. This is CLI only.

Next: Ask, Teach, Task, for how you use a pack.