Telha User Guide (Lite)¶

Telha is the Enterprise World-State Layer for AI: it stores your organization's knowledge as versioned, provable state. Every fact carries three time axes, every answer carries provenance down to the exact source bytes, every generated claim is verified against evidence, and when Telha is not sure, it asks the right person in Teams or Slack.

This is the lite guide: enough to run Telha, load data, ask questions, and operate it day to day. The full reference documentation ships with PR-085; design detail lives in .ai/specs/.

Contents¶

Foundations

1. Getting Started: run the server, create keys, write and read your first records.
2. Core Concepts: world state, the three time axes, provenance, confidence, verification, clarifications.
3. Working with Data: ingestion formats, connectors, the query language, snapshots and compare.
4. AI Answers: /v1/generate, claim verification, traces, MCP tools, and chat.
5. Administration: backups, retention, SSO, encryption, and the operator CLI.

How people use Telha (by role)

1. End Users: asking questions in chat, reading verified answers, and answering the cards when Telha asks you.
2. Analysts & Power Users: the three time questions, evidence discipline, query recipes, MCP from your own tools.
3. Stewards: the review queue, conflicts and overrides, reading skip reasons, judgment calls.
4. Developers: what goes in Telha vs your database, SDK patterns, ingesting from your app, building AI features.

Not sure where to start? End users read pages 6 then 2. Analysts read 7 then 3. Whoever runs the server reads 1 then 5. Developers read 9 then 3.

The one-paragraph mental model¶

You write records (or connect SharePoint, Exchange, Slack, Salesforce, or upload documents). Telha stores them as append-only versions with valid time (when it was true), transaction time (when Telha learned it), and source-record time (what the document claimed). You query at any point in time, diff any two moments, and ask natural-language questions whose answers are generated from planned evidence and then verified claim by claim. Nothing is ever silently overwritten: corrections add new versions, and the old belief remains reconstructable forever.

Support surfaces at a glance¶