🛠 This page is for engineering teams self-hosting their own Lightdash instance.
What it is
Organization AI admins can download a sanitized JSON dump of any AI agent thread to share with Lightdash support when troubleshooting agent behaviour, without exposing customer data by default.Prerequisites
- A self-hosted Lightdash instance with a valid
LIGHTDASH_LICENSE_KEY. - AI Analyst enabled (
AI_COPILOT_ENABLED=trueand a configured provider). - The thread debug dump feature enabled by setting
AI_COPILOT_THREAD_DUMP_ENABLED=trueon your Lightdash deployment. It is disabled by default.
Who can download a dump
Only organization-level AI admins can download thread dumps. Concretely, the user must havemanage permission on the organization’s AI agents. Project-scoped agent managers cannot download dumps, even for threads inside a project they manage.
Eligible admins download a dump from Settings → Ask AI → Threads, using the per-row download action next to each thread.
For a refresher on the underlying roles, see Workspace roles.
What the dump contains
Each dump is a single JSON file that includes:- The full conversation — every user prompt, agent response, and tool call, including tool arguments.
- Agent configuration as of download time — the agent’s instructions, model, and feature toggles.
- The Lightdash version the instance was running at download time.
- UUIDs only — no user names or email addresses are included.
- Chart and dashboard artifact configs produced by the agent during the conversation.
Accepted risks
The dump is a support-oriented artifact, not a fully anonymized export. Before enabling the feature or sharing a file, make sure your team understands the following:- Downloaded files leave Lightdash’s retention window. Once a dump is on disk, Lightdash cannot delete or expire it. The customer owns its lifecycle from that point on.
- End users are not notified that their agent conversations may be shared. Organizations with internal privacy-notice obligations (for example, EU works councils) should document this internally before enabling the feature.
- The agent configuration block is a point-in-time snapshot. It reflects the agent’s settings at the moment of download, not at the time of the conversation. If instructions, model, or toggles changed after the thread ran, the block shows the current values. Per-turn model configuration is historical, so the specific model used for each response in the conversation is preserved even if the agent’s default model has since changed.