It gets an excerpt. Never the file.
Copilot reads the context around the thing you asked about and explains what broke, when, and why. Everything else — parsing, clustering, the timeline, every finding — is computed locally whether or not it is configured at all.
Lives in Copilot — the drawer on the right
Context sent
412 lines around the finding
Cluster summary · 312 shapes
Never: the full file
What crosses the line, exactly.
The excerpt is scaled to your token budget — roughly half of it, clamped between 1k and 120k characters — and can be scoped to a time window you pick. Anomaly excerpts are pulled with a sliding window around the finding.
The full file is never sent, under any configuration.
Your key. Your account.
There is no Logier inference service and no proxy in the middle. Keys are stored locally and are never included in settings sync — the payload names the fields it sends, and API keys are not among them.
- OpenAIYour key, your account
- AnthropicClaude models
- GoogleGemini
- Ollama / LM StudioRuns on your machineoffline
- Any OpenAI-compatible endpointSelf-hosted or internaloffline
The answers that matter most are the ones it declines.
Any model will produce a confident paragraph about a log file. The useful behaviour is knowing when the file doesn’t contain the answer and saying so — then naming the log context that would. Three real shapes of that:
What deployed at 14:01?
A.Nothing in this file records a deploy. The first timeout is at 14:02:07 — if your deploy tooling writes to a separate log, opening it alongside this one would let me line them up.
Is the database down?
A.I can't tell from these entries. Every failing line here is the web tier reporting an upstream timeout; there are no database logs in this file to confirm or rule that out.
How many users were affected?
A.This log has no user or session identifiers on the failing lines, so any number I gave you would be invented. The request IDs are present, if a count of failed requests is useful instead.
Illustrative exchanges — the shape of the behaviour, not transcripts. The same principle governs generated reports: the prose may be written by the model, but every figure renders from a deterministic payload computed locally.
faq
Common questions
keep reading
Where this fits
Bring your own key. Or no key at all.
Point Copilot at a hosted provider, or at a model running on your own machine. Everything else in Logier works either way.