gl-admin
—Stored in browser localStorage. Sent as X-Admin-Token to both chat-service and savings-service (same value).
Edits land in the dev chat service immediately (60s cache TTL).
Version history
Linear, append-only. Activate any row to make it live; View loads it into the editor without changing the active version.
| Version | Label | Author | When | Action |
|---|
Flat-agent LLM config
Model + reasoning effort + verbosity used by the dashboard / extension agent loops.
Read-only. Names, descriptions, and parameter schemas the LLM sees for each tool.
| Card | Issuer | Network | Rotating | Next refresh | Updated |
|---|---|---|---|---|---|
| No results yet. | |||||
LLM model · cost projection
Per-query token counts come from a 3-run benchmark average. Pick a different model to see what the same 11 query types would cost. Only the model is editable; query mix and activity scenarios are fixed for an apples-to-apples comparison.
Per-query cost & weighted mix
Product-search-heavy mix (60% product search/research, 15% post-search actions, 12% browsing, 13% meta). Weighted = (% of mix) × (avg cost).
| Query | % of mix | Avg cost | Weighted | Rationale |
|---|
Scenarios — 1,000 users
Edit % of users per activity level to model your traffic mix. The column must sum to 100%. The TOTAL row is what 1,000 users actually cost at the selected model.
| Activity level | Queries / user / day | % of users | Users (of 1,000) | $ / user / day | $ / day (segment) | $ / month (segment, 30d) |
|---|
Caveats
- Mix percentages are an estimate. Once production telemetry exists, recompute the weighted avg with the real distribution.
- "1,000 users" is daily-active. If only ~30% of signed-up users are actually active per day, divide daily figures by ~3.
- Cache hit rate is assumed to be similar across models. Different models have separate caches, so the first request after a model switch pays full price until the new cache warms up.
- Sub-tool LLM costs (Pinecone embeddings, the gpt-5-mini call inside filter_products) are NOT included. Estimated ≤$0.001 per call.
Send broadcast notification
Posts one broadcast to the selected environment's alert service. Recipients are created lazily as users open their inbox.
target:
Feedback responses are sent only to the user(s) you list below (not broadcast to everyone).
Pinecone ↔ Mongo sync
Make Pinecone an exact mirror of Mongo for credit cards / memberships — adds missing docs and deletes orphan vectors.
Credit Cards
- Mongo
- —
- Pinecone
- —
- Missing (add)
- —
- Orphans (delete)
- —
Memberships
- Mongo
- —
- Pinecone
- —
- Missing (add)
- —
- Orphans (delete)
- —
Reset onboarding (dev)
Pick a teammate and reset their onboarding — clears their feature-tour progress (profile service) and wipes their onboarding chat history so the flow restarts from the welcome card. Roster is fixed in the backend.