The Two-Tier Answer System
Last updated: March 24, 2026
Why two tiers?
A single-tier system would mean either always doing the expensive forum search (slow, 15-30s for every question) or never doing it (missing crucial edge-case knowledge that only appears in BGG threads).
Two tiers let me be fast on simple questions and thorough on hard ones.
Loading diagram...
Tier 2 adds community knowledge on top of official rules -- it never replaces the rulebook.
Confidence score
After Tier 1 synthesis, the AI model returns a confidence estimate alongside the answer. This is a self-reported score from the model, calibrated against the quality and coverage of the retrieved chunks. If the top-K chunks do not well cover the question, confidence is low.
The threshold is configurable. I tune it per-game based on how well the rulebook is structured (dense rulebooks score higher confidence on the same question than sparse ones).
Auto-escalation is silent
You never see the escalation happen. If your question triggers Tier 2, you just wait a few extra seconds. The response will arrive with community citations ([T1], [T2]) alongside the PDF citations, which tells you it was a Tier 2 answer.
When Tier 2 finds nothing
Tier 2 can fail gracefully. If the BGG community has no threads about the relevant rule (rare for popular games, more common for new releases), I return the Tier 1 answer with a note that no community discussion was found.
Performance breakdown
| Phase | Tier 1 | Tier 2 |
|---|---|---|
| Vector search | 100-200ms | 200-400ms (larger search) |
| Forum retrieval | -- | 2-8s (network + scraping) |
| AI synthesis | 3-5s | 5-20s (deeper reasoning) |
| Total | 5-7s | 7-35s |