Backend 2.0 Resilience Engine, 4-Tier Hierarchy & 18-Finding Hardening Pass
A multi-tier resilience architecture delivering continuous chat widget availability, sub-10ms zero-cost fast paths, multi-model automatic failover, and high-precision hybrid retrieval.
Key Updates & Specifications
- Tier 0 Fast Path: instant responses (<10ms, $0 cost) for greetings, custom Q&A rules, FAQ cache hits, and zero-model human escalation chips
- Multi-Model Failover: automatic fallback from primary gemini-3.5-flash-lite to secondary gemini-2.5-flash-lite across the API key pool during provider rate-limits or 5xx outages
- Redis-Backed Circuit Breaker: 3-strike failure detector opening 30-second protective cooldowns with atomic counter reset to bypass upstream timeouts
- Tier 2 Extractive Grounded Fallbacks: zero-LLM extractive citation summaries with regex word-boundary ranking during model disruptions
- SQL-Enforced Cosine Floor: relevance floor executed directly in database query before importance-weighted ranking to prevent irrelevant context inclusion
- Additive Page-Affinity Scoring: eliminates score distortion and rank inversions across search results
- Multi-Turn Follow-Up Synthesis: automatically folds previous conversation context into queries for short follow-up questions
- Worker Thread Document Ingestion: offloads PDF and DOCX parsing to dedicated worker threads with timeout termination, table structure preservation, and continuation headers
- Real-Time Lead Alerts: instant, privacy-safe NEW_LEAD dashboard notifications fanned out to workspace owners upon visitor contact capture