Skip to main content

Release Notes: Feb 14-20, 2026

175 commits | Chatbot UI + knowledge architecture

TL;DR

The chatbot got a complete UI redesign with follow-up suggestions, citation auto-open, and escalation support. A portable medical knowledge package was created, the graph went through v10-v11 quality repair cycles, a two-tier semantic query cache was introduced, and the golden eval started at ~80% ending the week at 87.9%.

Chatbot UI Redesign​

  • Two-phase layout with collapsible QA sections
  • Follow-up suggestion chips with escalation support
  • Citations auto-open on response complete, auto-close on input focus
  • Skeleton loader and motion animations
  • Full i18n translations for the redesign

Medical Knowledge Architecture​

  • Created portable medical_knowledge package (conditions, treatments, departments)
  • LLM entity validation with taxonomy context and synonym detection
  • DIAGNOSES graph traversal wired into query pipeline
  • Center-to-department mapping and algorithmic doctor name deduplication

Graph Quality (v10-v11)​

  • Golden-only gate: skip graph storage for non-authoritative pages (ADR-0029)
  • Removed Graphiti library — direct neo4j.AsyncDriver
  • Frozen relationship enforcement in AllowlistFilter
  • Dual-label Centers as Department nodes
  • Comprehensive repair scripts

RAG Pipeline Improvements​

  • Two-tier semantic query cache (ADR-0031) with settings UI
  • LLM-first query normalization and clickable graph citations
  • Escalated search through full RAG pipeline
  • Per-sub-query reranking for multi-hop queries

Self-Healing Database​

  • Database diagnostics with self-heal actions
  • Canonical questions repair for all chunks (not just first 10)
  • pgvector diagnostic accuracy fixes

Key ADRs​

  • ADR-0027: Multilingual prompts + GPT-4.1 best practices
  • ADR-0029: Remove Graphiti, use direct Neo4j driver
  • ADR-0030: LLM structured entity extraction replaces dictionary gating
  • ADR-0031: Two-tier semantic query cache

Evaluation​

Golden eval: ~80% to 87.9% across the week. Taxonomy pipeline redesign spec written (6 sub-projects that would drive the next 4 weeks of work).