Changelog

A running log of what we ship. Each phase represents a focused sprint of improvements to retrieval, agents, UX, or infrastructure.

Apr 2026Phase v0.83.0Verification

Verification Hardening + Memory Efficacy

Expert verification mode with frontier model re-verification for high-stakes claims
Source-aware external query construction across 7 data sources with per-source relevance tuning
CRAG retrieval quality gate — supplements with external sources when KB results are poor
Verified-fact-to-memory promotion — high-confidence claims auto-promote to empirical memories
15-bug structural fix sprint — embedding singleton, healthcheck rewrite, verification wiring
Major dependency upgrades: langgraph 1.1, neo4j 6.1, TypeScript 6.0, Vite 8

Mar 2026Phase 51Memory

Memory Salience Improvements

6-type memory classification (empirical, decision, preference, project_context, temporal, conversational)
Per-type decay models: power-law for decisions, exponential for conversations, no decay for empirical facts
Recency-weighted access counts with reinforcement
Source authority weighting (user-stated > LLM-extracted > web search)
Migration tooling for existing knowledge bases

Mar 2026Phase 50Workflows

Visual Workflow Builder

DAG execution engine with cycle detection
SVG canvas with drag-to-reposition and type-colored nodes
4 built-in workflow templates
Live execution status tracking

Mar 2026Phase DDesktop

Electron Desktop App

Native macOS + Windows packaging
Docker lifecycle management from within the app
System tray integration and auto-updater

Mar 2026Phase A-CPlatform

Platform Foundation

Unified Docker Compose with first-run setup wizard
BYOK model configuration with 5 provider support
Web search fallback with auto-learn ingestion
A2A protocol support for agent-to-agent communication

Mar 2026Phase 38-39Performance

Performance & Security Hardening

MCP server optimization (60-80% verification latency reduction)
Expert verification mode (Grok 4) with per-message selection
Privacy hardening (CORS, port binding, encryption, audit TTL)
Quantized reranker (91MB to 23MB, 3-4x faster)

Mar 2026Phase 37UX

UX Polish

Brand identity with teal accent color system
Simple / Advanced mode for progressive disclosure
Settings reorganized into tabs with user experience presets
First-run onboarding dialog

Mar 2026Phase 34Pipeline

Advanced RAG Pipeline

6-stage adaptive retrieval pipeline
Query decomposition with parallel sub-retrieval
MMR diversity reordering + late interaction scoring
Semantic query cache with quantized embeddings
KB admin endpoints + GUI

Mar 2026Phase 33Auth

Multi-User Auth

Opt-in JWT authentication with tenant isolation
Per-user API key encryption
Marketing website launched at cerid.ai
Usage metering dashboard

Mar 2026Phase 32Retrieval

Core Retrieval Quality Uplift

Cross-encoder reranker (ONNX-accelerated, 3 modes)
Switchable client-side embeddings (Matryoshka, zero-migration)
Contextual chunking with LLM-generated situational summaries

Mar 2026Phase 29UX

Response Formatting & Verification UI

Rich Markdown rendering with custom component overrides
Inline claim verification with popovers and footnotes
Collapsible code blocks + table of contents

Mar 2026Phase 28UX

Power-User Features

Right-click context menus on toolbar icons
Drag-and-drop ingestion on KB pane and chat input
Infrastructure settings + search tuning sliders
Proactive model switch on ignorance detection

Mar 2026Phase 27Infrastructure

Infrastructure Hardening

LAN access with multi-interface IP detection
Pre-flight validation (ports, env vars, disk space)
Guided setup.sh installer
Optional Caddy HTTPS gateway

Mar 2026Phase 26Agents

Knowledge Graph & RAG Agents

Neo4j-backed knowledge graph with relationship extraction
Circuit breakers on all LLM gateway and graph DB calls
Distributed request tracing
9-agent system (query, triage, curator, rectify, audit, maintenance, hallucination, memory, self-RAG)