v3.0.0 — Helpdesk agent + full reviewer-docs refresh

v3.0.0 — Helpdesk agent + full reviewer-docs refresh

Tag: v3.0.0 · Date: 2026-05-31

Summary

Ships the bounded helpdesk agent end-to-end (backend LangGraph + Vue Ask/Agent mode + HITL ticket filing), refreshes all documentation with v3 architecture diagrams, agent UI screenshots, and versioned asset layout, and completes a full reviewer-docs refresh pass — releases hub, operations manual, Architecture/Design refresh with Azure retrieval parity, and aligned reviewer introductions across README, docs landing, Reviewer Guide, and Case Study.

Re-cut on the release branch tip after the docs refresh arc (PRs #50–#58) so the v3.0.0 commit includes the consolidated reviewer-facing surface.

Highlights

Helpdesk agent (backend)

• LangGraph agent endpoints — /api/helpdesk/agent/{start,resume,confirm,abort} + SSE streams
• Supervisor + specialists — clarifier, classifier, writer, solution nodes with bounded tool budgets
• Agent tools — KB retry, web search, GitHub-issue search, HITL-gated file-ticket
• SQLite checkpointing — multi-turn session persistence across clarifying questions
• Durable chat persistence — one assistant bubble per agent journey with activity trace
• Observability — chatbot_helpdesk_agent_* Prometheus metrics + LangSmith tracing

Helpdesk agent (frontend)

• Ask / Agent mode toggle — independent chat modes with mode-specific input placeholders
• Activity timeline — multi-step agent trace rendered in a single assistant bubble
• HITL flows — radio/pill clarifying questions, outcome chips, ticket review modal with redaction warning
• Design tokens v2 — rebuilt light/dark palettes, sidebar recency grouping, always-visible copy button

Documentation refresh pass (PRs #50–#58)

• Releases hub (#50) — canonical docs/release-notes/index.md covers v1.0/v2.0/v3.0.0; duplicate docs/README.md removed.
• Helpdesk shipped-vs-target (#51) — ADR-006 records the LLM supervisor migration; helpdesk/index.md adds a shipped-vs-target table so deterministic supervisor and bounded specialists are not over-claimed.
• Architecture + Design refresh (#52) — full current state with collapsed v1/v2 history; LANGGRAPH + WEB_RESEARCH folded into DESIGN; helpdesk agent flow promoted next to RAG flow; Azure AI Search retrieval documented in parity with AWS Bedrock KB.
• Operations manual (#53) — docs/operations-manual/ consolidates CI/CD, security, release, OAuth, TLS, E2E, tenant config, performance.
• Eval refresh (#54) — eval_baseline_v2 reframed to distinguish shipped baseline from planned trajectory rig.
• Security + roadmap refresh (#55) — dependency floor pinned; PRODUCT_ROADMAP.md marks Phase 6d shipped under v3.0.0.
• First-visit polish + reviewer intros (#56, #57) — README, docs landing, Reviewer Guide, Case Study, Architecture aligned around one canonical introduction; sections reordered around product features and architecture; collapsible primary nav; scripts/check_docs.py extended to enforce intro + capability rows under tox -e docs.
• v3.0.0 release notes + changelog roll-up (#58) — folded [Unreleased] into [v3.0.0] and strengthened release-notes lede.

Documentation & assets

• v3 architecture diagrams — overview, detailed, and full topology (RAG + agent subgraphs)
• v3 product screenshots — agent mode, HITL questions, ticket draft, GitHub Issues
• Versioned asset layout — docs/assets/{architecture,product,auth}/{v1,v2,v3}/
• Agentic rebuild roadmap — AGENTIC_HELPDESK_REBUILD.md

Security & hygiene

• Tool attribution guard workflow
• Frontend dev-tool CVE remediation (vite, esbuild, vitest)
• Python dependency bumps (authlib, langchain, streamlit)
• Pydantic settings tightening + env template CI guard
• CD workflow startup fix on qa / release (PR #49)

Architecture

See architecture/v3/overview.png, detailed.png, and topology.png.

Known limits

• Supervisor routing is deterministic (hand-coded), not yet LLM-driven — see AGENTIC_HELPDESK_REBUILD.md
• SQLite checkpointing (Postgres migration planned in Phase 1b)
• RAGAS context precision (~0.50) remains the main quality bottleneck

What's next

Agentic Helpdesk Rebuild — phased migration to LLM supervisor + specialists, Postgres checkpointing, real-event SSE, trajectory eval, and live campus router.

Full release-notes index

docs/release-notes/

v1.0 — Upstream chabot baseline

v1.0 — Upstream chabot baseline

Tags: v1.0 (release ladder), v0.1 (historical fork tag — same commit) · Date: 2024 (upstream fork point)

Summary

Initial fork of the UC Berkeley ETS chabot campus chatbot — a Streamlit + FastAPI + LangChain stack querying OpenSearch and Bedrock directly.

Highlights

• Streamlit UI — single-page chat against a campus knowledge base
• FastAPI backend — chat endpoints, PostgreSQL sessions
• LangChain retrieval — direct OpenSearch vector search + Bedrock LLM
• AWS deployment — Elastic Beanstalk + Nginx configs for co-hosting API and Streamlit
• LangSmith — basic trace integration

Architecture

See architecture/v1/architecture.png for the upstream diagram.

What came next

v2 transformed this into a multicloud RAG platform with Vue SPA, provider registry, LangGraph, RAGAS eval, and CI/CD. v3 added the bounded helpdesk agent.

v2.0 — RAG platform transformation

v2.0 — RAG platform transformation

Tag: v2.0 · Date: 2026-05-19

Summary

Major platform evolution from the upstream chabot fork into Campus RAG Assistant — a production-style enterprise RAG platform with Vue product UI, multicloud provider boundaries, LangGraph orchestration, RAGAS evaluation, and GitHub Actions CI/CD.

Highlights

Product & UI

• Vue 3 SPA — streaming chat, session sidebar, source panels (Sources + Content tabs), feedback, OAuth handoff
• Opt-in web research — per-message toggle with visible disclaimer banner
• GitHub OAuth — dev OAuth on API port with one-time handoff to Vue

RAG engineering

• LangGraph pipeline — condense → multi-query → retrieve → rerank → generate → format
• Dual RAG engines — RAG_ENGINE=chain (true SSE) and langgraph (observable stages)
• Provider registry — AWS Bedrock KB, Azure AI Search + OpenAI, mock mode for CI/local
• Multi-query + RRF fusion — improved retrieval coverage

Platform & ops

• Alembic migrations — schema versioning for PostgreSQL
• Prometheus metrics — /api/metrics, request IDs, structured logging
• GitHub Actions CI/CD — tox (lint, backend, frontends) + gitleaks on every PR
• MkDocs site — GitHub Pages documentation

Evaluation & quality

• RAGAS golden-set harness — 10-question regression baseline
• Phase 5 retrieval tuning — AWS context_recall improved to 0.80 (passes gate)
• LangSmith traces — per-node spans on LangGraph path; waterfall screenshots in README

Documentation

• Architecture diagrams (v1/v2), screenshot gallery, ADRs (001–004), portfolio case study, reviewer guide

Architecture

See architecture/v2/overview.png and detailed.png.

What came next

v3 added the bounded helpdesk agent (LangGraph supervisor, HITL ticket filing, Ask/Agent mode) on top of this RAG platform.