Skip to content

Product phased roadmap

Last updated: 2026-05-31

Delivery phases for Campus RAG Assistant — what is shipped on main versus optional follow-ups. Design context: DESIGN.md.

Roadmap index

Doc Purpose
This file Phases, priorities, what’s next
DESIGN.md — LangGraph KB path RAG_ENGINE, graph nodes, latency, config flags
DESIGN.md — Opt-in web research research_mode=web (Vue toggle + API), Tavily config
EVALUATION.md RAGAS vs LangSmith
archive/SPRINT_2026-05-18_LANGGRAPH.md Completed AWS KB validation sprint (log)
AGENTIC_HELPDESK_REBUILD.md LLM supervisor migration, Postgres checkpointing, trajectory eval — live track
archive/PHASED_IMPROVEMENT_ROADMAP.md Campus / production scale (Redis HA, EB) — optional

Quick dev commands

tox -e lint,backend,agent-eval,frontend-vue   # CI-style checks (mock RAG + helpdesk trajectory)
PIP_SYNC=0 ./scripts/run-backend-venv.sh
./scripts/run-frontend-vue.sh

CI: GitHub Actions on push/PR to main (docs/operations-manual/ci-cd.md). Live AWS / LangGraph: set RAG_ENGINE=langgraph in local .env only (not required for tox).

Goals

Goal How we measure success
Runnable without cloud Clone → mock mode → login → chat with sources in <15 min
Credible AI engineering Providers, RAGAS harness, LangSmith traces, LangGraph
Measurable quality RAGAS golden set, LangSmith traces, documented baselines
Clean repo story README attribution; upstream LICENSE retained

Phase map

flowchart LR
  P0[Phase0_Publish]
  P1[Phase1_CommitAndDemo]
  P2[Phase2_PlatformProviders]
  P4[Phase4_LangGraph]
  P6W[Phase6b_WebResearch]
  P3[Phase3_Eval]
  P5[Phase5_RAGQuality]
  P6[Phase6_SSE_Agentic]
  P0 --> P1 --> P2 --> P4 --> P6W --> P3 --> P5 --> P6
Phase Focus Status
0–2 Publish repo, platform + providers + tox Done
4 LangGraph KB graph (RAG_ENGINE); per-node traces Done — DESIGN.md — LangGraph KB path
6b Opt-in web research (research_mode=web) Done — DESIGN.md — Opt-in web research
3 RAGAS baseline + README quality + LangSmith screenshots Done (lite) — EVALUATION.md; strict gates on release only
5 Retrieval nodes (multi-query, filters, rerank) Done — re-run RAGAS for full gates (baseline)
6d Bounded helpdesk agent (Ask/Agent mode, tools, HITL, SQLite checkpoint) Done — tagged v3.0.0; helpdesk/index.md
6 LangGraph SSE (6a); bounded rewrite loop (6c) Optional

Campus production (Redis HA, tenant budgets, Elastic Beanstalk): archive/PHASED_IMPROVEMENT_ROADMAP.md (separate phase numbering).

Completed on main (summary)

Phase 3 — Evaluation (done, lite)

Tool Role
RAGAS Golden dataset + baseline; RAGAS_QUALITY_GATE=1 on release
LangSmith Per-node traces; screenshots in README Quality section

Detail: EVALUATION.md.

Phase 4 — LangGraph (done)

Explicit RAG pipeline; same { message, metadata } contract as chain. Detail: DESIGN.md — LangGraph KB path. Sprint log: archive/SPRINT_2026-05-18_LANGGRAPH.md.

Phase 5 — RAG quality (done)

Item Status
Multi-query + RRF fusion Shipped
Metadata filters Shipped
Rerank node (FlashRank + keyword) Shipped
RAGAS tuned run eval_baseline_v2.md — recall passes; faithfulness/precision below gate

Optional next: ingestion/chunking, stricter RAGAS gates, chain vs graph parity run.

Phase 6b — Web research (done)

Opt-in research_mode=web, Tavily optional, disclaimer UI. DESIGN.md — Opt-in web research.

Phase 6d — Helpdesk agent (shipped — v3.0.0)

Slice Status Description
6d (shipped) Done on main Bounded helpdesk loop: tools, multi-turn pause/resume, HITL gate, four outcomes, Vue Ask/Agent mode, Prometheus metrics. Supervisor is deterministic today — helpdesk/index.md — Today vs target.
6d (target) Live track LLM supervisor + specialists, compiled StateGraph, AsyncPostgresSaver, enforced budgets, trajectory eval — AGENTIC_HELPDESK_REBUILD.md, ADR-006.

Specs: CONVERSATION_FLOW.md, HELPDESK_AGENT.md.

Phase 6 — Optional (RAG path)

Slice Description
6a LangGraph astream_events → same SSE shape as chain
6c Bounded grade_documents / rewrite loop (RAG_AGENTIC_ENABLED)

What to defer

Defer Why
Multi-agent swarms Cost, flakiness
Knowledge-graph RAG Only if hybrid RAG fails eval
Semantic cache After exact cache + baselines (campus track)
Production Redis HA / tenant budgets Org roadmap
Stripping UC LICENSE Requires OTL permission

What’s next

Priority Focus
Now Agentic Helpdesk Rebuild — Phase 5 (live campus router) shipped behind CAMPUS_ROUTER_ENABLED=false; next is calibrating the floor with trajectory eval before flipping the flag on in demos
Optional Phase 6a LangGraph SSE; grow golden set (10 → 30–50); faithfulness/precision via ingestion/chunking
Later Campus scale track — PHASED_IMPROVEMENT_ROADMAP.md (Redis HA, exact/semantic cache, tenant budgets)