Ask. Answer. Resolve. Promote. — the full knowledge lifecycle in one MERN platform.
Curio Wiki — feature-by-feature guides to authentication, the Q&A forum, reputation & moderation, the FAQ + AI chatbot, the admin dashboard, and the architecture. This README is the quick overview; the Wiki is the deep dive.
Curio is a self-contained community knowledge platform that unifies everything a team needs to capture and grow its collective knowledge:
- a curated FAQ fronted by an AI chatbot,
- a structured query intake that gates out noise before it ever reaches the community,
- a Q&A forum where members answer each other, vote, and discuss, and
- a reputation + governance layer that rewards good contributors and keeps the knowledge base clean automatically.
The big idea: instead of a wiki that decays as it ages, Curio is a knowledge base that improves itself over time — the best community answers are promoted back into the canonical FAQ, duplicates are merged, stale content is flagged, and unused threads are archived, all without manual gardening.
Built as an open-source internship project with zero paid infrastructure: the externally-dependent pieces (AI model, database, hosting) are isolated behind swappable modules so production infrastructure can be dropped in later as a config change.
| Pillar | What it does |
|---|---|
| FAQ + AI Chatbot | Category-organized FAQ with hybrid keyword + semantic search. Both the search bar and the AI chatbot answer from the FAQ first and ask before checking the community forum — if there's no FAQ match they prompt "Not in the FAQ — do you want me to check the forum?" and only search the forum (then redirect to the matching thread) once you say yes. |
| Ask a Query | Structured intake with gibberish detection, unfinished-question detection, an opt-in "Refine with AI" pass, duplicate detection, admin-curated categories/tags (users pick from the list or "Others" — no free-form tagging), a required joining date and contact email, and screenshots. Posting is always attributed — no anonymous posting. |
| Q&A Forum | A support-ticket model: any member answers (but not on their own question); only the poster — or a moderator/admin — rates/closes answers, and discussion stays poster ↔ answerer (no peer voting or cross-talk). Question voting, bookmarks, full-text + semantic forum search, a My Questions / My Answers filter, and reporting included; resolved questions sink to the bottom of the list. |
| Solution Marking Engine | The poster (or a moderator/admin) marks an answer "helpful" to close the thread; admins can mark an answer "Admin Verified" (a label on the answer itself, not a user badge); resolved, high-value threads can be promoted into the canonical FAQ. |
| Reputation & Badges | Points, tiered reputation badges (shown under each author's name in the forum), admin-awarded custom badges, and admin-issued moderation flags. Admins carry no points or badges — they moderate, they don't farm reputation. |
| Admin & Governance | A telemetry dashboard, moderation queue, duplicate merge, a moderators roster, 15-minute rollback of deletions, admin-curated taxonomy, bans/roles, full badge control, FAQ management (with a duplicate guard), and an audit log. |
| Automated Maintenance | Scheduled jobs (LRU archival, staleness checks, orphan cleanup, soft-delete purge, badge recalc, embedding refresh) — each also triggerable from the admin panel. |
Access: the whole app sits behind a login gate — visitors land on a sign-in / sign-up screen and are taken into the app only after authenticating.
A query moves through a clear lifecycle, with quality gates at the front door so the community only ever sees real, well-formed questions.
1. Intake & quality gates — when a member asks a question:
- Gibberish detection — a fast heuristic (word-likeness scoring) blocks nonsense submissions; borderline cases escalate to an AI check. Blocked attempts count as spam strikes and are flagged to the moderation queue so admins can see repeat offenders.
- Unfinished-question detection — a submission that's real words but clearly half-typed (too short, placeholder text, or trailing off mid-sentence) is blocked with a prompt to finish it. Because it's an honest mistake rather than abuse, it does not count as a spam strike.
- Refine with AI — an opt-in "Refine with AI" pass cleans up spelling/clarity without changing meaning; the original text is preserved.
- Duplicate detection — the question is embedded and compared (cosine similarity) against existing ones. Above the similarity threshold the user is shown the likely duplicate and can either jump to it or post anyway (which flags it for moderator review rather than silently rejecting).
- Context — a category and tags chosen from an admin-curated list (users can't invent their own; an "Others" tag covers the gaps), a required joining date and contact email, and screenshots. Posting is never anonymous — every question is attributed to its author.
2. The forum & resolution — once posted, a query is Open:
- Any logged-in member can post an answer (Markdown supported), flipping the question to Answered — except the asker, who can't answer their own question.
- Support-ticket interaction model — no user ↔ user cross-talk. The forum is a conversation between a member and the question poster, not a free-for-all: only the poster (or a moderator/admin) can rate answers, and discussion under an answer is limited to the poster and that answer's own author. Other members can read and answer, but can't vote on or comment on each other's answers. (Questions themselves can still be up/down voted, bookmarked, and reported — reporting opens a short reason form.)
- The poster (or a moderator/admin) marks the best answer "User found helpful," which closes the question for answers, records it as the solution, and rewards the answerer; un-marking reopens the thread.
- Admin verification — an admin can mark any answer "Admin Verified." Verification lives on the answer itself: verified answers are labelled ✅ Admin Verified under the answer and pinned to the top of the thread. It is not a badge on the answerer's profile.
- Editing window — authors can edit their own question or answer for 15 minutes after posting; afterwards the content is locked.
- Escalation: a member holding the Expert badge — or any moderator — can flag a question as "Needs admin attention," routing it to the admin attention queue.
- Browsing & search — the forum list keeps unanswered questions up top and sinks resolved ones to the bottom; questions are searchable (keyword + semantic). The knowledge-base search and the chatbot only reach into the forum after the user opts in (see Finding answers below) — the forum is never silently mixed into FAQ results.
- Attachments — screenshots are shown as "N attachments added — click to view" and open in an in-app zoomable viewer (no forced downloads).
- Resolved, high-value threads can be promoted into the FAQ, turning a one-off answer into permanent, searchable knowledge.
3. Status lifecycle: Open → Answered → Resolved → (Archived). Unused resolved threads are LRU-archived over time and quietly un-archived the moment someone opens them again. If an answer is later deleted (including by a moderator/admin), the thread's status is automatically reconciled — a question never shows as "resolved" once it has no answers left, and a deleted accepted answer clears the solution. Deletions are reversible by an admin or moderator for 15 minutes via the admin Rollback view.
The FAQ is always the first place we look — and we only reach into the community forum with your permission. The search bar and the chatbot follow the same two-step contract:
FAQ search bar (the Browse FAQs page)
- As you type, Curio searches the curated FAQ semantically and shows matching entries.
- If nothing matches, the page asks: "Not in the FAQ — do you want me to check in the forum?" with a Yes, check the forum button. The forum database is not touched until you click it.
- On Yes, Curio searches the community forum and lists matching questions; clicking one redirects you to that forum thread. If the forum has nothing either, it points you to raise a query.
AI chatbot (the floating assistant)
- The assistant answers from the FAQ first and cites the FAQ entry it used.
- If the FAQ has no confident answer, it does not silently search the forum — it replies "I couldn't find this in the FAQ. Do you want me to check the community forum?" with Yes / No buttons.
- On Yes, it searches the resolved community Q&A and, on a match, redirects you to the forum thread (a clickable citation). On No, it stays put and suggests browsing the FAQ or raising a query.
This keeps FAQ answers authoritative and makes any jump into community-sourced content an explicit, opt-in choice. Under the hood the chatbot's /api/chatbot/ask accepts a check_forum flag: the first call returns source_tier: "await_forum", and the follow-up with check_forum: true returns the community match (or a fallback).
Reputation is earned, not given — and it drives the badge system.
Points are awarded for real contribution:
| Action | Points |
|---|---|
| Your answer is accepted as the solution | +15 |
| Your answer is upvoted | +2 each |
Points reward answering, not asking — posting or resolving your own question earns nothing. Downvotes never deduct an author's points (no griefing), and you can't vote on or accept your own content.
Reputation badges unlock automatically as points accumulate:
| Badge | Unlocks at |
|---|---|
| 🥉 Helper | 30 pts |
| 🥈 Contributor | 100 pts |
| 🥇 Expert | 200 pts |
| 🏆 Legend | 300 pts |
Every profile shows the full badge catalog (earned vs. locked, with the points needed) plus a reputation tier and progress toward the next tier. The highest badge a member holds is shown under their name throughout the forum, so good contributors are visible where they contribute. Admin accounts are exempt from the reputation system entirely — they show no points and no badges anywhere (forum author chips, profiles, the home dashboard, and the topbar).
Admin-controlled badges (see below) layer on top of the automatic ones:
- Admin Verified ✅ — a verification mark on an answer (shown under the answer and pinned to the top of the thread), applied when an admin marks the answer verified. It is a property of the answer, not a badge awarded to the answerer.
- Custom badges — admins can create and award a free-form badge (name + icon + reason) and revoke it later.
- Moderation flags —
⚠️ Warning, 🚫 Restricted, ☠️ Suspended — issued and revoked by admins to govern behavior. Restriction gates a user behind post-approval; suspension bans them.
Becoming a moderator — when a member reaches the Expert tier they're notified that they can apply to moderate, and can send a request from their Settings. Admins decide who actually becomes a moderator (and can grant or revoke it for anyone, regardless of badge tier). A moderator can delete/restore queries and answers, mark answers helpful, escalate questions for admin attention, and re-categorise/re-tag a question from the forum — without the full admin toolkit.
Admins get a dedicated dashboard and inline controls across the app — moderation is always one click from where the content lives.
Dashboard & insight
- System overview — live KPIs (users, open questions, resolution rate, moderation load, AI status).
- Needs Attention — actionable cards (questions escalated for admin attention, flagged content, pending approvals, open questions) that deep-link straight to the work.
- Queries by Category — central grouping of all technical queries by topic with open / answered / resolved counts and deep links.
- Attention queue — questions escalated by Expert members/moderators, grouped by category and ordered by the asker's joining date, then the question's posting date; the queue lists each asker by their email id (with their joining date) — click an email to open the question.
- Audit log — every privileged action is recorded and reviewable.
- Quieter notifications — admins don't get pinged for routine answer/like/comment activity (they moderate, not farm reputation).
Content & community control
- Moderation queue — review reported/flagged/duplicate items; resolve or dismiss.
- Duplicate merge & amalgamation — merge a duplicate into its canonical thread (answers move over automatically), or cluster related questions for review.
- Inline powers in the forum — verify answers, mark helpful, re-categorise/re-tag, and delete right where the content lives.
- Approve any answer as the solution — not just the asker; admins can mark the accepted answer on any thread.
- Delete any query or answer — with automatic status reconciliation, plus a 15-minute Rollback view to restore anything deleted by mistake (by anyone).
- Categories & Tags — curate the category and tag lists that users may choose from.
- FAQ management — create, edit, mark-outdated, or delete FAQ entries (with a near-duplicate guard that warns before adding a FAQ that already exists), and promote resolved threads into the FAQ.
People & reputation
- Bans (timed or permanent) and role management (promote/demote admins).
- Moderators — grant or revoke moderator access for any user (independent of badge tier); pending Expert requests are flagged in the user list, and a dedicated Moderators roster lists everyone with moderation powers.
- Full badge control — verify answers, award custom badges, and revoke custom or moderation badges.
- Self-moderation guard — admins cannot ban, suspend, or badge their own account; the controls are hidden on your own profile and disabled on your own row.
Automation — six scheduled maintenance jobs (LRU archival, staleness checks, orphan cleanup, soft-delete purge, badge recalculation, embedding refresh) run on a cron schedule and can also be triggered on demand from the admin panel.
- Frontend: React (Vite), React Router, Axios — a left-sidebar app shell in the "Frozen Precision" light design system (Inter type, glacial-blue palette), with Markdown-rendered threads and a personalized home dashboard.
- Backend: Node.js, Express (REST + JWT with access/refresh rotation)
- Database: MongoDB + Mongoose
- AI: Google Gemini free tier via
@google/genai(gemini-2.5-flashchat,flash-litecheap checks,gemini-embedding-001@ 768 dims). Runs in deterministic mock mode offline when no key is set. Multiple keys (AI_API_KEYS, comma-separated, from different accounts) are rotated round-robin to spread quota, and a rate-limited key is skipped on the next request. - Vector search: in-app cosine similarity over embeddings stored in MongoDB
- Scheduler:
node-cron - Testing/CI: Jest, Supertest,
mongodb-memory-server, ESLint, GitHub Actions
See PLANNING.md for full architecture and TASK.md for the milestone tracker.
This is an npm-workspaces monorepo — dependencies are isolated per workspace in local node_modules, nothing global.
Prerequisites: Node.js ≥ 20, and MongoDB (local or via the bundled Docker Compose).
npm install # install root + all workspaces
cp .env.example .env # set MONGODB_URI + JWT secrets; leave AI_API_KEY empty for mock mode
npm run seed # load the offline demo dataset (incl. the imported FAQ set)
npm run dev # Express on :5000, Vite on :5173Or run everything with one reproducible command:
docker compose up --buildUseful scripts
| Command | Description |
|---|---|
npm run dev |
Server + client in watch mode |
npm run lint |
Lint all workspaces |
npm test |
Server tests (in-memory Mongo, mocked AI) |
npm run build |
Production client build |
npm run seed |
Seed demo data + import FAQs |
Two modules are the only touchpoints for external infrastructure, so swapping in production services is configuration, not a refactor:
server/config/db.js— the single place that knows the database connection.server/config/ai.js— the single place that calls the AI provider (chat, cheap calls, embeddings) + the request queue and backoff.
npm run seed starts the platform from a clean slate — a single admin account plus the curated, offline-embedded FAQ set, so search and the RAG chatbot work with zero live AI calls. There are no demo users, points, or forum questions; everyone else self-registers and the forum fills up organically.
| Role | Password | |
|---|---|---|
| Admin | admin@example.com |
admin12345 |
A scripted ~8-minute walkthrough lives in the Demo loop checklist at the bottom of TASK.md: register → ask a question (quality gates fire) → answer → resolve → reputation/badges → admin moderation/merge/ban → FAQ promotion → RAG chatbot → trigger a maintenance job.
Contributions welcome — see
CONTRIBUTING.md.
Built by the Curio team:
| # | Name | |
|---|---|---|
| 1 | Sai Akshay Oggisetty · Team Lead | akshay2.mitmpl2024@learner.manipal.edu |
| 2 | Vishal Autade | autadevishal185@gmail.com |
| 3 | Namrata Saha | namratas.2026@gmail.com |
| 4 | Devanshi Mittal | devanshimittal106@gmail.com |
| 5 | Sharvani Shriya | shanaya12698@gmail.com |
| 6 | Aruna Thangavel | 21f3000883@ds.study.iitm.ac.in |
| 7 | Tanvish Desai | tanvishdesai.05@gmail.com |
| 8 | Hansraj Jat | hansrajjat5055@gmail.com |
| 9 | Chandana Kuncharapu | kuncharapuchandana003@gmail.com |
| 10 | Anshika Pathak | anshikapathak820@gmail.com |
| 11 | Tejash Burle | tejash.burle@gmail.com |
| 12 | Sahitya Kumar | sahityakumar0063@gmail.com |
MIT © 2026 vicharanashala