📖 About
Dejavue is a Discord bot for help and Q&A channels. It catches duplicate questions the moment
they're posted, drives a clean mark-solved workflow, and turns every solved thread into a
searchable archive — inside Discord, on a public SEO-friendly website, and (on the top tier) as an
MCP server your AI tools can query.
It's useful for any server where the same questions keep coming back:
- Developer tools & SaaS communities — the same setup error, asked weekly
- Game studios — crash reports and "how do I…" threads
- Open-source projects — questions your docs don't answer yet
How it works
- Someone posts in a monitored channel → Dejavue reads the question and checks for duplicates
(keyword on Free, semantic on Plus+). If it finds similar solved posts, it replies with the
matches — plus, on Plus and up, an AI-drafted answer.
- Someone resolves it: click Mark as solved (a modal captures the answer), right-click the
helpful reply → Apps → Mark as Answer, or hit Use this answer & close to borrow a previous
answer. The person who asked can always resolve their own post; anyone else needs the
Manage Threads, Timeout Members, Manage Server or Administrator permission.
Dejavue defines no roles of its own.
- Dejavue archives it: swaps the
unsolved → solved tag, stores the accepted answer plus the
full transcript, embeds it for semantic search, writes an AI summary (Pro/Max), and — if the
server opted in — publishes it to the public website.
- Everyone finds it again with
/dejavue search, on the website, or through the MCP server.
✨ Features
🔍 Duplicate detection & search
- Keyword (Free) or semantic (Plus+) duplicate detection on new posts, debounced with a
starter-message retry so it never misses the question text
/dejavue search over the solved-answer archive — private to the searcher, with broad /
balanced / exact match presets or a custom match threshold- Use this answer & close borrows a matched answer, posts it, and marks the post solved as a
duplicate — or Not a duplicate dismisses the suggestion
✅ Solving & archive
- Button + modal + message context menu + slash command, all permission-gated to OP/mod/admin
- Per-forum
solved / unsolved tag automation (created automatically on setup)
- Full thread transcript captured on solve (speakers privacy-aliased on the public website)
- Two channel modes: question (Q&A — find duplicates, mark answers) and knowledge
(archive everything, no prompts)
🤖 AI (credit-metered via OpenRouter)
- Answer drafting from past solved threads when a duplicate is found
- Thread summarization into a clean, canonical answer for the public website
- Knowledge-gap clustering — groups recurring questions so you know what docs to write
- Auto-FAQ generation and maintenance (respects manual edits)
🌐 Public website
- Per-server subdomain (
{slug}.dejavue.app) or your own custom domain
- Each monitored channel becomes its own category; pages show the answer plus the full discussion
QAPage / FAQPage JSON-LD, DB-driven sitemap, near-zero-JS pages for SEO- Opt-in per server (default off); usernames aliased as "Original poster" / "Helper N"
🔌 MCP server (Pro & Max)
- Publishes your answers as a Streamable HTTP MCP server at
{slug}.dejavue.app/mcp
- Add it to Claude, ChatGPT, Cursor, or any MCP client so your AI answers from your community's
solved questions instead of guessing
📊 Ops
- Analytics: resolution rate, time-to-resolution, top helpers
- Stale-question nudges that ping a helper role on a timer
- Re-scan to import a channel's entire history into the archive, and to drop deleted posts
🚀 Usage
Invite the Bot
Quick Setup
- Add the bot to your server using the link above
- Monitor a channel:
/dejavue setup channel:#help — this also creates the solved /
unsolved tags for forum channels
- Import your history:
/dejavue rescan channel:#help indexes past threads so duplicate
detection works from day one
- Done. Dejavue now replies to duplicates and archives every solved thread.
Next steps
/dejavue dashboard — your home screen: every indexed channel, how fresh it is, and buttons
through to everything below/dejavue settings — duplicate sensitivity, stale-question nudges, the off-topic guard and
cleanup, one page each with a plain-English explanation/dejavue website — set up your public site: name, address, theme, privacy and imprint/dejavue insights — stats, analytics, recurring-question clusters and the auto-FAQ
📝 Commands
Everything lives under one /dejavue command, plus a Mark as Answer message context menu.
Every admin screen links to every other one, so /dejavue dashboard is enough to find the rest.
| Command | Description | Who |
|---|
/dejavue dashboard | See what Dejavue is watching — channels, index size, sync status and your plan | admin |
/dejavue setup <channel> [mode] | Start indexing a channel — forum, text or announcement | admin |
/dejavue search <query> [match] [threshold] | Search every answer Dejavue has archived (results only you can see) | anyone |
/dejavue insights | Stats, resolution rate, top helpers, most-asked topics and the auto-FAQ | anyone |
/dejavue settings | Duplicate detection, stale-question nudges, the off-topic guard and cleanup | admin |
/dejavue website | Your public knowledge-base site — name, address, theme, privacy and imprint | admin |
/dejavue rescan [channel] | Re-read a channel's full history and drop posts that were deleted | admin |
/dejavue privacy | Opt out of having your messages archived, in every server | anyone |
/dejavue help | What Dejavue does, and every command explained | anyone |
| Mark as Answer (right-click a message → Apps) | Mark that reply as the accepted answer | OP/mod/admin |
/dejavue customize was renamed to /dejavue website. The old name still works for now and
points you at the new one.
Settings
/dejavue settings opens one page per feature. None of them are locked behind a paid plan.
| Page | What it controls |
|---|
| 🔁 Duplicate detection | How closely a new question must match an archived one before I suggest it. Presets, or an exact minimum match % |
| ⏰ Stale-question nudges | Whether to chase up unanswered questions, after how long, and which role to ping |
| 🧭 Channel routing & off-topic guard | Gentle "might fit better in #other" suggestions, and the stricter guard that warns on — and optionally closes — clearly off-topic posts |
| 🧹 Cleanup & branding | Whether my "Got your answer?" prompt is deleted when a post is solved, and whether the "Powered by Dejavue" footer is shown (always on for Free) |
Options
| Option | Values | Description |
|---|
mode | question, knowledge | Forums only. question runs the Q&A loop (duplicates + mark-answer); knowledge archives everything with no prompts. Default: question |
match | broad, balanced, exact | How closely search results must match. Default: balanced |
threshold | 0–100 | Custom minimum match percentage — overrides the match preset |
⚠️ The search threshold and the duplicate-detection sensitivity in /dejavue settings are both
percentages, but they are not on the same scale — search compares a question against retrieval
chunks, duplicate detection compares two questions, and the same genuine match scores much higher
in the second. Tune each against its own results, not against the other's number.
Important: Bot Permissions
Dejavue needs View Channels, Manage Channels, Send Messages, Embed Links, Read Message History,
Manage Threads and Send Messages in Threads. Manage Channels is what lets it create and swap
the solved / unsolved forum tags — without it, tag automation fails.
View Audit Log is optional. It is used once, to work out who added the bot so the setup
walkthrough can be sent to them by DM; without it, that DM goes to the server owner instead.
The Message Content privileged intent must be enabled, since Dejavue reads question text to
detect duplicates.
📊 Telemetry
The hosted bot records aggregate, per-server operational metrics (threads indexed, AI
credits used, embedding tokens spent, subscription events) so the service can be run and
billed. The allowlist in packages/analytics makes it structurally impossible to emit
message content, thread titles, search queries, or Discord user ids. distinct_id is the
guild id.
Published knowledge bases additionally count page traffic in the visitor's browser
(components/KbAnalytics.astro), in PostHog's cookieless mode: no cookie, no local
storage, no persistent visitor id, no person profile, autocapture and session replay both
off. Visitors are counted by a hash PostHog computes server-side that rotates daily. The
apex marketing site is not instrumented at all. Because this runs client-side,
POSTHOG_API_KEY is rendered into the page — correct for a PostHog project key
(phc_…), which is public by design; never put a personal key (phx_…) there.
Free-plan knowledge bases also show one ad (ADS_PROVIDER=ethicalads). It is
cookieless and non-behavioural, and because loading it discloses the reader's IP to a
third party it is gated behind an explicit opt-in — nothing is requested until the visitor
agrees, and declining falls back to an in-house promo.
Both are disclosed in the privacy policy. Self-hosting sends nothing: leave
POSTHOG_API_KEY unset and no client is created, and leave ADS_PROVIDER at none.
💳 Tiers
The hosted bot is billed per server through Discord Premium Apps. AI usage is metered in AI
credits (1 credit = 1,000 model tokens, input + output): a drafted answer costs ≈ 0.8 credits and a
thread summary ≈ 3.
Only two things actually cost us money per server: indexing (embedding your messages) and
AI generation. So those are the only two things the tiers meter. Semantic search, analytics,
nudges and theming cost nothing to serve, so every server gets them.
| Free | Plus €4.99 | Pro €9.99 | Max €24.99 |
|---|
| Forum channels | 3 | 15 | 50 | unlimited |
| Tracked (non-forum) channels | 3 | 15 | 50 | unlimited |
| Indexed messages | 2,500 | 25,000 | 250,000 | unlimited |
| Duplicate detection / search | semantic | semantic | semantic | semantic |
| Stale-question nudges | ✓ | ✓ | ✓ | ✓ |
| Analytics | full | full | full | full |
| Knowledge-base theming | ✓ | ✓ | ✓ | ✓ |
| AI-drafted answers | — | ✓ | ✓ | ✓ |
| AI summary / clustering / auto-FAQ | — | ✓ | ✓ | ✓ |
| AI credits / mo | — | 250 | 2,500 | 12,000 |
| Website page rendering | raw | AI-summarized | AI-summarized | AI-summarized |
| MCP server | — | — | ✓ (30 req/min) | ✓ (30 req/min) |
| Ads on the public site | shown | none | none | none |
| "Powered by Dejavue" branding | shown | removed | removed | removed |
One-time purchases (any tier): Custom domain — €29.99 · AI top-up packs of 500 / 1,000 /
2,000 / 5,000 credits (consumable, stacks, never expires). Importing your existing history is free,
up to each tier's indexed-message limit.
💡 Self-hosting has no tiers. Set DEV_FORCE_TIER=max and every feature is unlocked — you
just pay your own OpenRouter bill for the AI parts, or run embeddings on CPU for free.
🛠 Built With
🐳 Self Hosting
Dejavue is fully self-hostable — no license key, no phone-home, no feature gating. You have two
options: Docker (recommended) or Manual Installation.
Requirements: Docker, or Node.js 22+ with a Postgres 16 database that has the pgvector
extension available.
Docker (Recommended)
The whole stack — Postgres, migrations, bot, worker and web — comes up with one command.
1. Clone the repository
git clone https://github.com/lkaesberg/Dejavue.git
cd Dejavue
2. Create the config file
cp .env.example .env
nano .env
Fill in at minimum:
DISCORD_TOKEN=<Discord Bot Token>
DISCORD_CLIENT_ID=<Discord Bot Client ID>
OPENROUTER_API_KEY=<OpenRouter API Key> # optional: only needed for AI features
DEV_FORCE_TIER=max # unlock every feature when self-hosting
3. Start the stack
docker compose up -d --build
DATABASE_URL is auto-pointed at the bundled postgres service, migrations run on start, and the
CPU embedding model is cached in a named volume. The web app listens on port 4321.
4. Register the slash commands
docker compose exec bot pnpm --filter @dejavue/bot register
5. Encrypt the database at rest (optional)
Postgres has no built-in encryption, so scripts/encrypt-db-volume.sh encrypts the storage under it,
the way managed database providers do. It moves the Postgres volume into a LUKS2 container
mounted at the volume's own folder, so Docker, Coolify and the compose files keep using the same
volume and the official pgvector image is unchanged. It works for this compose file and for
Coolify alike (run it in Coolify's server terminal).
sudo bash scripts/encrypt-db-volume.sh setup # asks for a passphrase, then encrypts
sudo bash scripts/encrypt-db-volume.sh status
sudo bash scripts/encrypt-db-volume.sh cleanup # once verified: delete the unencrypted copies
Setup asks for a passphrase. Paste one from your password manager, or let the script generate
one and save it there; it asks you to paste it back before changing anything. The same
passphrase is stored in a key file on the server so the volume unlocks by itself at boot. Your
password-manager copy is the way back in if that file is ever lost. Dejavue is offline for the
copy, usually under a minute. Setup checks that every table has the same row count afterwards,
and rollback undoes it until you run cleanup. Needs root, systemd and a Linux host.
Manual Installation
Requirements: Node.js v22 or higher
1. Clone and install
git clone https://github.com/lkaesberg/Dejavue.git
cd Dejavue
corepack enable # provides pnpm (version pinned in package.json)
pnpm install
2. Create the config file
cp .env.example .env
nano .env
3. Start Postgres and migrate
pnpm db:up # Postgres (pinned pgvector/pgvector:pg16) in Docker
pnpm db:migrate # extension + tables + HNSW index
pnpm db:verify # proves pgvector cosine search works
Point DATABASE_URL at your own Postgres instead if you already run one — it needs the pgvector
extension.
4. Run it
pnpm dev # Postgres + migrations + bot + worker + web, with live reload
Or run the services individually: pnpm dev:bot / pnpm dev:worker / pnpm dev:web.
💡 make setup does steps 1–3 in one go, and make help lists every available task.
Dev container: open the repo in VS Code / Cursor and "Reopen in Container" — the
.devcontainer starts Postgres alongside a Node 22 workspace,
installs dependencies, and runs migrations automatically (ports 4321/5432/8090 forwarded).
Discord Application Setup
- Create an application in the
Discord Developer Portal
- Bot → Privileged Gateway Intents: enable Message Content only. Leave Presence and Server
Members off.
- OAuth2 → URL Generator: scopes
bot + applications.commands; permissions: View Channels,
Manage Channels, Send Messages, Embed Links, Read Message History, Manage Threads, Send Messages
in Threads. Open the generated URL to add your bot.
- Register commands:
pnpm --filter @dejavue/bot register — uses DISCORD_DEV_GUILD_ID for
instant guild-scoped commands, or registers globally if unset.
- Website domains (optional): point
*.yourdomain.tld (wildcard DNS + TLS) at the web app and
set KB_BASE_DOMAIN. Custom domains CNAME to the same host.
- Monetization (only if you're selling access): enable it in the portal, create the SKUs, and
put their ids in
.env. Self-hosters should just set DEV_FORCE_TIER=max instead.
Configuration Options
| Option | Description |
|---|
DATABASE_URL | Postgres connection string (needs the pgvector extension) |
DISCORD_TOKEN | Your Discord bot token |
DISCORD_CLIENT_ID | Your Discord application's client ID |
DISCORD_DEV_GUILD_ID | (Optional) Register commands to one guild for instant updates |
OPENROUTER_API_KEY | (Optional) Enables AI drafts, summaries, clustering and the auto-FAQ |
OPENROUTER_MODEL | Model for generative features (default deepseek/deepseek-v4-pro) |
KB_BASE_DOMAIN | Base domain for public knowledge bases (default dejavue.app) |
KB_PUBLIC_URL | Public URL of the marketing/website root |
KB_REVALIDATE_SECRET | Signs the private-KB gate cookie. Required in production — the app refuses to boot without it rather than fall back to a constant published in this repo. openssl rand -hex 32 |
TRUSTED_PROXY_HOPS | How many reverse proxies sit in front, each appending to X-Forwarded-For. Decides which entry is trustworthy for rate-limit keys. 0 (default) ignores the header and uses the socket peer — correct for a directly exposed app. Traefik alone = 1; Cloudflare + Traefik = 2. Setting it too high is the dangerous direction |
EMBEDDING_PROVIDER | local (default, free CPU) or openrouter |
EMBEDDING_MODEL | Embedding model id — see the table below |
EMBEDDING_SERVICE_URL | (Local provider) Set on bot and web only, to http://worker:8090: they ask the worker to embed instead of each loading the ~1.6 GB model. Both compose files set it |
QUOTA_CREDITS_PLUS / _PRO / _MAX | Monthly AI credits per tier (250 / 2,500 / 12,000) |
QUOTA_EMBED_TOKENS_FREE / / / |
Embeddings
Embeddings are pluggable via EMBEDDING_PROVIDER:
| Provider | EMBEDDING_MODEL | Cost | Notes |
|---|
local (default) | bge-small-en-v1.5 / multilingual-e5-small | free, on-CPU | downloads a small ONNX model |
openrouter | openai/text-embedding-3-large | per-token API | reuses OPENROUTER_API_KEY |
Both produce 384-dim vectors (EMBEDDING_DIM) so they share one vector(384) column — switching
is a config change. On startup the bot re-embeds threads with the new model, and search is scoped to
the active model id so stale vectors are never mixed in. Point EMBEDDING_BASE_URL /
EMBEDDING_API_KEY at https://api.openai.com/v1 to call OpenAI directly instead of OpenRouter.
💡 Fully offline AI: with EMBEDDING_PROVIDER=local and no OPENROUTER_API_KEY, Dejavue runs
duplicate detection, semantic search and the whole archive without sending anything to a third
party. Only the generative features (drafts, summaries, FAQ) need an LLM.
🏗 Architecture
packages/core shared types, env, tier/entitlement logic, clustering
packages/db Drizzle schema + migrations (pgvector) + repositories
packages/ai embeddings (Transformers.js) + OpenRouter LLM client
packages/queue pg-boss queues + job definitions
apps/bot discord.js gateway client (commands, events, interactions)
apps/worker pg-boss jobs (embed, summarize, cluster, FAQ, nudge, backfill, reconcile)
apps/web Astro SSR — public website (subdomain/custom-domain) + /mcp endpoint
The bot handles live query-embedding and interactions; the worker does the heavy and
scheduled work. They coordinate only through pg-boss and Postgres, so you can scale or restart either
independently. Tier is derived purely from Discord entitlements (hourly LIST reconcile + per-process
cache) and gates every feature in one place.
See docs/PLAN.md for the full design rationale and milestone history.
🔌 MCP Server
Once a Max server has set a website slug and opted in, its published answers are queryable at
https://{slug}.dejavue.app/mcp. Add it to any AI client as a Streamable HTTP MCP server. It
exposes one tool:
search_knowledge_base(query: string, limit?: number) → relevant solved Q&As with source links
🔒 Privacy
Dejavue is open source so you don't have to take our word for any of this — read the code, or host it
yourself.
- Publishing is opt-in per server and off by default. Nothing leaves Discord until an admin turns
it on.
- Public pages never show Discord identities. Authors appear as neutral aliases ("Original
poster", "Helper 1"); no usernames, display names, or avatars. Only the numeric user ID is stored
internally.
- No tracking cookies and no visitor profiling. Knowledge-base traffic is measured in a
cookieless mode (a daily-rotating, server-side hash — no cookie, no local storage, no persistent
id). Fonts and scripts are self-hosted; the analytics SDK is bundled, not loaded from a CDN.
- Free-tier knowledge bases show one ad from a cookieless, non-behavioural network. Because the
request discloses the reader's IP to that provider, it is opt-in: nothing loads until the
reader agrees, and declining shows an in-house promo instead. Paid plans have no ads at all.
- Any member can opt out.
/dejavue privacy stops Dejavue from storing anything they post, in
every server, and deletes what it already had of theirs.
- Removing the bot removes the data. The public knowledge base goes offline at once; the server's
content is deleted 30 days later, and adding the bot back before then restores it.
- No user accounts on the website or the knowledge bases.
- Payments happen entirely inside Discord — we never see payment details.
- AI is the only third-party processor, used only when AI features are enabled, and can be turned
off entirely (see the offline note above). Requests ask OpenRouter for providers that neither store
nor train on the data (
data_collection: deny).
Full details: Privacy Policy · Terms
👥 Contributing
Contributions are welcome. See CONTRIBUTING.md for the workflow, and
SECURITY.md for reporting vulnerabilities.
Before pushing:
make check # typecheck + tests
Developer
📄 License
Distributed under the GNU Affero General Public License v3.0. See LICENSE for the
full text.
In short: you're free to use, modify and self-host Dejavue, including commercially. If you run a
modified version as a network service, you must make your source available to its users under the
same license.
