Conflux

Team knowledge that compounds.

Conflux is an open source team knowledge base with AI-powered search, persistent Q&A, and automatic expert discovery.

Live Demo · Quick Start · How it works

---

The problem

Knowledge lives in the wrong places. Docs nobody can find. Slack threads that scroll away. Answers in one person's head who just quit.

Notion gives you writing. Stack Overflow for Teams gives you Q&A. Neither tells you who on your team actually knows about a topic. Neither gets smarter over time.

Conflux does all three.

---

Features

• Ask, don't search- type a question, get a direct answer synthesised from your team's documents with cited sources
• Persistent Q&A- post questions, get AI-drafted answers, let teammates verify them. Verified answers enter the search index permanently
• Automatic expert discovery- expertise map built silently from writing and answering behaviour. No profiles to fill in
• Rich document editor- Tiptap-powered with slash commands, code blocks, and markdown shortcuts
• Workspace isolation- full data isolation per team at the API layer

---

How it works

RAG pipeline

Documents are chunked at the paragraph level and embedded using Gemini's gemini-embedding-001 model into PostgreSQL with pgvector. When someone searches, the query is embedded and a cosine similarity search retrieves the top 5 most relevant chunks. These are passed to Gemini with a strict grounded-answer prompt - the model only answers from retrieved context, never from training data.

Q&A feedback loop

Posted questions trigger the RAG pipeline automatically, generating an AI draft answer immediately. When a human verifies and accepts an answer, it gets embedded and stored as a new chunk in the search index. Every human correction makes future searches more accurate. The knowledge base compounds.

Expert graph

Every contribution increments a per-user, per-tag expertise score silently:

Action	Points
Write a document	+5 per tag
Post a human answer	+2 per tag
Answer accepted	+10 per tag

When AI retrieval confidence is low (which is determined by cosine similarity distance) the top scorer for the relevant tags is surfaced as a suggestion.

---

Tech stack

Framework	Next.js 15 (App Router)
Database	PostgreSQL via Neon
Vector search	pgvector
ORM	Prisma
Auth	BetterAuth
AI	Google Gemini (gemini-embedding-001 + gemini-2.5-flash-lite)
Editor	Tiptap
Styling	Tailwind CSS
Deployment	Vercel

---

Quick start

Prerequisites

• Node.js 18+
• A Neon database (free tier)
• A Google AI Studio API key (free)
• A Google Cloud OAuth app

1. Clone and install

git clone https://github.com/yourusername/conflux.git
cd conflux
npm install

2. Set up environment variables

cp .env.example .env

3. Enable pgvector and run migrations

On your Neon database, run:

CREATE EXTENSION IF NOT EXISTS vector;

Then:

npx prisma migrate dev
npx prisma generate

4. Run locally

npm run dev

Open http://localhost:3000.

---

Environment variables

DATABASE_URL=
BETTER_AUTH_SECRET=
BETTER_AUTH_URL=http://localhost:3000
NEXT_PUBLIC_APP_URL=http://localhost:3000
GOOGLE_CLIENT_ID=
GOOGLE_CLIENT_SECRET=
GEMINI_API_KEY=

---

Deployment

Conflux runs on Vercel + Neon - both free tiers are sufficient.

1. Push repo to GitHub
2. Import to Vercel
3. Add environment variables in Vercel dashboard
4. Add your Vercel URL to Google OAuth authorised redirect URIs
5. Deploy

---

License

GPL-3.0. See LICENSE.