Your team's private AI workspace for document intelligence
Each team gets their own subdomain — sales.docsflow.app, legal.docsflow.app — with role-based access, AI-powered search, and every answer traced back to the source document.
- Create your workspace — Pick a subdomain (
your-team.docsflow.app) and invite your team - Upload documents — PDFs, Word, Excel, PowerPoint, images, and text files
- Ask questions in plain English — The AI searches across all your documents and returns answers with exact source citations
- Control who sees what — 5-tier access levels (Public → Executive) plus Admin/User/Viewer roles per workspace
Every answer includes a confidence score and clickable source links. No hallucinated claims — if the AI can't find it in your documents, it says so.
| Choose Your Subdomain | AI Chat with Sources |
|---|---|
 |
|
Pick sales.docsflow.app or any custom subdomain |
95% confidence — answers traced to source documents |
| AI Persona Setup | Workspace Onboarding |
|---|---|
 |
 |
| Guided 6-step setup tailors the AI to your business | Industry-specific configuration with custom AI assistant |
| Feature | What it does |
|---|---|
| Isolated team workspaces | Each subdomain is a fully separate environment — database-level row security ensures zero data leakage between teams |
| Hybrid AI search | Combines meaning-based search (vector similarity) with keyword matching for higher accuracy than either alone |
| Multi-model AI failover | If one AI provider goes down, queries automatically route to the next (Llama 3.3 → GPT-4o-mini → Mixtral → Gemini) |
| Confidence scoring | Every response includes a grounded confidence score — low-confidence answers are flagged rather than presented as fact |
| Custom AI persona | Each workspace can customize the AI's role, tone, and focus areas for their industry |
| Source attribution | Every claim in every answer maps back to a specific document section — click to verify |
User (sales.docsflow.app)
│
├── Clerk Auth + Edge Middleware (tenant resolution)
│
▼
Next.js 15 API Layer
│
├── /api/chat ──────────── RAG Pipeline
│ │
│ ┌─────────┴──────────┐
│ │ │
│ Pinecone LLM Failover
│ (hybrid search) (4-model cascade)
│ namespace per tenant
│
├── /api/documents ─── Upload → Parse → Chunk → Embed → Index
│ (Vercel Blob) (LangChain) (OpenAI) (Pinecone)
│
├── /api/queue ─────── Background processing (Supabase job queue)
│
└── /api/stripe ────── Subscription billing + usage tracking
│
▼
Supabase PostgreSQL (Row-Level Security on all tables)
The retrieval pipeline processes queries through multiple stages:
- Query Classification — Detects intent, complexity, and routes to the appropriate model tier
- Conversation Memory — Loads recent messages and reformulates vague follow-ups into standalone queries
- Hybrid Search — Vector similarity + BM25 keyword matching with Reciprocal Rank Fusion
- Hierarchical Retrieval — For large collections: ranks documents first, then searches within top matches
- Confidence Scoring — Threshold-based abstention — the AI won't guess if context is insufficient
- Source Attribution — Every claim maps to a specific document section with page numbers
| Layer | Mechanism |
|---|---|
| Database | Supabase Row-Level Security — queries physically cannot return another tenant's data |
| Vectors | Pinecone namespace separation per tenant — cross-tenant search is impossible |
| Auth | Clerk session tokens carry tenant context, validated on every request |
| Routing | {tenant}.docsflow.app resolved in edge middleware before any API call |
| Layer | Technology |
|---|---|
| Framework | Next.js 15 (App Router, Turbopack) |
| Language | TypeScript |
| Auth | Clerk (SSO, SAML, webhook sync) |
| Database | Supabase (PostgreSQL + RLS) |
| Vector Store | Pinecone (namespace-per-tenant) |
| LLM | Llama 3.3 70B → GPT-4o-mini → Mixtral → Gemini 2.0 Flash (4-tier failover) |
| Embeddings | OpenAI text-embedding-3-small (1536d) |
| Payments | Stripe (subscription tiers + usage metering) |
| File Storage | Vercel Blob |
| Queue | Supabase-backed job queue with cron worker |
| UI | Tailwind CSS, Radix UI (shadcn/ui), Framer Motion |
| Deployment | Vercel (serverless + edge middleware) |
docsflow/
├── app/ # Next.js App Router
│ ├── api/ # API routes (chat, documents, queue, auth, admin, stripe)
│ ├── dashboard/ # Protected dashboard pages
│ └── ... # Auth pages, onboarding, landing
├── components/ # React components (chat, admin, documents, UI primitives)
├── lib/ # Core business logic
│ ├── rag/ # RAG pipeline
│ │ ├── core/ # Embeddings, retrieval, generation, sparse vectors
│ │ ├── storage/ # Pinecone adapter (implements VectorStorage interface)
│ │ └── workflows/ # Query, ingest, delete workflows
│ ├── auth/ # Auth provider factory (Clerk/Supabase)
│ ├── subscription/ # Tier enforcement and billing
│ └── constants.ts # Centralized domain and URL configuration
├── types/ # TypeScript types (database relations, global)
├── tests/ # E2E and integration tests
├── scripts/ # Database migrations and tooling
└── docs/ # Architecture documentation
| Decision | Rationale |
|---|---|
| Pinecone over pgvector | Namespace isolation maps naturally to multi-tenancy; managed scaling without ops burden |
| 4-model LLM failover | No single provider dependency — automatic cascade on rate limit or outage |
| Supabase job queue over Redis | RLS applies to jobs, single data layer, right-sized for early-stage SaaS |
| Hybrid search (dense + sparse) | Keyword matching catches exact terms that embedding similarity misses |
| skipGeneration in query workflow | Chat route handles its own LLM call with persona context, so the RAG pipeline skips redundant generation |
git clone https://github.com/nicuk/docsflow.git
cd docsflow
npm install
cp .env.example .env.local
# Fill in API keys (Clerk, Supabase, Pinecone, OpenRouter, Stripe)
npm run devOpen http://localhost:3000.
MIT — see LICENSE for details.
Built by Nic Chin