Production-grade enterprise RAG platform — workspace-scoped document intelligence with hybrid search, streaming AI chat, LLM-as-judge evaluation, and SaaS billing.
Product walkthrough + UI mockups.
Pricing — Free / Pro ($19) / Team ($49) |
Mobile — landing + chat preview |
Upload docs → ask questions → get cited answers. Workspaces, eval dashboard, and Stripe billing included.
Most "chat with your PDF" demos are single-user, flat document lists, and zero access control. IntelliVault is different.
It's a full SaaS product built around team workspaces — isolated containers where organisations upload documents, invite collaborators with specific roles, and get AI answers grounded strictly in their knowledge base. Every response is cited, every quality metric is measured, and every API call is billing-aware.
| Problem | Knowledge workers waste hours searching internal documents, policies, and handbooks — with no guarantee the AI answer is accurate |
| Solution | A production RAG pipeline with hybrid retrieval, contextual chunking, query rewriting, re-ranking, and automated quality evaluation |
| For | Engineering teams, legal teams, HR departments, and any organisation that needs to make large document collections searchable and conversational |
- Multi-tenant workspaces — isolated document collections, chat sessions, and member management per workspace
- Role-based access control — Owner / Editor / Viewer roles enforced at the API layer on every route, never just the frontend
- Document ingestion pipeline — PDF, DOCX, TXT, and Markdown parsed, chunked, embedded, and indexed automatically
- Streaming RAG chat — token-by-token streaming via Vercel AI SDK v7 with tool-calling architecture
- Source citation chips — every assistant response shows which documents backed it
- Session persistence — conversations saved to DB, titled by LLM, resumable from history sidebar
- Email invite system — invite collaborators by email with tokenized links; works with or without an existing account
- SaaS billing — Stripe-powered Free / Pro / Team tiers with usage limits enforced server-side (not just UI guards)
- Hybrid search — pgvector cosine similarity + Postgres BM25 full-text search, fused via Reciprocal Rank Fusion (RRF)
- Contextual chunking — document name + detected section heading prepended before embedding; raw chunk stored for display
- Query rewriting — follow-up questions rewritten into standalone queries using last 6 messages of conversation history
- Cohere re-ranking — hybrid search returns top 20 chunks; Cohere cross-encoder re-ranks to top 5 before LLM
- LLM-as-judge evaluation — three parallel GPT-4o-mini calls score every response on Context Relevance, Faithfulness, and Answer Relevance
- Eval dashboard — per-workspace averages + per-session breakdown + lowest-scoring sessions panel
- 30-day billing cycles — rolling 30 days from subscription date, not calendar month boundary
- Audit logging — every action inside a workspace logged with actor, timestamp, and metadata
- HyDE — embed a hypothetical answer instead of the raw query for better semantic matching
- Multi-document cross-referencing agent — reason across multiple documents in one response
- Chunk-level thumbs up/down feedback → future retrieval fine-tuning signal
- Scheduled email digests of workspace activity
- Real-time member presence in workspace chat
Every assistant response is automatically scored. Current workspace averages:
| Metric | Before | After (full pipeline) | Meaning |
|---|---|---|---|
| Context Relevance | 0.10 | 0.88 | Right chunks retrieved |
| Faithfulness | 0.70 | 0.97 | No hallucination |
| Answer Relevance | 0.70 | 0.97 | Question directly answered |
The 8.8× improvement in Context Relevance came from contextual chunking alone — before vs after re-uploading with document name + section heading prepended.
- Next.js 16 (App Router, Server Components, Route Groups)
- TypeScript 5 — strict mode throughout
- Tailwind CSS v4 — Linear-inspired dark design system
- Vercel AI SDK v7 —
useChat,streamText, tool-calling,UIMessagetypes
- Next.js Route Handlers — all API logic, no separate server
- Drizzle ORM — type-safe queries, relations, migrations
- Zod — request validation on every route
- Neon — serverless PostgreSQL with connection pooling
- pgvector —
vector(768)column with HNSW index for cosine similarity - Postgres full-text search —
tsvectorgenerated column with GIN index for BM25
- OpenAI
gpt-4o-mini— LLM inference, query rewriting, LLM-as-judge evaluation - Google Gemini
gemini-embedding-001— 768-dim text embeddings viaembedMany() - Cohere
rerank-v3.5— cross-encoder re-ranking on top-20 retrieved chunks
- Clerk — authentication, session management,
useUser()/auth()server-side - Stripe — checkout sessions, webhook-driven plan management, customer portal
- Resend — transactional email for workspace invites
- React Email — typed email templates
- unpdf — PDF text extraction
- mammoth — DOCX text extraction
- LangChain
RecursiveCharacterTextSplitter— 1000-char chunks, 200-char overlap
- Vercel — production deployment, serverless functions, Edge runtime
- Neon — serverless Postgres with auto-scaling
┌─────────────────────────────────────────────────────────────────┐
│ CLIENT (Browser) │
│ Next.js App Router · Server Components · Vercel AI SDK useChat │
└──────────────────────────────┬──────────────────────────────────┘
│ HTTP / SSE streaming
┌──────────────────────────────▼──────────────────────────────────┐
│ API LAYER (Next.js) │
│ Route Handlers · Clerk auth() · Drizzle ORM · Zod validation │
└──────┬──────────────────┬───────────────────┬───────────────────┘
│ │ │
┌──────▼──────┐ ┌────────▼───────┐ ┌───────▼────────┐
│ Neon DB │ │ External AI │ │ Stripe │
│ pgvector │ │ OpenAI │ │ Checkout │
│ BM25 fts │ │ Gemini │ │ Webhooks │
│ Drizzle │ │ Cohere │ │ Portal │
└─────────────┘ └────────────────┘ └────────────────┘
File Upload (PDF/DOCX/TXT/MD)
│
▼
Text Extraction unpdf · mammoth · raw text
│
▼
Chunking RecursiveCharacterTextSplitter
1000 chars · 200 overlap
│
├─────────────────────────────────────┐
│ │
▼ ▼
Contextual Enrichment Keyword Indexing
"Document: X | Section: Y tsvector GENERATED ALWAYS
--- [raw chunk]" GIN index for BM25
│
▼
Gemini Embedding 768-dim vector per chunk
│
▼
chunks table content · embedding · fts
HNSW index · GIN index
User question
│
▼
Query Rewriting GPT-4o-mini + last 6 messages
│
├──────────────────────────┐
│ │
▼ ▼
Vector Search BM25 Search
cosineDistance() ts_rank(fts, query)
top 20 by similarity top 20 by keyword rank
│ │
└────────────┬─────────────┘
│
▼
Reciprocal Rank Fusion score = Σ 1/(rank + 60)
top 20 fused results
│
▼
Cohere Re-ranking cross-encoder reads query + chunk
top 5 most relevant
│
▼
LLM Synthesis GPT-4o-mini · grounded · streamed
│
▼
LLM-as-judge Eval 3× parallel calls → scores saved to DB
workspaces
├── workspace_members role: owner | editor | viewer
├── workspace_invites token · 7-day expiry · status enum
├── documents status: processing | ready | failed
│ └── chunks vector(768) HNSW + tsvector GIN
│ contextual embedding · raw content
├── chat_sessions LLM-generated title
│ └── messages source_chunk_ids jsonb
│ context_relevance · faithfulness · answer_relevance
└── activity_logs actor · action · metadata
billing (user-level, not workspace-level)
userId · plan · stripeCustomerId · messageCount · messageResetAt
intellivault-enterprise/
├── src/
│ ├── app/
│ │ ├── (marketing)/ # Public routes — landing, pricing, sign-in/up
│ │ ├── (app)/
│ │ │ ├── dashboard/ # Workspace switcher + create workspace
│ │ │ ├── billing/ # Plan management, usage meters, success page
│ │ │ └── workspace/[id]/
│ │ │ ├── layout.tsx # Auth guard + sidebar + membership check
│ │ │ ├── documents/ # Upload dropzone + document table
│ │ │ ├── chat/ # RAG chat + session history sidebar
│ │ │ │ └── [sessionId]/
│ │ │ ├── eval/ # LLM-as-judge evaluation dashboard
│ │ │ ├── members/ # Member list + invite modal
│ │ │ └── settings/ # Rename + danger zone
│ │ └── api/
│ │ ├── workspaces/[id]/
│ │ │ ├── documents/[docId]/
│ │ │ ├── chat/sessions/
│ │ │ ├── members/
│ │ │ ├── invites/
│ │ │ └── eval/
│ │ ├── invites/[token]/
│ │ └── billing/
│ │ ├── checkout/ # Stripe checkout session
│ │ ├── portal/ # Stripe customer portal
│ │ └── webhook/ # Stripe webhook handler
│ ├── components/
│ │ ├── documents/ # UploadDropzone · DocumentTable · StatusBadge
│ │ ├── eval/ # ScoreCard · SessionTable · WorstSessions
│ │ ├── workspace-home/ # StatsCard · RecentDocuments · RecentChats
│ │ ├── dashboard/ # WorkspaceCard · CreateWorkspaceModal
│ │ └── ai-elements/ # Chat UI primitives (Message · Response · Citation)
│ ├── schema/ # Drizzle table definitions + enums + relations
│ ├── lib/
│ │ ├── search.ts # Hybrid search: vector + BM25 + RRF + rewrite
│ │ ├── rerank.ts # Cohere cross-encoder re-ranking
│ │ ├── eval.ts # LLM-as-judge scoring (3 parallel calls)
│ │ ├── billing.ts # getUserBilling · checkLimit · incrementCount
│ │ ├── plans.ts # PLAN_LIMITS config
│ │ ├── stripe.ts # Stripe singleton client
│ │ ├── embedding.ts # Gemini embed + embedMany
│ │ ├── chunking.ts # Splitter + detectHeading + addContext
│ │ └── parsers/ # pdf · docx · text · index router
│ └── emails/ # React Email templates
├── migrations/ # Drizzle SQL migrations
├── drizzle.config.ts
└── middleware.ts # Clerk route protection
- Node.js 20+
- Neon PostgreSQL database
- Clerk application
- OpenAI API key
- Google AI Studio API key
- Cohere API key
- Resend API key
- Stripe account with two products created (Pro + Team)
git clone https://github.com/itsyashbisht/IntelliVault-Enterprise.git
cd IntelliVault-Enterprise
npm installCopy .env.example to .env.local and fill in all values:
cp .env.example .env.localRun once in your Neon SQL editor:
-- Enable vector extension
CREATE EXTENSION IF NOT EXISTS vector;
-- Create enums
CREATE TYPE workspace_role AS ENUM ('owner', 'editor', 'viewer');
CREATE TYPE invite_status AS ENUM ('pending', 'accepted', 'expired');
CREATE TYPE document_status AS ENUM ('processing', 'ready', 'failed');
CREATE TYPE message_role AS ENUM ('user', 'assistant');
CREATE TYPE plan_type AS ENUM ('free', 'pro', 'team');
-- Push schema
npx drizzle-kit push
-- After push — add BM25 full-text index
ALTER TABLE chunks
ADD COLUMN fts tsvector
GENERATED ALWAYS AS (to_tsvector('english', content)) STORED;
CREATE INDEX chunks_fts_idx ON chunks USING GIN(fts);
-- Add eval score columns
ALTER TABLE messages
ADD COLUMN context_relevance NUMERIC(3,2),
ADD COLUMN faithfulness NUMERIC(3,2),
ADD COLUMN answer_relevance NUMERIC(3,2);- Create two products in your Stripe dashboard (Pro at $12/mo, Team at $29/mo)
- Copy price IDs to env variables
- Create a webhook endpoint pointing to
/api/billing/webhook - Listen for:
checkout.session.completed,customer.subscription.deleted,invoice.payment_failed - Copy webhook signing secret to
STRIPE_WEBHOOK_SECRET
npm run dev| Variable | Description | Required |
|---|---|---|
NEON_DATABASE_URL |
Neon PostgreSQL connection string (neon-serverless) | ✅ |
NEXT_PUBLIC_CLERK_PUBLISHABLE_KEY |
Clerk publishable key | ✅ |
CLERK_SECRET_KEY |
Clerk secret key | ✅ |
OPENAI_API_KEY |
OpenAI API key (GPT-4o-mini) | ✅ |
GOOGLE_GENERATIVE_AI_API_KEY |
Google AI Studio key (Gemini embeddings) | ✅ |
COHERE_API_KEY |
Cohere API key (re-ranking) | ✅ |
RESEND_API_KEY |
Resend API key (invite emails) | ✅ |
STRIPE_SECRET_KEY |
Stripe secret key | ✅ |
STRIPE_WEBHOOK_SECRET |
Stripe webhook signing secret | ✅ |
NEXT_PUBLIC_STRIPE_PRO_PRICE_ID |
Stripe price ID for Pro plan | ✅ |
NEXT_PUBLIC_STRIPE_TEAM_PRICE_ID |
Stripe price ID for Team plan | ✅ |
NEXT_PUBLIC_APP_URL |
Your app's public URL | ✅ |
- Sign up at the landing page → Clerk handles authentication
- Create a workspace — give it a name; you're automatically the owner
- Upload documents — PDF, DOCX, TXT, or Markdown; status updates from
processing → ready - Start chatting — ask any question; IntelliVault searches your documents and streams a cited answer
- Invite teammates — enter their email; they receive a tokenized invite link; roles assigned on acceptance
- View eval scores — open the Eval tab to see retrieval quality metrics per session
- Upgrade plan — visit
/billingto upgrade from Free to Pro or Team
# Run dev server
npm run dev
# Push schema changes
npx drizzle-kit push
# Generate migrations (production)
npx drizzle-kit generate
npx drizzle-kit migrate
# Type check
npx tsc --noEmit
# Lint
npm run lintHero section with product mockup, feature grid, and pricing CTA
All workspaces with doc count, member count, and role badge
Upload dropzone + document table with status badges (processing / ready / failed)
Streaming response with source citation chips below each assistant message
Workspace-average score bars + per-session breakdown + lowest-performing sessions
Current plan, usage meters (messages / documents / workspaces), Stripe portal link
| Route | Methods | Access | Purpose |
|---|---|---|---|
/api/workspaces |
POST GET |
Authenticated | Create and list workspaces |
/api/workspaces/[id] |
GET PATCH DELETE |
Member / Owner | Single workspace CRUD |
/api/workspaces/[id]/members |
GET PATCH DELETE |
Member / Owner | Member management |
/api/workspaces/[id]/invites |
POST GET DELETE |
Owner / Editor | Email invite system |
/api/workspaces/[id]/documents |
POST GET |
Member | Ingest and list documents |
/api/workspaces/[id]/documents/[docId] |
DELETE |
Owner / Editor | Delete document + chunks |
/api/workspaces/[id]/chat |
POST |
Member | Streaming RAG chat |
/api/workspaces/[id]/chat/sessions |
POST GET |
Member | Session management |
/api/workspaces/[id]/eval |
GET |
Owner / Editor | Evaluation metrics |
/api/invites/[token] |
POST |
Authenticated | Accept workspace invite |
/api/billing/checkout |
POST |
Authenticated | Stripe checkout session |
/api/billing/portal |
GET |
Authenticated | Stripe customer portal |
/api/billing/webhook |
POST |
Stripe (signed) | Handle payment events |
Nine core tables in dependency order:
workspaces Central tenant container
├── workspace_members User ↔ workspace join (role-bearing)
├── workspace_invites Token-based email invites
├── documents Document metadata (status enum)
│ └── chunks Text chunks + vector(768) + tsvector
├── chat_sessions Conversation threads per workspace
│ └── messages User/assistant turns + eval scores
└── activity_logs Immutable audit trail
billing User-level (not workspace-level)
userId · plan · Stripe IDs · messageCount · messageResetAt
Key design decisions:
chunks.ftsisGENERATED ALWAYS AS (to_tsvector('english', content)) STORED— Postgres maintains it automaticallymessages.source_chunk_idsisjsonb— stores chunk UUIDs used per response for citation UI- All workspace-child tables use
ON DELETE CASCADE— deleting a workspace cleans everything atomically billing.messageResetAtuses 30-day rolling window, not calendar month
- Hybrid search architecture — combining dense (vector) and sparse (BM25) retrieval with RRF fusion
- Cross-encoder vs bi-encoder — why re-ranking improves precision after approximate retrieval
- Contextual embeddings — why raw chunk content alone produces poor retrieval for domain documents
- LLM-as-judge evaluation — building measurable quality metrics instead of subjective testing
- Stripe webhook reliability — why the webhook handler is the source of truth, not the success redirect URL
- Multi-tenant data isolation — enforcing workspace boundaries at every DB query, not just middleware
| Challenge | Root Cause | Solution |
|---|---|---|
| Context relevance at 0.10 | Pure vector search misses exact terms | Hybrid BM25 + vector + RRF |
| Chunks lacked context | Embedding short text without document context | Prepend doc name + section heading |
| Follow-up questions failing | Pronouns and references in short queries | Query rewriting with conversation history |
| No quality signal | Subjective manual testing | LLM-as-judge eval dashboard |
| Wrong billing reset | Calendar month vs subscription date | 30-day rolling from messageResetAt |
- HNSW index on
chunks.embedding— approximate nearest-neighbor, O(log n) lookup - GIN index on
chunks.fts— inverted index, fast full-text lookups - Retrieve 20, re-rank to 5 — broad recall from hybrid search, precision from Cohere
Promise.alleverywhere — parallel DB queries instead of sequential
- Direct DB calls in server components — no unnecessary HTTP round-trips to own API
neon-serverless(WebSocket driver) for transactions — notneon-http- Usage limits at 402 server-side — not frontend-only guards
- Webhook signature verification —
stripe.webhooks.constructEventwith raw body bytes
- Membership check on every workspace-scoped route —
workspace_memberslookup before any data access - Cross-tenant document check before delete —
AND workspace_id = ?on every document query - Stripe webhook signed — rejects requests without valid
Stripe-Signatureheader - Clerk session on every API route — no unauthenticated access to any data
| Enhancement | Impact | Complexity |
|---|---|---|
| HyDE (Hypothetical Document Embedding) | Better semantic retrieval | Medium |
| Multi-document reasoning agent | Cross-reference across docs | High |
| Chunk feedback (thumbs up/down) | Training signal for future fine-tuning | Low |
| Scheduled activity digests | Engagement + retention | Medium |
| Real-time presence in workspace | Collaboration UX | High |
| Fine-tuned embedding model | Domain-specific retrieval | High |
Contributions are welcome. Please follow this process:
# Fork the repository
# Create a feature branch
git checkout -b feature/your-feature-name
# Make your changes with clear commits
git commit -m "feat: add your feature"
# Push and open a pull request
git push origin feature/your-feature-nameBefore submitting:
- Run
npx tsc --noEmit— no TypeScript errors - Run
npm run lint— no lint errors - Test your feature against a real Neon database
This project is licensed under the MIT License.
MIT License — Copyright (c) 2026 Yash Bisht
Permission is granted to use, copy, modify, and distribute this software.
Yash Bisht Full-Stack AI Engineer
Built in ~7 weeks from schema to deployed SaaS product.
IntelliVault Enterprise · Built with Next.js · pgvector · OpenAI · Gemini · Stripe · Vercel
If this helped you build something, give it a ⭐