Draftly

Enterprise Multi-Tenant AI Proposal & RFP Generator (RAG)

Transform hours of manual proposal writing into seconds.
Draftly is a full-stack SaaS application that leverages Retrieval-Augmented Generation (RAG), vector embeddings, and dual-LLM fallback strategies to generate tailored, 10-section RFP proposal drafts from organizational knowledge bases.

_{Built with multi-tenant isolation, async job queues, pgvector similarity search, and automated test coverage.}

---

Table of Contents

• Engineering & Architecture Highlights
• Application Preview & Screenshots
• Problem & Solution
• Key Features
• Architecture & Data Flow
• Tech Stack
• Security & Multi-Tenancy
• Billing & Subscription Engine
• License & Contact

---

Engineering & Architecture Highlights

Key engineering decisions and architectural implementations in Draftly:

• Production RAG Engine: Implements end-to-end vector retrieval with pgvector and Google AI text-embedding-004 (768-dimensional embeddings), chunking documents into ~500-word segments with 50-word overlap for high context recall.
• Fault-Tolerant Dual-LLM Pipeline: Primary generation uses Google Gemini 2.5 Flash. On encountering rate limits (HTTP 429), the task seamlessly falls back to Groq llama-3.1-8b-instant, ensuring high uptime.
• Strict Multi-Tenant Isolation: Custom Django REST Framework permissions (IsOrgMember, OrgDocQuotaPermission, OrgProposalQuotaPermission) ensure zero cross-tenant data leakage and enforce strict row-level security on all database queries.
• Asynchronous Worker Architecture: Non-blocking document ingestion and proposal generation executed via Celery with Redis as broker/result backend, featuring status polling and exponential backoff retries.
• Full Stripe Subscription Integration: End-to-end monetization featuring Stripe Checkout, Customer Portal, and webhook signature verification with idempotency protection via StripeEvent logging.
• Robust Automated Test Suite: Full test coverage with pytest on the backend and Vitest + React Testing Library + MSW (Mock Service Worker) on the frontend, protected by pre-push hooks.

---

Application Preview & Screenshots

1. Proposal Generator & RFP Workspace

Paste RFP requirements, select past company case studies, customize generation tone, and trigger AI drafting.

---

2. Interactive 10-Section Proposal Editor

WYSIWYG section-by-section proposal editor featuring real-time status polling, prompt refinement, and PDF / DOCX export capabilities.

---

3. RAG Knowledge Base & Document Library

Multi-format document ingestion (PDF, DOCX, TXT), vector chunking status indicators, metadata tracking, and semantic search.

---

4. Real-Time Analytics & Quota Monitor

Comprehensive org analytics displaying proposal volume trends, word counts, active vector chunk metrics, and monthly tier quota usage.

---

Problem & Solution

Problem	Draftly Solution
Time-Consuming RFPs: B2B sales teams spend 15+ hours drafting proposals manually.	Sub-60s Draft Generation: Draftly builds structured, 10-section proposals in seconds.
Hallucinations & Generic Reponses: Off-the-shelf LLMs output generic, ungrounded text.	Grounding via RAG: Proposals are strictly synthesized from company case studies & reference docs.
Formatting Inconsistencies: Proposals lack standard corporate structure.	Deterministic 10-Section Schema: Guarantees standard executive summary, methodology, pricing, etc.
Data Privacy Risks: Mixing client context across organizations.	Tenant-Isolated Vector Store: pgvector cosine search is strictly scoped by org_id.

---

Key Features

1. RAG Knowledge Base Ingestion

• Upload PDF, DOCX, and TXT company documents.
• Automatic text extraction via PyMuPDF and python-docx.
• Smart word-based chunking with configurable overlap.
• Batched embedding generation stored directly in PostgreSQL using pgvector.

2. AI Proposal Generation Engine

• Automatically parses RFP requirements and generates 10 tailored sections:
  1. Executive Summary
  2. Understanding Requirements
  3. Proposed Solution
  4. Relevant Experience
  5. Team Qualifications
  6. Project Timeline
  7. Methodology
  8. Pricing Structure
  9. Why Choose Us
  10. Appendix & Terminology
• Automatic LLM fallback strategy (Gemini 2.5 Flash ➡️ Groq Llama 3.1 8B).

3. Rich Editor & Export Tools

• Section-by-section Tiptap rich-text editor in React.
• Export finalized proposals directly to PDF (styled layout) or DOCX (editable Word document).

4. Tiered Quota & Billing Management

• Subscription tiers: Free, Solo, Studio, and Agency.
• Real-time enforcement of document uploads and monthly proposal generation limits.
• Automated monthly quota resets on the 1st of each month (UTC).

---

Architecture & Data Flow

                               ┌────────────────────────────────────────────────────────┐
                               │                    React 18 SPA                        │
                               │  Vite • Zustand • Tiptap • Recharts • Axios Interceptor │
                               └───────────────────────────┬────────────────────────────┘
                                                           │ HTTPS (JWT Auth)
                               ┌───────────────────────────▼────────────────────────────┐
                               │                  Django REST Framework                 │
                               │     Auth • Multi-Tenant Permissions • REST Endpoints   │
                               └──────────┬──────────────────┬──────────────────┬───────┘
                                          │                  │                  │
                                          ▼                  ▼                  ▼
                                   ┌──────────────┐   ┌──────────────┐   ┌──────────────┐
                                   │  PostgreSQL  │   │ Celery Queue │   │ External AI  │
                                   │  + pgvector  │   │  + Redis     │   │ APIs & Stripe│
                                   └──────────────┘   └──────────────┘   └──────────────┘

Async Document & RAG Pipeline Flow

[User Upload] ──> [Django REST API] ──> [Save File] ──> [Celery Task]
                                                             │
   ┌─────────────────────────────────────────────────────────┴─────────────────────────────────────────┐
   │ 1. Extract raw text (PyMuPDF / python-docx)                                                       │
   │ 2. Split into ~500-word overlapping chunks                                                         │
   │ 3. Batch embed via Google AI text-embedding-004                                                  │
   │ 4. Store vectors in PostgreSQL (pgvector) scoped to tenant org_id                                 │
   └─────────────────────────────────────────────────────────┬─────────────────────────────────────────┘
                                                             │
[Proposal Request] ──> [Embed RFP Query] ──> [Cosine Vector Search] ──> [Prompt Gemini / Groq] ──> [10-Section Draft]

Core Data Models Schema

Model	Purpose	Key Attributes / Relationships
Organization	Tenant boundary & billing	subscription_tier, doc_quota, proposal_quota, Stripe IDs
User	Tenant member	org (FK), email, role (admin/member), is_active
Document	Raw uploaded knowledge	org (FK), uploaded_by (FK), file_type, status (processed/failed)
Chunk	Embedded text segment	document (FK), org (FK), content, embedding (vector(768))
RFP	Target proposal request	org (FK), created_by (FK), title, raw_text
Proposal	Final generated output	rfp (FK), org (FK), sections (JSONField), status (draft/final)

---

Tech Stack

Backend & Infrastructure

• Framework: Python 3.11, Django 5.0, Django REST Framework
• Task Queue: Celery 5.3, Redis 7 (Broker & Backend)
• Database: PostgreSQL 16 with pgvector extension
• Authentication: djangorestframework-simplejwt (JWT Access/Refresh tokens)
• Error Monitoring: Sentry SDK (Backend + Celery workers)

AI & Machine Learning

• Vector Embeddings: Google AI models/text-embedding-004 (768 dimensions)
• Primary LLM: Google Gemini 2.5 Flash
• Fallback LLM: Groq llama-3.1-8b-instant
• Text Parsing: PyMuPDF (fitz), python-docx

Frontend

• Framework: React 18, Vite 5
• State Management: Zustand
• WYSIWYG Editor: Tiptap Editor (@tiptap/react)
• Data Visualization: Recharts
• HTTP Client: Axios with auto-refresh token interceptors
• Styling: Modern Vanilla CSS Design System with dark mode support

---

Security & Multi-Tenancy

Draftly enforces security at every layer of the application stack:

1. Row-Level Organization Scoping: All models explicitly reference an Organization. Queries automatically inject .filter(org=request.user.org) via permissions.
2. Permission Guardrails: Custom permission classes inspect subscription quotas before executing heavy Celery tasks.
3. Vector Store Scoping: Vector similarity queries in pgvector append standard SQL WHERE org_id = %s conditions to prevent cross-tenant context injection.
4. JWT Lifecycle: Short-lived access tokens with automatic token rotation via refresh endpoints.

---

Billing & Subscription Engine

Draftly features a fully implemented Stripe billing engine in apps/billing/:

• Tier Resolution: Resolves subscription tiers directly from Stripe line-item price IDs.
• Webhook Handling: Handlers for checkout.session.completed, customer.subscription.updated, customer.subscription.deleted, invoice.paid, and invoice.payment_failed.
• Idempotency: Prevents duplicate webhook processing using a dedicated StripeEvent event tracking model.

---

License & Contact

Distributed under the MIT License. See LICENSE for more information.

Hassan Zafar