Summarist.ai

Summarist.ai: AI-Powered PDF Summarization & Chat Platform

Overview

Summarist.ai helps users understand long PDFs faster by turning dense documents into structured AI summaries and interactive document conversations. Instead of manually scanning pages for key ideas or searching through a PDF for specific answers, users can upload a document, receive a clean summary, and ask follow-up questions grounded in the actual PDF content.

The project started as a simple PDF summary generator, but evolved into a complete document intelligence SaaS with authentication, subscriptions, usage limits, background processing, realtime progress updates, semantic search, and a unified vault for saved summaries and chats.

Impact

• Reduced manual reading effort by converting long PDFs into structured summaries with overview, key points, sections, and action items
• Helped users move from static document reading to interactive exploration through AI-powered PDF chat
• Supported PDFs up to 32MB with reliable upload handling through UploadThing
• Built a full RAG pipeline using 3,072-dimensional Gemini embeddings and pgvector similarity search
• Improved user experience during long-running AI tasks with realtime progress updates and polling fallback
• Created a subscription-ready SaaS foundation with Clerk authentication, Polar billing, monthly usage limits, and webhook-based user/subscription sync

What Users Can Do

• Upload PDFs: Drag and drop or select PDF files up to 32MB
• Generate AI Summaries: Create structured summaries with title, read time, overview, key points, sections, and action items
• Chat with PDFs: Ask questions about uploaded documents and get context-aware answers using RAG
• Track Real-time Processing: See live progress while PDFs are parsed, chunked, embedded, and indexed
• Use a Unified Vault: Browse all generated summaries and PDF chat sessions in one place
• View Rich Summaries: Read summaries in a clean interactive viewer with collapsible sections
• Export Summaries: Download summaries as Plain Text, Markdown, or Word-compatible .doc files
• Manage Usage by Plan: Free, Pro, and Unlimited plans with separate limits for summaries and PDF chats
• Securely Access Content: Use protected dashboard routes powered by Clerk authentication

Why I Built This

I built Summarist.ai to solve the biggest pain points around reading and understanding long PDF documents:

• Reading long documents takes time: Research papers, reports, and documentation can be difficult to process quickly
• Static summaries are limited: A summary helps, but users often need answers to specific questions
• PDF search is not enough: Keyword search misses context and semantic meaning
• AI output needs structure: Raw AI responses are hard to scan, so summaries need predictable formatting
• Document workflows need persistence: Users should be able to return to previous summaries and chats anytime
• AI features need reliable infrastructure: Long-running parsing and embedding tasks need durable background processing

Tech Stack

Frontend

• Next.js 16: App Router, Server Components, API routes, and server actions
• React 19: Component-based UI architecture
• TypeScript: Type-safe frontend, backend, and database interactions
• Tailwind CSS 4: Utility-first styling system
• Shadcn UI / Radix UI: Accessible UI primitives and reusable components
• Framer Motion / Motion: Smooth page transitions and micro-interactions
• Lenis: Smooth scrolling experience
• Sonner: Toast notifications for upload, processing, and error states

Backend & Services

• Google Gemini 2.5 Flash: Summary generation and chat responses
• Gemini Embedding 001: 3,072-dimensional embeddings for semantic search
• LangChain: PDF loading and text splitting
• NeonDB PostgreSQL: Serverless relational database
• pgvector: Vector storage and similarity search
• Drizzle ORM: Type-safe schema and queries
• Inngest: Durable background jobs for PDF chat processing
• Inngest Realtime: Live processing updates from background jobs
• Clerk: Authentication, user sessions, and user webhooks
• UploadThing: Secure PDF upload handling
• Polar: Checkout, subscriptions, customer portal, and billing webhooks

Key Features

AI-Powered PDF Summarization

• Users upload a PDF through UploadThing
• The server extracts PDF text using LangChain's PDFLoader
• Extracted text is sent to Gemini 2.5 Flash with a strict structured prompt
• Gemini returns JSON containing:
  • title
  • readTime
  • overview
  • keyPoints
  • sections
  • actionItems
• The app strips markdown code fences if Gemini wraps the JSON response
• The summary is saved to NeonDB and displayed in a polished viewer
• If JSON parsing fails, the UI gracefully falls back instead of crashing

Chat with PDF — Full RAG Pipeline

1. Upload: User uploads a PDF and chooses chat mode
2. Record Creation: A chat_pdf database record is created with status processing
3. Inngest Event: The app sends a pdf/chat.uploaded event
4. Parse: Inngest fetches the PDF and loads page-level documents using LangChain PDFLoader
5. Chunk: Text is split into 1,000-character chunks with 200-character overlap
6. Embed: Each chunk is embedded using gemini-embedding-001
7. Store: Chunk text, page number, and vector embedding are stored in pdf_chunks
8. Ready State: The PDF status changes to ready
9. Ask Question: User messages are embedded and compared against stored chunks
10. Retrieve Context: Top 5 relevant chunks are fetched using pgvector similarity search
11. Generate Answer: Gemini receives the retrieved context and user question
12. Save History: User and assistant messages are saved in chat_messages

Real-time PDF Processing

• Inngest publishes progress updates for every processing stage:
  • processing
  • parsing
  • chunking
  • embedding
  • ready
  • error
• Each PDF gets its own realtime channel: pdf-processing:<chatPdfId>
• The frontend subscribes using @inngest/realtime
• If realtime fails or closes, the UI falls back to polling /api/chat/[id]/status
• Users get smooth loading states instead of waiting blindly during long embedding jobs

Unified Vault

• The vault combines saved summaries and PDF chat sessions into one feed
• Summary items come from pdf_summaries
• Chat items come from chat_pdf
• Items are sorted by latest activity:
  • chats use updatedAt
  • summaries use createdAt
• Each vault card links to either the summary viewer or chat interface

Subscription & Usage Limits

• Polar handles checkout, subscriptions, customer portal, and webhook events
• Plans include:
  • Free: 2 summaries/month and 2 PDF chats/month
  • Pro: 10 summaries/month and 10 PDF chats/month
  • Unlimited: High-limit access for summaries and chats
• Usage is calculated monthly using database counts
• Both summary generation and PDF chat creation are gated by plan limits
• Polar webhooks update subscription status, product ID, customer ID, and billing period

Technical Implementation

Summary Generation Flow

1. User selects summary mode and uploads a PDF
2. UploadThing stores the file and returns the file URL
3. Server action verifies the authenticated Clerk user
4. PDF text is extracted using LangChain PDFLoader
5. Gemini generates a structured JSON summary
6. The app extracts the generated title from the JSON
7. Summary is saved in the pdf_summaries table
8. Dashboard and vault paths are revalidated
9. User is redirected to the summary viewer page

Chat PDF Flow

1. User selects chat mode and uploads a PDF
2. Server action verifies the user and checks chat usage limits
3. A chat_pdf record is created with status processing
4. Inngest receives a pdf/chat.uploaded event
5. Background function parses the PDF, chunks text, embeds content, and stores vectors
6. Processing status updates are published in realtime
7. Once ready, the user can ask questions
8. Each user question is embedded and matched against the stored PDF chunks
9. Gemini answers using only the retrieved document context
10. Chat history is persisted and loaded on future visits

Database Schema

• users: Stores app user records mapped to Clerk IDs
• subscriptions: Stores Polar customer, subscription, product, and status data
• pdf_summaries: Stores generated summaries, titles, file names, and source file URLs
• chat_pdf: Stores uploaded PDFs prepared for chat, including processing status
• pdf_chunks: Stores chunked PDF text, page numbers, and pgvector embeddings
• chat_messages: Stores user and assistant messages for each PDF chat

Technical Challenges & Solutions

Challenge 1: Reliable PDF Text Extraction

• Problem: PDFs can vary in structure, length, and formatting
• Solution: Used LangChain's PDFLoader to load PDF content page-by-page, then combined text for summaries or preserved page metadata for chat chunks

Challenge 2: Structured AI Output

• Problem: AI responses can be inconsistent or wrapped in markdown fences
• Solution: Designed a strict JSON prompt and added post-processing to strip code fences before parsing. The summary viewer also has a graceful fallback for invalid JSON

Challenge 3: Building a RAG Pipeline

• Problem: Chat answers need to be grounded in uploaded documents, not generic AI knowledge
• Solution: Implemented a custom RAG pipeline using Gemini embeddings, pgvector similarity search, and context injection into Gemini chat prompts

Challenge 4: pgvector with Drizzle ORM

• Problem: Drizzle does not provide first-class pgvector support out of the box
• Solution: Created a custom Drizzle vector type for vector(3072) and used raw SQL for vector inserts and similarity queries

Challenge 5: Long-running PDF Processing

• Problem: Parsing, chunking, embedding, and storing large PDFs can take time
• Solution: Moved chat PDF processing into an Inngest background function with retries, status updates, and error handling

Challenge 6: Realtime Processing Updates

• Problem: Users need feedback while background jobs are running
• Solution: Used Inngest Realtime to publish per-document progress events and added a polling fallback for reliability

Challenge 7: Gemini Chat History Format

• Problem: Gemini requires chat history to alternate strictly between user and model messages
• Solution: Rebuilt history from stored messages by including only complete user → assistant pairs and skipping incomplete turns

Challenge 8: User Sync Reliability

• Problem: Clerk webhooks may not always create the database user before the first app action
• Solution: Added ensureFreeUserExists fallback checks across dashboard, vault, upload, chat, credits, and delete actions

Challenge 9: Subscription Lifecycle Handling

• Problem: Billing state must stay synced with the app database
• Solution: Integrated Polar webhooks for subscription activation, updates, cancellation, revocation, and customer state changes

After Launch & Impact

• Built a complete AI SaaS platform with authentication, uploads, payments, usage limits, and persistent user data
• Expanded the original PDF summarizer into a document intelligence tool with RAG-based PDF chat
• Implemented a production-style vector search pipeline using pgvector and Gemini embeddings
• Built durable background processing with Inngest and realtime progress updates
• Integrated multiple third-party services into one coherent system: Clerk, UploadThing, Gemini, NeonDB, Inngest, and Polar
• Improved user experience with a unified vault, structured summary viewer, chat history, and export options
• Gained practical experience with AI workflows, vector databases, subscription systems, and serverless architecture

Future Plans

• Add public shareable summary links
• Support more file formats such as DOCX, TXT, and EPUB
• Add batch PDF upload and processing
• Add search, filters, and tags in the Vault
• Add PDF export for generated summaries
• Allow users to edit and annotate summaries
• Add source citations directly inside chat answers
• Improve OCR support for scanned PDFs
• Add team workspaces and shared document libraries
• Build a public API for third-party integrations