ChatPDF docs ChatPDF turns uploaded PDFs into searchable, cited conversations. This is how the product works and how to run it yourself. Get started Documentation Frontend Next.js App Router Backend Convex Auth Clerk Upload and index PDFs Follow a document from upload through OCR, chunking, and embeddings. Ask grounded questions Routed, hybrid retrieval returns answers backed by verifiable citations. Self-host the app Wire up Convex, Clerk, OpenAI, Mistral OCR, and storage. What ChatPDF is ChatPDF is a protected document workspace for uploading PDFs, indexing their contents, and asking questions against the evidence inside those files. Answers stay grounded in the uploaded documents instead of generic model knowledge, and every claim can be traced back to a page. For readers Upload research papers, contracts, manuals, or reports and ask questions in natural language. For builders Run it on Next.js, Convex, Clerk, OpenAI, and Mistral OCR. The core loop Every document moves through the same path: it is uploaded, processed into searchable evidence, and then queried with grounded, cited answers. Upload a PDF The dashboard accepts text-based and scanned PDFs up to 100 pages. Process the document Every file is OCR'd with Mistral OCR 4, then chunked, embedded, and summarized. Ask questions A router picks a retrieval strategy, hybrid search finds evidence, and the answer streams in. Verify citations Each answer points back to verbatim, page-level source material in the PDF viewer. The ingestion pipeline. Every PDF is OCR'd, chunked, embedded, and summarized before it becomes available for chat. flowchart TB U(["PDF uploaded"]) --> S["Stored in Convex file storage
+ document record created"] S --> O["Mistral OCR 4
OCR · every PDF · ≤ 100 pages"] O --> X["Extract page text
from OCR output"] X --> C["Chunk text
~450 words · 75 overlap
page spans kept for citations"] C --> E["Embed chunks
text-embedding-3-small · 1536-dim"] X --> P["Per-page summaries
LLM"] P --> D["Document summary
LLM"] E --> R(["Ready for chat"]) D --> R O -. "transient error" .-> RT{"Retry?
≤ 3 attempts"} RT -. "yes · 15s / 60s backoff" .-> O RT -. "no" .-> F(["Failed"]) class R okNode class F badNode classDef okNode fill:#0e1f15,stroke:#34d399,color:#d1fae5; classDef badNode fill:#23100f,stroke:#f87171,color:#fecaca; How it fits together A Next.js frontend talks to a Convex backend that owns the database, server functions, background OCR jobs, and the streaming chat endpoint. External AI and storage services are called from Convex. High-level architecture. Convex coordinates everything; OpenAI and Mistral OCR are called from the backend while PDFs stay in Convex storage. flowchart LR subgraph client["Browser"] NX["Next.js App Router
landing · docs · dashboard"] end subgraph backend["Convex backend"] FN["Queries · mutations · actions"] HTTP["HTTP action
SSE chat stream"] DB[("Database
vector + text search indexes")] JOB["Scheduler
OCR jobs · retries"] end NX <-->|"reactive queries"| FN NX <-->|"POST · stream"| HTTP FN --> DB HTTP --> DB JOB --> DB CL["Clerk
Google · GitHub"] --- NX CL --- FN JOB --> MOCR["Mistral OCR 4"] JOB --> FS["Convex file storage"] FN --> OAI["OpenAI
embeddings · chat"] HTTP --> OAI NX -. "events" .-> PH["PostHog"] Where to go next Quickstart Install dependencies and run the app locally. Upload your first PDF Walk through a real upload, processing run, and question. Retrieval ranking Routing, hybrid search, rank fusion, and citation validation. Citations and sources How answers stay auditable and page-linked.

Troubleshooting

Common issues

Sign-in fails or redirects loop

Dashboard shows no data after sign-in

Uploads or OCR processing fail

Answers look weak

Build or typecheck fails