Document Assistant
mihairusu88/document-assistant
Chat RAG por conversa com visualização interativa da pipeline
Ver no GitHubAplicação RAG estilo ChatGPT que permite fazer upload de documentos e fazer perguntas respondidas pelo próprio documento. Combina Next.js 16, Claude, Qdrant e embeddings Voyage em uma pipeline visual que mostra cada etapa do processamento (extração → chunking → embedding → armazenamento → recuperação → geração).
Instalação
npx convez-ai mihairusu88/document-assistant1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Análise de documentos técnicos e PDFs em conversa privada
- Assistente de pesquisa com isolamento de contexto por thread
- Prototipagem de aplicações RAG com pipeline visual
- Comparação em tempo real de respostas com/sem retrieval
Veredito

Projeto bem estruturado em Next.js 16 com TypeScript e RAG production-ready. Stack moderno (Voyage + Qdrant + Claude) e arquitetura limpa por serviços, mas requer setup inicial de múltiplas dependências externas (Supabase, Qdrant, Anthropic).
README
A tradução para o português deste README ainda está na fila.
Document Assistant
A ChatGPT-style RAG chat app: upload a document into a conversation, then ask questions answered from that document. Retrieval is scoped per conversation, so each thread only draws on the files you added to it. A live pipeline drawer visualises every RAG step (extract → chunk → embed → store → retrieve → generate), and a toggle lets you compare answers with retrieval on vs. off.
Stack
Next.js 16 (App Router, Turbopack) · React 19 · TypeScript · Tailwind v4 · hand-authored shadcn/ui · Zod · AI SDK v6
Supabase (Google OAuth + Postgres, accessed via @supabase/supabase-js + RLS) · Voyage AI embeddings (voyage-3-large) · Qdrant vector store · Anthropic Claude (claude-sonnet-4-6) for generation.
How it works
Ingest (POST /api/documents, NDJSON stream) Answer (POST /api/chat, UI message stream)
1. Extract text ExtractTextService 5. Retrieve RetrievalService (Qdrant search)
2. Chunk ChunkingService 6. Generate ChatService (Claude, streamed)
3. Embed EmbeddingsService (Voyage)
4. Store StoreChunksService (Qdrant)
Chunks are stored in Qdrant tagged with their conversationId; retrieval filters on it so answers are grounded only in the current conversation's documents. Conversations and messages are persisted to Supabase Postgres (RLS-scoped to the signed-in user).
Project structure
src/
app/
(auth)/ # sign-in / sign-up / password reset
(app)/ # authed shell: /chat and /chat/[conversationId]
api/ # chat, conversations, conversations/[id], documents
auth/callback # OAuth code exchange
components/
layout/ # appSidebar, userMenu (app chrome)
ui/ # hand-authored shadcn primitives (camelCase filenames)
services/ # business + RAG pipeline (one service per step) + supabase persistence
views/ # route-aligned feature modules
chat/ # chatWindow, conversationSidebar, the RAG drawer + zustand store
conversations/hooks/ # useDocumentUpload (ingest stream reader)
users/ # auth server actions + forms
lib/supabase/ # browser / server clients
types/ # api envelope + UI DTOs
utils/ # cn, files, apiClient
proxy.ts # Next 16 middleware: session refresh + route guard
supabase/schema.sql # run by hand in the Supabase SQL Editor
The feature folder is
views/, notpages/—pagesis a reserved Next.js directory (the legacy Pages Router). Conventions: camelCase filenames, feature folders named to match routes.
Getting started
npm install
cp .env.example .env.local # fill in Supabase, Qdrant, Voyage, and Anthropic values
npm run dev # or: npm run dev:proxy (if you're behind a TLS-intercepting proxy)
Then, once per project:
- Database — open the Supabase Dashboard → SQL Editor and run
supabase/schema.sql(createsconversations+messageswith RLS). - Auth — enable the Google provider in Supabase (Authentication → Providers), set the OAuth client's redirect URI to
https://<project-ref>.supabase.co/auth/v1/callback, and set the Site URL tohttp://localhost:3000.
The UI compiles and renders with placeholder env values; sign-in, retrieval and generation activate once real credentials are present.
Environment
See .env.example. Public: NEXT_PUBLIC_SUPABASE_URL, NEXT_PUBLIC_SUPABASE_ANON_KEY, NEXT_PUBLIC_SITE_URL. Server-only secrets (never NEXT_PUBLIC_): QDRANT_URL, QDRANT_API_KEY, VOYAGE_API_KEY, ANTHROPIC_API_KEY.
Scripts
dev · dev:proxy · build · start · lint · typecheck
There is no test suite — npm run typecheck and npm run lint are the correctness gates.
Screenshots
Dark theme



Light theme


