How to vibe codeScite

Prerequisites

• Node.jsfree

  Required runtime for Next.js and ingestion scripts

• GitHubfree

  Repository hosting and CI/CD deployment connection

AI coding tools

• Claude Code

  Best-in-class terminal agent for scaffolding complex RAG ingestion pipelines and vector storage

• Cursor

  Ideal for iterative UI work on the chat interface and search dashboard

Build guide

1. 01Project Scaffolding & Database Schema

   Initialize a Next.js application with TypeScript, Tailwind CSS, and shadcn/ui components. Configure Supabase with pgvector extension enabled for storing open-access academic paper embeddings.

   Create a new Next.js 16 project using TypeScript, App Router, and Tailwind CSS. Initialize shadcn/ui configuration with base styles. Set up Supabase client utilities and write migration SQL files to create tables for 'papers' (id, doi, title, abstract, full_text, published_date, source_url), 'chunks' (id, paper_id, content, embedding vector(1536)), and 'citations' (id, citing_paper_id, cited_paper_id, context_sentence, intent_classification). Ensure pgvector extension is enabled in the migration script. Write an initialization test verifying database connectivity.

   ⧉copy promptopen in Claude CodeCopen in Codexopen in Cursor

2. 02Open Access Ingestion Pipeline

   Build a background ingestion script that fetches open-access papers and preprints from external open APIs (e.g., PubMed Central Open Access subset, arXiv API, Unpaywall), chunks their text, and stores vectors.

   Build a TypeScript ingestion script in lib/ingestion/ that queries the PubMed Central Open Access API and arXiv API for recent papers in a specific field. Parse the XML/PDF responses to extract title, abstract, sections, and reference lists. Implement a text-chunking utility that splits academic text into 500-token chunks with 50-token overlap. Use OpenAI text-embedding-3-small to generate vector embeddings for each chunk, and batch insert papers, chunks, and embeddings into the Supabase database. Add robust error handling for rate limits and malformed XML.

   ⧉copy promptopen in Claude CodeCopen in Codexopen in Cursor

3. 03Smart Citation Intent Classifier

   Implement an LLM-based classification pipeline that analyzes citation contexts within papers and classifies them into supporting, contrasting, or mentioning categories.

   Create an AI classification module in lib/classifier/ that takes a target paper's reference list and the sentences surrounding each citation marker in downstream papers. Using the Vercel AI SDK and Anthropic Claude Haiku, prompt the model to classify each citation intent as 'supporting', 'contrasting', or 'mentioning' based strictly on the sentence context. Store these classified citations in the 'citations' table with their respective intent tags and exact context snippets. Write an automated test suite with sample citation sentences to verify classification accuracy.

   ⧉copy promptopen in Claude CodeCopen in Codexopen in Cursor

4. 04Full-Text Hybrid Search & Vector Retrieval

   Develop a hybrid search engine combining Postgres full-text search (BM25 style via tsvector) and vector similarity search over paper chunks, enhanced with Cohere Rerank.

   Implement a search service in lib/search/ that performs hybrid search across the ingested scholarly database. Combine PostgreSQL full-text search (tsvector on title, abstract, and chunks) with pgvector cosine similarity search. Pass the combined result set of top 50 matches through the Cohere Rerank API to re-order results by relevance. Expose this via a Next.js API endpoint /api/search supporting query filters by date range, author, and citation intent counts.

   ⧉copy promptopen in Claude CodeCopen in Codexopen in Cursor

5. 05Grounded RAG Assistant Chat Interface

   Build a conversational RAG assistant UI where users can ask research questions and receive answers strictly grounded in retrieved literature with inline verifiable citations.

   Build a chat interface component using Tailwind CSS and shadcn/ui resembling a research assistant dashboard. Implement the corresponding API route at /api/assistant using Vercel AI SDK's streamText with Claude Sonnet. The backend must first query the hybrid search engine for relevant paper chunks, inject them into the system prompt with strict grounding instructions ('Only answer based on the provided literature; cite exact sentence IDs inline'), and stream the response back. Render inline citation badges that open a modal preview of the cited paper chunk and its Smart Citation status.

   ⧉copy promptopen in Claude CodeCopen in Codexopen in Cursor

6. 06Collections & Reference Manager Integration

   Add user collection management to save papers, organize them into folders, and provide a basic export format compatible with reference managers like Zotero (BibTeX / RIS).

   Implement a collections system using better-auth for user session management. Create database tables for 'collections' and 'collection_papers'. Build a dashboard UI allowing users to create folders, save papers from search or assistant results, and view collection details. Add an export feature that generates BibTeX (.bib) and RIS format strings for any collection so users can import their saved papers directly into Zotero or Mendeley.

   ⧉copy promptopen in Claude CodeCopen in Codexopen in Cursor

Cost vs paying for Scite