react-native-pageindex

It takes documents — PDFs, Word files, CSVs, Excel spreadsheets, and Markdown — and builds two indexes: a hierarchical forward index (tree structure with chapters, sections, and page ranges) and an inverted reverse index (term-to-node mappings for fast search). You can then search the index locally or run multi-turn conversational Q&A over the indexed content.

Yes. The keyword mode builds and searches the reverse index using TF-IDF scoring with zero LLM calls and zero network access. Build the index once, and the entire search and retrieval loop works offline indefinitely. The LLM mode requires an API call during indexing, but once the index is built, keyword search over it is fully offline.

React Native developers, mobile architects, AI product teams, and anyone building document-driven features in mobile or edge applications. It's particularly useful for teams serving users in low-connectivity environments — healthtech, agritech, education, enterprise field tools.

Traditional RAG converts text into vector embeddings and stores them in a vector database for similarity search. react-native-pageindex skips embeddings entirely. It uses LLM reasoning to build a structured tree, then creates a TF-IDF inverted index for search. No vector database, no embedding model, no similarity computation. The tradeoff is that keyword search is exact-match rather than semantic — but it's instant, offline, and has zero infrastructure requirements.

Any provider that can respond to a text prompt. The library defines a simple LLMProvider interface — you implement a function that takes a prompt and returns a response. The README includes ready-to-use examples for OpenAI, Anthropic Claude, Google Gemini, and Ollama (fully local). You can wire up any model or endpoint.

Yes. The built-in chat pattern follows a standard RAG loop: retrieve relevant nodes via searchReverseIndex, build grounded context from page text or node summaries, pass it to your LLM with conversation history, and return cited answers with source metadata. The library provides the retrieval and context layers — you bring the LLM.

All three. The core library is pure TypeScript with no platform-specific dependencies. It works in React Native, Node.js, and browser environments. The demo app is a Vite + React app that runs entirely in the browser. PDF extraction uses pdfjs-dist which works in browsers and Node.js.

Keyword mode is near-instant — a 100-row CSV indexes in under a second with zero API calls. LLM mode depends on your provider and document size. A 32-page PDF takes roughly 3-4 minutes with gpt-4o-mini due to the multi-step reasoning pipeline (TOC detection, verification, summary generation). Once indexed, search is always instant.

The library is at version 0.1.x, following semantic versioning. The API surface is stable and fully typed. It's designed for production use in architectures where offline search and structured document retrieval are needed. As with any early-version library, it's worth evaluating against your specific requirements.

When your users don't have reliable internet. When you need search results without network latency. When documents are sensitive and shouldn't leave the device. When you want retrieval without infrastructure overhead. When you're building for mobile-first, offline-first, or edge environments. If your users always have fast, reliable connections and you're already invested in cloud search infrastructure, a cloud service may be more appropriate.