--- title: Fitting RAG in Your Pocket: Local Retrieval in React Native description: A hands-on walkthrough builds a fully on-device Retrieval-Augmented Generation pipeline in React Native, using a small Qwen2.5-0.5B GGUF model via... canonical: https://daily.dev/posts/fitting-rag-in-your-pocket-local-retrieval-in-react-native-lte8vojis twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Fitting RAG in Your Pocket: Local Retrieval in React Native | daily.dev og:description: A hands-on walkthrough builds a fully on-device Retrieval-Augmented Generation pipeline in React Native, using a small Qwen2.5-0.5B GGUF model via... og:url: https://daily.dev/posts/fitting-rag-in-your-pocket-local-retrieval-in-react-native-lte8vojis og:image: https://api.daily.dev/og/posts/lte8vOJIs.png og:image:alt: Fitting RAG in Your Pocket: Local Retrieval in React Native og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Fitting RAG in Your Pocket: Local Retrieval in React Native **[Nitro Modules](https://daily.dev/sources/margelo)** · 24 min read · 0 upvotes · 0 comments ## Summary A hands-on walkthrough builds a fully on-device Retrieval-Augmented Generation pipeline in React Native, using a small Qwen2.5-0.5B GGUF model via @react-native-ai/llama for generation, a multilingual MiniLM embedding model for vectors, and sqlite-vec compiled directly into nitro-sqlite for native KNN search. It covers chunking WhatsApp exports by day, handling context window limits, fixing follow-up question grounding, and extends the same vector-search machinery to build an on-device photo search app using CLIP (via ONNX Runtime) for text-to-image, image-to-image, and camera-to-image search. The piece closes with a note on GraphRAG as a next direction for relationship-based queries. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How do I run retrieval-augmented generation entirely on a mobile device without a server or API key? Combine a small on-device LLM with a local vector database inside SQLite. Use Qwen2.5-0.5B-Instruct (4-bit GGUF, 469 MB) via @react-native-ai/llama for generation, embed text chunks with paraphrase-multilingual-MiniLM-L12-v2 (384 dimensions, ~119 MB), and store vectors in sqlite-vec compiled directly into nitro-sqlite for native KNN search, all offline. _daily.dev surfaces practical patterns like this for developers building offline-first AI features._ ### Why does my on-device LLM reject a long prompt instead of just truncating it? Because the context window is a fixed KV cache budget, and llama.cpp refuses requests that exceed it rather than truncating them, reporting a 'Context is full' error. For example, Qwen2.5-0.5B natively supports 32,768 tokens, but on a phone the context may be capped at 4,096 to leave RAM for the model and embedder, so a 21,000-token chat export gets rejected outright. _developers debugging context-window errors in mobile LLM apps can track fixes like this on daily.dev._ ### Which embedding model should I use for multilingual text search on a phone? Use paraphrase-multilingual-MiniLM-L12-v2 rather than an English-only model like all-MiniLM-L6 if users' text might mix languages. The multilingual model produces 384-dimensional vectors at roughly 119 MB in 4-bit quantization and trained on 50+ languages, whereas the English-only model collapses non-English messages to near-identical vectors, making search return noise on bilingual chats. _choosing the right embedding model for multilingual search is easier when comparisons like this surface on daily.dev._ ## Similar posts on daily.dev - [Making Very Small LLMs Smarter With RAG](https://daily.dev/posts/making-very-small-llms-smarter-with-rag-ivzwrmg6q) · Docker · 1 upvotes · 0 comments - [RAG Explained Simply with a Real Project](https://daily.dev/posts/rag-explained-simply-with-a-real-project-pgy15jzg7) · freeCodeCamp · 23 upvotes · 1 comments - [How to Build a Robust RAG System with Minimal Resources](https://daily.dev/posts/how-to-build-a-robust-rag-system-with-minimal-resources-h9d0rodtw) · Machine Learning Mastery · 1 upvotes · 1 comments - [AI Crash Course: Retrieval Augmented Generation \(RAG\)](https://daily.dev/posts/ai-crash-course-retrieval-augmented-generation-rag--f2som1u5k) · Telerik · 1 upvotes · 0 comments --- Tags: [#react-native](https://daily.dev/tags/react-native), [#rag](https://daily.dev/tags/rag), [#vector-search](https://daily.dev/tags/vector-search), [#sqlite](https://daily.dev/tags/sqlite), [#local-ai](https://daily.dev/tags/local-ai) [View this post on daily.dev](https://daily.dev/posts/fitting-rag-in-your-pocket-local-retrieval-in-react-native-lte8vojis) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Fitting RAG in Your Pocket: Local Retrieval in React Native","url":"https://daily.dev/posts/fitting-rag-in-your-pocket-local-retrieval-in-react-native-lte8vojis","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/fitting-rag-in-your-pocket-local-retrieval-in-react-native-lte8vojis"},"datePublished":"2026-08-23T12:19:47.738Z","dateModified":"2026-08-23T12:42:02.462Z","description":"A hands-on walkthrough builds a fully on-device Retrieval-Augmented Generation pipeline in React Native, using a small Qwen2.5-0.5B GGUF model via...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/f7d62cbb2a7f840eee843ff4dfbd417e?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/f7d62cbb2a7f840eee843ff4dfbd417e?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Nitro Modules","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Nitro Modules","logo":"https://media.daily.dev/image/upload/s--5qrQGNNm--/f_auto,q_auto/v1787487502/logos/margelo?_a=BAMAMicg0","url":"https://daily.dev/sources/margelo"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/fitting-rag-in-your-pocket-local-retrieval-in-react-native-lte8vojis","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"react-native,rag,vector-search,sqlite,local-ai","timeRequired":"PT24M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Nitro Modules","item":"https://daily.dev/sources/margelo"},{"@type":"ListItem","position":3,"name":"Fitting RAG in Your Pocket: Local Retrieval in React Native"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/fitting-rag-in-your-pocket-local-retrieval-in-react-native-lte8vojis#faq","mainEntity":[{"@type":"Question","name":"How do I run retrieval-augmented generation entirely on a mobile device without a server or API key?","acceptedAnswer":{"@type":"Answer","text":"Combine a small on-device LLM with a local vector database inside SQLite. Use Qwen2.5-0.5B-Instruct (4-bit GGUF, 469 MB) via @react-native-ai/llama for generation, embed text chunks with paraphrase-multilingual-MiniLM-L12-v2 (384 dimensions, ~119 MB), and store vectors in sqlite-vec compiled directly into nitro-sqlite for native KNN search, all offline. daily.dev surfaces practical patterns like this for developers building offline-first AI features."}},{"@type":"Question","name":"Why does my on-device LLM reject a long prompt instead of just truncating it?","acceptedAnswer":{"@type":"Answer","text":"Because the context window is a fixed KV cache budget, and llama.cpp refuses requests that exceed it rather than truncating them, reporting a 'Context is full' error. For example, Qwen2.5-0.5B natively supports 32,768 tokens, but on a phone the context may be capped at 4,096 to leave RAM for the model and embedder, so a 21,000-token chat export gets rejected outright. developers debugging context-window errors in mobile LLM apps can track fixes like this on daily.dev."}},{"@type":"Question","name":"Which embedding model should I use for multilingual text search on a phone?","acceptedAnswer":{"@type":"Answer","text":"Use paraphrase-multilingual-MiniLM-L12-v2 rather than an English-only model like all-MiniLM-L6 if users' text might mix languages. The multilingual model produces 384-dimensional vectors at roughly 119 MB in 4-bit quantization and trained on 50+ languages, whereas the English-only model collapses non-English messages to near-identical vectors, making search return noise on bilingual chats. choosing the right embedding model for multilingual search is easier when comparisons like this surface on daily.dev."}}]} ```