iAny

Feed AI from anything. The more you feed it, the smarter it gets — 100% offline, on your device.

iAny is an offline-first RAG (retrieval-augmented generation) PWA. You feed it text and documents; it chunks, embeds and indexes them in a real Postgres database running inside your browser. Then you ask questions — in Khmer or English — and a local LLM answers grounded in your own knowledge, with sources. Nothing ever leaves your device.

Tech stack

Layer	Technology
App	React 19 + Vite + TypeScript, installable PWA (vite-plugin-pwa)
Database	PGlite (WASM Postgres) + pgvector, persisted to IndexedDB, running in a Web Worker
Embeddings	EmbeddingGemma 300M via Transformers.js, truncated to 256 dims (Matryoshka)
Generation	Gemma 4 E2B via Transformers.js on WebGPU
Retrieval	Hybrid search: HNSW cosine similarity + Postgres full-text search, fused with reciprocal rank fusion
Sync (planned)	ElectricSQL — the schema is already sync-ready (UUIDs, updated_at, soft deletes)

Architecture

UI (React PWA, service worker precaches the full app shell incl. WASM)
 │
 ├─ AI Worker ─── Transformers.js
 │     ├─ EmbeddingGemma 300M  (ingest + query embedding, WebGPU or WASM)
 │     └─ Gemma 4 E2B          (answers, WebGPU only — search mode elsewhere)
 │
 └─ DB Worker ─── PGlite (idb://iany-db)
       ├─ chunks.embedding vector(256)  → HNSW index (semantic search)
       ├─ chunks.tsv tsvector           → GIN index (keyword search)
       └─ hybrid ranking via reciprocal rank fusion

• src/ai/ — AI worker, RPC client, Khmer-aware chunker
• src/db/ — DB worker, schema, hybrid search, documents, knowledge packs
• src/rag/ — ingest (feed) and ask (retrieve + generate) pipelines
• src/i18n/ — English and Khmer UI translations
• src/views/ — Chat, Library, Packs, Settings

Khmer support

Khmer is written without spaces between words, which breaks conventional full-text search. iAny uses Intl.Segmenter's ICU dictionary segmentation (built into every modern browser) to split Khmer into words at ingest time (ភ្នំពេញជារាជធានី → ភ្នំពេញ ជា រាជធានី) and indexes the result with Postgres FTS. Sentence segmentation (for chunking) handles the Khmer khan (។) the same way. EmbeddingGemma is multilingual (100+ languages including Khmer), so semantic search works in both languages out of the box, and cross-language retrieval (ask in Khmer, find English content) comes free.

Knowledge packs

A pack is iAny's portability primitive: a self-contained JSON bundle of documents + chunks + embeddings. Because every iAny install pins the same embedding model and dimensions, packs import with zero re-embedding and are instantly searchable. Packs serve as manual backup/device transfer today and are the foundation for pack sharing and a future marketplace. Source text is always included (RAG needs it, and it makes packs survivable across future embedding-model upgrades).

Models

All of iAny's models download once from Hugging Face and are cached on-device, then work fully offline. The app asks the browser for persistent storage so neither the models nor your knowledge base get evicted.

Answering + search (PWA):

• EmbeddingGemma 300M (q4, ~200 MB) — required; loads on first feed/search.
• iAny Khmer (~270M, the small default) or Gemma 4 E2B (q4f16, ~1.5 GB, needs WebGPU) — the answering model. Without a generator iAny runs in search mode: you still get the most relevant passages, just not a generated answer.

Open Khmer models — 🤗 https://huggingface.co/sengtha

iAny publishes its Khmer models open source. Browse them all on the profile above; the full catalog (purpose, format, license lineage) is in docs/MODELS.md.

Model	Purpose
iany-khmer-tiny-v1-ONNX · Qwen3-0.6B-khm-ft3-…-GGUF	Khmer answering LLM (PWA / mobile)
whisper-tiny-khmer	Khmer speech-to-text (STT)
khmer-tts-female-v2	Khmer text-to-speech (TTS)
khmer-ocr	Khmer OCR (detector + recognizer)

Models and datasets carry their own licenses, separate from iAny's Apache-2.0 code license — each Hugging Face repo card is authoritative. Help improve them by contributing at /voice, /scan, and /sign.

Development

npm install          # ONNXRUNTIME_NODE_INSTALL_CUDA=skip npm install if the
                     # onnxruntime-node postinstall can't reach the network
npm run dev          # dev server
npm run build        # typecheck + production build
npm run preview      # serve the production build
node scripts/generate-icons.mjs   # regenerate PWA icons from public/icon.svg

Deployment (Cloudflare Workers)

iAny deploys on Cloudflare Workers (wrangler.jsonc): static assets plus a small edge worker (worker/index.ts) that acts as a pull-through model mirror. The app downloads models from /models/* on its own origin; on a cache miss the worker fetches the file from Hugging Face through Cloudflare's network, streams it to the client and stores it in R2 in the same pass. This matters because many client networks (mobile carriers, filtered Wi-Fi) cannot reach huggingface.co directly — after the first download, models are served entirely from R2 with free egress.

npx wrangler login                         # once
npx wrangler r2 bucket create iany-models  # once
npm run deploy                             # build + wrangler deploy

Set localStorage iany.modelHost to override the model source (e.g. https://huggingface.co to bypass the mirror during local development). Settings → "Test model download connections" probes each hop on-device.

Or connect the repo with Workers Builds for deploy-on-push: build command npm run build, deploy command npx wrangler deploy, and set the build environment variable ONNXRUNTIME_NODE_INSTALL_CUDA=skip (the onnxruntime-node postinstall otherwise tries to download CUDA binaries and can fail CI).

When the v2 API lands, add "main": "worker/index.ts" plus R2/D1 bindings to wrangler.jsonc; Worker routes automatically take precedence over static assets.

Sharing models between devices (no internet)

Settings → AI models → Export: bundles a downloaded model's cached weights into a single .iany-model file you can move via Quick Share, USB or SD card, then Import on the receiving device — the model is ready with zero download. One person downloads once and can seed many offline devices.

Cloud backup (v2, beta)

Settings → Cloud backup encrypts your knowledge base on-device (AES-GCM; key derived from a recovery code that never leaves the device) and stores the ciphertext via /api/backup/* in R2. Restoring on a new device just needs the recovery code. The server cannot read backups — losing the code means losing the backup. Free during beta; the future credits system will gate this endpoint.

Roadmap

1. v1 (done) — offline feed/ask, hybrid Khmer/English retrieval, knowledge packs, model mirror, device-to-device model sharing.
2. v2 (in progress) — encrypted cloud backup/restore (done, beta); next: credits & payments, live multi-device sync via ElectricSQL.
3. v3 — knowledge pack marketplace (buy/sell packs, platform fee).

License

iAny's source code is licensed under Apache-2.0 — free to use, modify, and build on, including for commercial products, with an explicit patent grant. See NOTICE for third-party attributions.

The code license covers code only. Machine-learning models and datasets carry their own licenses (e.g. community datasets are CC-BY-SA-4.0, some models follow Google's Gemma Terms). See NOTICE.

Open source & commercial

iAny is fully open source and a sustainable project. Revenue comes from building on it, support, a hosted service + credits, B2B/B2G contracts, and grants — not from restricting the code. Anyone may build derivative products. See COMMERCIAL.md for the full model, and CONTRIBUTING.md to get involved.

Trademark

The Apache-2.0 license grants rights to the code, not the name. "iAny" and the iAny logo are trademarks of E-KHMER Technology Co., Ltd. You're free to build and even sell products from this code, but please give your product its own name and don't imply it's the official iAny or endorsed by it. Nominative use ("built with iAny", "based on iAny") is fine.

Security

Found a vulnerability? Please report it privately — see SECURITY.md. Do not open a public issue.