dazzle_flutter 1.0.0-beta.6

Dazzle SDK for Flutter. Embedded in-process database with HNSW vector search and a ChatAgent runtime for on-device LLM agents. Links the same libdazzle.so / Dazzle.xcframework the native Android and i [...]

Changelog

All notable changes to dazzle_flutter. This package follows the Dazzle SDK release line; see the repo CHANGELOG for cross-stack release notes.

1.0.0-beta.6 — 2026-05-12

Paper-side updates (applies to every binding — the SDK runs the same harness everywhere)

• §5.9.5 cross-platform extension. End-to-end RAG reproduction now spans 5 physical mobile SoCs across 2 operating systems: Unisoc T760 / Cortex-A76 (Moto G35 5G, Android 14), QCOM SD662 / Cortex-A73 (Moto G30, Android 11), HiSilicon Kirin 659 / Cortex-A53 (Huawei P20 Lite, Android 9 / EMUI 9), MediaTek Helio G80 / Cortex-A75 (Huawei Y9a / FRL-L23, Android 10 / EMUI 10), and Apple A14 / Firestorm (iPhone 12 Pro, iOS 26). Four Android microarchitecture generations plus Apple Firestorm. Bootstrap CIs (B=10000, paired-qid resampling) star-mark every F1_short ratio at 95% confidence across the spread.
• §5.9.6 quantization sensitivity sweep. Q4_K_M vs Q5_K_M on the two v8.2 HNSW chips (Moto G35 5G + Huawei Y9a). Headline: the em_contains metric is flat between quant levels (deltas ≤ 0.025 on every cell, inside the per-cell CI half-width). Latency reveals a bandwidth-vs-compute split — the faster A76 chip pays a +50% wall-clock tax for Q5, the slower A75 chip pays ≤+13% because it was already bandwidth-bound on Q4 weights. Disk footprint cost is +6.3% on the 0.5B GGUF and +15% on the 1.5B GGUF.
• REPRODUCIBILITY §4a + §4b. New recipes: per-device chunked instrumentation for §5.9.5 (any new chip can be added with a small set of am instrument invocations) and GGUF-swap for §5.9.6 (no rebuild needed to A/B Q4 vs Q5 vs future quant variants).
• PDFs regenerated. research/paper/arxiv-build/paper.pdf and paper_es.pdf rebuilt from the updated paper sources.

Inherited from the underlying native bindings

• All Android (sdk/android) and iOS (sdk/ios) fixes that ship with this Flutter plugin: the Kirin-pass-15 n_batch = n_ctx universal fix (now in both the SDK core and the Android experiment JNI), the iOS launch-watchdog dispatch fix, the DazzleServer.vectorIndex initialCapacity parameter, and the G80/EMUI-10 freeze unblock that was specific to the experiment-app JNI path.

1.0.0-beta.5

Added — Flutter Web (WebAssembly runtime)

• DazzleWeb, DazzleWebHash, DazzleWebVectorIndex — exported from the package's main library. Backed by dazzle.wasm (~236 KB) built from the same C++ source as the native iOS / Android binaries. Same on-device promise extended to the browser: in-process HNSW vector search + hash KV, no remote server.
• Persistence via the Origin Private File System (OPFS). DazzleWeb.persist() writes a binary snapshot; initialize() restores it on boot. Multi-user isolation via opfsFileName:.
• Setup contract: add a <script type="module"> tag to your app's web/index.html that imports assets/packages/dazzle_flutter/web/native/dazzle.js and assigns it to globalThis.dazzleModule. See the README "Flutter Web" section for the exact snippet.

Added — Flutter Desktop (Linux / macOS / Windows)

• DazzleDesktop, DazzleDesktopHash, DazzleDesktopVectorIndex — same API surface as DazzleWeb, backed by libdazzle_lite via dart:ffi. Compiled from the same C++ source as the WASM build, so behaviour is byte-for-byte identical across web and desktop.
• Plugin declares ffiPlugin: true for linux, macos, windows — pre-built native libraries ship inside the package so consumers don't need a host C++ toolchain.
• Persistence to a regular file on disk (default <cwd>/.dazzle/snapshot.bin, override with snapshotPath:).
• Snapshot binary format identical to Web — a snapshot saved by a Flutter Web app loads byte-for-byte on Flutter Desktop.

Fixed — iOS / Android (LLM stack)

• ToolCallParser now accepts arguments as a stringified JSON string (Qwen 0.5B fine-tune / OpenAI tool-call shape) in addition to the JSON-object shape (Gemma / Qwen 1.5B / Llama 3.x). The previous parser silently swallowed tool calls from models that emitted the OpenAI shape.
• dazzle_llama_new_context() pins n_batch = n_ubatch = n_ctx to prevent the SIGABRT inside llama_decode on prompts longer than the previous hardcoded 512-token batch. Reproduced on iPhone 12 Pro / iOS 26.3 with a 590-token prompt.

Scope (web / desktop)

• ✅ Hash KV + Vector index (HNSW) + binary snapshot persistence.
• ❌ Lists / Sets / SortedSets / Streams / standalone Strings — stay on iOS / Android mobile.
• ❌ On-device LLM clients (LlamaCppClient, LiteRtLmClient, FoundationModelsClient) — stay on iOS / Android mobile (these would need llama.cpp / LiteRT compiled to WASM, separate project).

1.0.0-beta.4

Added

• First public pre-release. Embedded in-process database with HNSW vector search and a ChatAgent runtime for on-device LLM agents.
• Five swappable LLMClient adapters:
  • LlamaCppClient — GGUF inference, Isolate worker + NativeCallable.listener for zero-copy C→Dart token streaming.
  • LiteRtLmClient — Android-only plugin bridge to LiteRT-LM.
  • FoundationModelsClient — iOS 26+ Apple Intelligence bridge.
  • OpenAICompatibleClient — pure Dart + package:http with SSE.
  • AnthropicClient — Claude 3.5/4 family via the Messages API.
• Hot-path FFI calls for HashKey.getAllDirect, SortedSetKey.rangeByScoreDirect, VectorIndex.searchDirect; method channel reserved for lifecycle only.
• ChatAgent.VectorRecallWindow performs real on-device retrieval (HNSW_SQ8) and prepends top-k semantically similar older turns to the LastN window each send().

Notes

• Same libdazzle.so / Dazzle.xcframework as the native Android / iOS SDKs — zero behaviour drift across platforms.
• LiteRT-LM and Foundation Models adapters require platform-specific setup; see README.md and samples/.