⚖️ Janadhikar — जनाधिकार

Your rights. In your own words. On your phone.

An on-device AI legal assistant for Indian citizens — ask about your rights by typing or speaking (English · Hindi · Hinglish) and get a clear, grounded answer with the exact law behind it.

📲 Download Janadhikar v1.0.0 (APK)

_{Install → open once on WiFi to fetch the AI model → use it fully offline.}

🔒 Private by design Every question, transcript & answer stays on the device. Network is used once — to download the model.

⚡ Fast & offline Sub-second retrieval from a local law database; the AI runs on-device after setup.

📜 Grounded in real law Every answer links to the exact bare-act text + the official PDF at the right page.

🏠 README · 📐 Architecture · 🤝 Contributing · 📚 Knowledge · 🧠 Cognee

---

✨ What it does

flowchart LR
    A["🙋 'article 15 kya hain'<br/>'police slapped me'<br/>🎙 or typed"] --> B["Janadhikar"]
    B --> C["📄 Clear answer<br/>quote → meaning → why it matters"]
    B --> D["🔗 Exact law + official PDF"]
    B --> E["🌐 English · Hindi · Hinglish"]

Loading

• Ask anything, any way — type or speak; English, Hindi, or Hinglish (article 15 kya hain → answered in Hindi).
• Rich, ChatGPT-style answers — the exact wording quoted, what it means in simple words, how it's interpreted, and why it matters — rendered as decorative Markdown (bold, blockquotes, highlights).
• Always shows the real law — a collapsible Sources card with the verbatim section, page, and a "📄 Read the official law" button that opens the government PDF in-app at the exact page.
• Smart routing — article 21, section 302, fundamental rights, how many articles, or explain this simpler are each handled the right way.
• 📖 Constitution overview — dates, drafters (Dr. B. R. Ambedkar), the original-vs-today structure table, the three pillars, borrowed sources.
• 🕘 History, ✎ new chat, select-a-word → meaning, copy, and a model picker (Fast / Balanced / Gemma).
• Never fails open — if the model errors or times out, you still get the exact verified law, never a hallucination.

---

🚀 Try it

Option A — install the APK (for users / testers)

1. Download janadhikar-v1.0.0.apk and install it (allow install from unknown sources).
2. Open it once on WiFi — it downloads the ~350 MB AI model with a progress bar.
3. Done — it now works fully offline. Type or tap the mic and ask.

First run needs internet + ~1 GB free storage. After that, no network is used. Requires Android 8.0+ (arm64). The APK is signed with the Janadhikar release key.

Verify the download (optional): shasum -a 256 janadhikar-v1.0.0.apk → 526b801b994b65ee788428ab220eb165151e6219c0fcc6ef3e37f3b9f084a82b

Option B — build from source (for developers)

git clone --recurse-submodules git@github.com:FiscalMindset/JanAdhikar.git
cd JanAdhikar
./gradlew :app:assembleDebug            # builds native (llama.cpp + whisper.cpp + sqlite-vec)
adb install app/build/outputs/apk/debug/app-debug.apk

# Optional: push models over USB instead of downloading in-app
HF_TOKEN=hf_xxx ./scripts/fetch_models.sh
./scripts/push_models.sh

A signed release build (./gradlew :app:assembleRelease) needs a keystore.properties at the repo root (gitignored) — see the release notes.

---

🤖 The models (pick one in Settings)

Model	Runtime	Download	Speed	Best for
Balanced — Qwen 2.5 1.5B (default)	llama.cpp (CPU, Q4_0)	auto, ~1 GB	Moderate	Best accuracy & Hindi, richest answers
⚡ Fast — Qwen 2.5 0.5B	llama.cpp (CPU, Q4_0)	auto, ~350 MB	Quick	Speed on budget phones (lower accuracy)
Gemma 3 1B / 4B	MediaPipe (accelerated)	manual .task	Fast	Devices without CPU dot-product

The 1.5B is the default because it follows the retrieved bare-act text faithfully — the 0.5B is faster per token but drifts and can mix up provisions, which is unacceptable for legal answers. Speed comes from Q4_0 repacking + Flash Attention + ARM dot-product kernels (≈ 4× faster than a naïve build) plus live token streaming — the answer is written out word-by-word as it generates, so it feels responsive instead of a long silent wait. On CPUs without dot-product the app safely falls back to Gemma / verbatim.

---

🧱 Tech stack

Area	Technology
UI	Kotlin · Jetpack Compose · Material 3
LLM	llama.cpp (Qwen 2.5 GGUF) · MediaPipe GenAI (Gemma) — via JNI
Voice (STT)	whisper.cpp (small, multilingual)
Retrieval	SQLite + sqlite-vec (vector KNN) + FTS5 (keyword), fused in one native lib
Embedder	paraphrase-multilingual-MiniLM-L12-v2 → TFLite (LiteRT), pure-Kotlin SentencePiece
Knowledge	Offline Python pipeline over official India Code / Constitution PDFs

See ARCHITECTURE.md for the full diagrams (startup, routing, retrieval, model selection, voice, native build, distribution).

---

📚 The law inside

5 statutes · ~1,742 provisions — the Constitution of India (466 articles), Bharatiya Nyaya Sanhita, Bharatiya Nagarik Suraksha Sanhita, Bharatiya Sakshya Adhiniyam, and the Motor Vehicles Act, 1988 — all with verbatim English + Hindi text, page numbers, and official source links. See knowledge_database.md.

---

🔐 Privacy & honesty

• Your questions never leave the phone. The only network call is the one-time model download from Hugging Face.
• Grounded, not fabricated. Answers are built from the retrieved law; the exact verified text is always one tap away; if the model produces garbage, the app shows the raw law instead.
• Auditable. The manifest, the prompt builder (PromptContract), and the output guard (OutputSanitizer) are small and reviewable.

---

_{Built to put justice information in every hand. ⚖️}