| # Merge Request: SQLite Persistence, Latency Fallbacks, Thread-Safe KV Cache Lock, and Mobile UI Enhancements |
|
|
| ## 📌 Overview |
| This Merge Request delivers critical enhancements to the **Kisan-Sathi** application to ensure thread safety, robust offline latency management, layout responsiveness, and clean system state resets. |
|
|
| --- |
|
|
| ## 🚀 Key Enhancements |
|
|
| ### 1. Latency & Timeout Fallbacks for local GGUF (`src/llm.py`) |
| To prevent the application from hanging on "Thinking..." when running local inference on slower CPU architectures (e.g., without AVX-512): |
| * **Model Loading Limit:** If the model loader (`init_llama_cpp()`) takes more than 35 seconds to load the GGUF file from disk, it automatically falls back to the mock backend. |
| * **Prefill Latency Limit:** If the initial prompt prefill (first token generation) takes more than 30 seconds, `llama_cpp` is disabled for subsequent prompts, falling back to the mock backend. |
| * **Optimized Thresholds:** Increased timeouts from 10s/15s to 30s/35s respectively, providing slow CPUs with sufficient headroom to process prompts successfully without premature timeouts. |
|
|
| ### 2. Thread-Safety Lock to Prevent KV Cache Corruption (`src/llm.py`) |
| FastAPI serves requests concurrently, which previously caused simultaneous calls to the shared in-process `llama.cpp` model: |
| * **The Issue:** Simultaneous inference requests corrupted the model's KV cache, resulting in repetitive token loops (`आपका2आपका2`) or garbled strings (`नमस्ते4आपका`). |
| * **The Fix:** Implemented a global `threading.Lock` (`_model_lock`) inside `generate_llama_cpp()`. Access to the local GGUF model is serialized: subsequent calls wait until the streaming response of the active call completes. |
|
|
| ### 3. Structured Digital Ledger Parser Fallback (`src/llm.py` & `src/ledger.py`) |
| * **The Issue:** When `llama_cpp` timed out during a ledger transaction parse query, it fell back to the mock chat backend. This returned a generic Hindi welcome message instead of valid JSON, crashing the ledger table. |
| * **The Fix:** Configured `generate_mock` to intercept structured extraction requests (checking for `"precise data extractor"` or `"json format"` system prompts) and parse fields like item, quantity, price, and type using regex rules to return a valid JSON payload. |
|
|
| ### 4. "Start from Scratch" Reset Feature (`src/db.py`, `app.py`, `assets/index.html`) |
| Added a global reset capability to clear user registrations and start clean: |
| * **Database Backend:** Implemented `db.clear_all_data()` to clear `profile`, `calendars`, `tasks`, and `ledger` tables. Deletes local JSON/CSV backup files to prevent auto-restoring data. |
| * **FastAPI Server:** Exposed a new `@app.post("/api/reset")` endpoint. |
| * **Frontend UI:** Added a localized **🔄 Reset / सब साफ करें** button inside header controls, resetting state, clearing chat views, and displaying the onboarding panel. |
|
|
| ### 5. Mobile Responsive UI CSS (`assets/index.html`) |
| * Added media queries (`@media (max-width: 600px)`) to adapt layouts for mobile views. |
| * Stacks header controls vertically, adjusts chat message widths to 95%, downsizes font/card metrics, and scales buttons/inputs for touch layouts. |
|
|
| ### 6. Proactive Nudge Markdown Parser (`assets/index.html`) |
| * Fixed rendering of raw markdown asterisks (`**`) inside the seasonal nudge banner by routing the API response through the client-side `parseMarkdown()` function. |
|
|
| ### 7. SQLite Database Migration & Cleanups |
| * Migrated local storage from JSON/CSV files to a native SQLite database (`kisan.db`), ensuring reliable offline persistence. |
| * Added support for individual transaction deletions and CSV exports. |
|
|
| --- |
|
|
| ## 🛠️ Verification & Tests |
| All 6 core component unit tests pass successfully: |
| ```bash |
| .venv\Scripts\python -m pytest |
| ``` |
| * **Output:** `6 passed in 8.85s` |
|
|
