Local LLM phi-2

[SPatil2026]

This PR transitions the form-filling engine from legacy heuristics to a Context-Aware Smart Batch Extraction system powered by the Local LLM. It significantly enhances accuracy, enables natural language corrections, and guarantees robust performance across various environments.

Key Changes

1. Smart Batch Extraction
   Context-Aware: Replaces regex extraction with extract_all_fields, which maps user speech to the complete form schema (Fields, Types, Options) in a single pass.
   Multi-Field Support: Users can now fill multiple fields simultaneously (e.g., "My name is Sam, and my email is sam@test.com").
   Pollution Control: Implements strict parser logic to discard hallucinated fields or extraneous text.

2. Robust Local LLM Service
   Smart Loading Strategy: Uses auto-fallbacks for hardware compatibility: 4-bit (bitsandbytes) → FP16 (GPU) → CPU.
   Setup Scripts: Fixes download_models.py with dependency checks and adds setup_env.bat for easier Windows setup.

3. Data Refinement & UX
   Smart Updates: Users can now naturally correct existing values (e.g., changing a phone number); the Agent refers to the master schema instead of just "remaining" fields.
   Strict Formatting: ValueRefiner enforces strict rules (e.g., removing spaces from Phone, lowercasing Email) on all inputs.
   Option Matching: Improves dropdown logic to support text labels and fuzzy matching (e.g., auto-correcting "Alumina" to "Alumni").