# Advanced Failover & Hybrid Intelligence Strategy

## 1. Cascading Key & Model Failover (Chain of Best)
The system implements a multi-tier failover strategy to ensure 100% uptime:

-   **Intra-Model Rotation**: If a model hits a rate limit (429), it immediately rotates to the next available API key in the pool.
-   **Exhaustive Search**: The system checks all keys. If all are on cooldown, it identifies the one with the soonest availability.
-   **Cross-Model Cascading**: If all keys for Model A are exhausted (Daily Quota reached), it switches to **Model B** (Next Best Model) and resets the key index to 0, ensuring a fresh attempt with all keys.
-   **Reverse Search Logic**: The cyclic rotation ensures that even if specific keys are throttled, the system eventually finds an entry point.

## 2. Hybrid Intelligence Extraction (LLM + Regex)
To prevent data loss during LLM downtime, the extraction pipeline is now strictly decoupled:

-   **Regex Baseline**: Every incoming message is first processed by high-performance Regex patterns in `app/utils/extractors.py`.
-   **LLM Augmentation**: The LLM runs in a `try-except` block. It validates findings and discovers "soft" intelligence (names, context) that Regex might miss.
-   **Guaranteed Persistence**: If the LLM crashes or stalls, the `IntelligenceExtractor` catches the error and returns the Regex findings. Intelligence is never lost.
-   **Validation**: LLM-extracted data is cross-validated against Regex patterns to filter out "hallucinated" phone numbers or UPI IDs.

## 3. High-Quality Static Fallbacks
When the system enters "Survival Mode" (all APIs down), it uses high-quality templates in `PersonaEngine`:

-   **Persona Consistency**: Replies like "Main drive kar raha hoon, ruko" or "Net problem hai" maintain the deceptive persona even without AI generation.
-   **Phase-Awareness**: Fallbacks vary based on the conversation stage (Hook, Engage, Extract, Stall).
-   **Time-Awareness**: If it's late at night, the static reply includes a sleep-deprived context ("Itni raat ko? Kal baat karein?").

## 4. Verification Check
Verified with `test_hybrid_failover.py`:
✅ Simulated LLM Crash during extraction.
✅ Baseline Regex intelligence (UPI, Bank, Phone) successfully captured.
✅ System stability maintained.