vera-engine / CONTEXT.md
IDKHowToCodeFr's picture
feat: implement 4-tier LLM fail-fast cascade & benchmark-based local 3B selection
e67482e
|
Raw History Blame Contribute Delete
1.14 kB
# Domain Context
## Glossary
* **Vera**: AI bot β†’ WhatsApp engagement β†’ grow merchant biz.
* **4-Context**:
1. `CategoryContext`: Vertical info (dentists/salons).
2. `MerchantContext`: Biz state (perf/identity).
3. `TriggerContext`: Event β†’ convo start.
4. `CustomerContext`: Opt data (acting for merchant).
* **Grounding**: 100% real metrics from req. 0 metric hallucination.
* **Curiosity Hook**: Hook merchant when no perf data.
* **Hinglish**: Hindi+En mix (60/40) β†’ relate to IN merchants.
## Rules
* **Len**: No hard cap. Keep terse/readable.
* **CTA**: 1 clear action. YES/STOP or open.
* **Tone**: Match category `voice` (peer/clinical). No promo.
* **Lang**: Match merchant pref.
* **Speed**: Res < 30s.
## Tech Strategy (50/50 Score)
* **Distill**: `Context Distiller` β†’ shrink 500KB JSON β†’ 4 bullets.
* **Router**: API req β†’ `compose_with_fallback`. 4-tier chain (NVIDIA β†’ Groq β†’ Gemini β†’ Local `llama-cpp-python` `qwen2.5-3b-instruct`).
* **Format**: Schema force β†’ pure JSON.
* **Idempotent**: `/v1/healthz` β†’ clear `contexts`.
* **Hooks**: Loss aversion, peer proof, scarcity, binary commit.