61 lines
2.8 KiB
Markdown
61 lines
2.8 KiB
Markdown
# AI Model Architecture — IT Pro Partner
|
||
|
||
**Updated:** July 31, 2026
|
||
|
||
Two separate concepts that must not be conflated:
|
||
|
||
---
|
||
|
||
## Fallback Chain *(auto-failover — config.yaml)*
|
||
|
||
Survives admin-ai outage. Uses direct API keys for F1–F4.
|
||
|
||
| Tier | Model | Provider | Notes |
|
||
|---|---|---|---|
|
||
| **Primary** | `deepseek-v4-pro` | `admin-ai` | Conductor & workhorse via LiteLLM v5 key |
|
||
| **F1** | `deepseek-v4-flash` | `deepseek` (direct) | Flash demoted -- fallback only, was hallucinating |
|
||
| **F2** | `gemini-3.6-flash` | `google` (direct) | Fast, free-tier eligible |
|
||
| **F3** | `grok-4.6` | `xai` (direct) | Verified working (NOT grok-4.5) |
|
||
| **F4** | `claude-sonnet-5` | `anthropic` (direct) | Premium survival fallback |
|
||
|
||
**Only fires when the model above it is unreachable.**
|
||
|
||
---
|
||
|
||
## Operational Models *(daily toolbox -- admin-ai only)*
|
||
|
||
All route through admin-ai. Shared $20/day budget via hermes-agent-v3 key.
|
||
|
||
| Role | Model | When |
|
||
|---|---|---|
|
||
| **Conductor** | `deepseek-v4-pro` | ALL standard work -- orchestrating, delegating, infra, coding |
|
||
| **Workhorse** | `deepseek-v4-pro` | Same model -- large context, big outputs, infra code |
|
||
| **Workhorse** | `deepseek-reasoner` | Multi-step analysis, debugging, planning |
|
||
| **Auditor** | `gemini-3.6-flash` | Sanity checks, every output (essentially free) |
|
||
| **Auditor** | `claude-sonnet-5` | Architecture, security, auth -- sparingly ($15/M output) |
|
||
|
||
---
|
||
|
||
## Admin-AI Virtual Key
|
||
|
||
- **Key Name:** `hermes-agent-v5` (regenerated Jul 31, 2026)
|
||
- **Daily Budget Cap:** $20/day shared across all operational models
|
||
- **Expiry:** Jul 31, 2027
|
||
- **Master Key:** stored in Hudu
|
||
- **Models:** deepseek-v4-pro, deepseek-v4-flash, deepseek-reasoner, gemini-3.6-flash, claude-sonnet-5, xai/grok-4.3, xai/grok-4-1-fast, grok-4.6
|
||
|
||
## Budget Targets
|
||
|
||
| Component | Daily est. |
|
||
|---|---|
|
||
| Conductor + Workhorse (ds-v4-pro) | ~$2.50 |
|
||
| Workhorse (ds-reasoner) | ~$0.70 |
|
||
| Auditors (gemini flash + claude sparingly) | ~$0.80 |
|
||
| **Total** | **~$4/day** (well under $20 cap) |
|
||
|
||
## Key Decisions
|
||
|
||
- **Jul 29, 2026:** DeepSeek V4 Flash demoted from conductor to F1 fallback. Was hallucinating tool calls and producing inconsistent results. V4 Pro promoted to primary conductor and workhorse.
|
||
- **Jul 30, 2026:** LiteLLM key regenerated — hermes-agent-v3 created in DB. However, the config.yaml was NOT updated with the new key (doc was written but config file not patched). Hermes silently fell back to DeepSeek direct since Jul 30 19:25 UTC.
|
||
- **Jul 31, 2026:** Discovered config.yaml still had stale key (`sk-DuZ..._AMg`, dead since rotation). Generated `hermes-agent-v5` with same model list + xAI voice models. Verified chat OK. Cleaned up stale keys (v3, v4, v4b). Updated model-chain doc to reflect actual state. Daily digest cron updated with HTML email + clean prompt for Telegram delivery.
|