Skip to content

OdooClaw 1.0.0 — Next Generation: local AI with own trained models

Latest
Latest

Choose a tag to compare

View all tags
@nicolasramos nicolasramos released this 14 Aug 13:49
· 11 commits to main since this release
v1.0.0
33ac0f5
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

OdooClaw 1.0.0 — Next Generation 🚀

100% local AI for Odoo. Your data stays on your machine. No subscriptions. No cloud.

This is the first stable release of the Next Generation: OdooClaw is now a complete
AI system with its own trained models, not just an agent that borrows any model.

✨ What's new in 1.0

🧠 Own trained models

  • OdooClaw Light 1.2B — fine-tuned on 26,968 real business examples (conversation,
    record creation, destructive-operation refusal). 95% accuracy in real conversation,
    600/600 record creations with full schemas.
  • OdooClaw Vision — invoice/document extraction straight from PDF, no cloud.
    794/800 fields correct in benchmark.
  • Models published on Hugging Face: GGUF (llama.cpp) + MLX (Apple Silicon) + Ollama.

📄 4-layer invoice OCR pipeline

Vision + deterministic fiscal layer + LLM header + arithmetic validation.
Cross-checks against Odoo reality: supplier, duplicates, totals. Never invents data.

⚖️ Dynamic billing rules (account_dynamic_rules)

Amplified for Odoo 16/17/18: tax and fiscal position decided from the OCR result,
nothing hardcoded. Configurable mechanism, zero fixed taxes in code.

🧠 Structured memory

Hot + cold: session context, business context, and long-term learning with its own
knowledge base (always-on recipe store: every successful tool execution teaches it).

🔧 134 tools · 5 MCP servers

Only the 3-5 relevant tools are injected per query: 245 tokens instead of 14,000.
Deterministic counting and hallucination rejection (NRA-556).

🔒 Enterprise security

  • ToolGuard validates every call, blocks destructive operations, denies by default
  • Agent inherits the real permissions of each user — if the user can't do it in Odoo,
    the agent can't either
  • Reply-token validation on every webhook (single-use, TTL)

🚀 Runs on any hardware

From a 2GB RAM CPU-only VPS (20 tok/s) to Apple Silicon (643 tok/s on M1 Ultra).
Less than 1GB of models. N-gram speculative decoding: +49% speed (NRA-541).

📦 One-shot installer

  • setup-local.sh: builds llama.cpp (Linux) or uses oMLX (Apple), downloads models,
    writes the gateway config
  • Doodba template v18.0.0: full Odoo + OdooClaw stack with Local AI runtime option
    (llama.cpp/MLX + models, gateway → host.docker.internal, no API keys needed)

🔄 Model-agnostic (as always)

If you already run OdooClaw with your own models, keep them. Or switch to ours.
Your choice. Nothing is hardcoded.

📊 Benchmarks

Hardware Speed
VPS 1 vCPU 20 tok/s
N100 72 tok/s
Mac Mini M1 146 tok/s
Strix Halo (Vulkan) 221 tok/s
RTX 5070 Ti 620 tok/s
M1 Ultra 643 tok/s

🧪 Tested

  • Go suite: 46/46 packages green
  • Odoo module: 23/23 tests green (Odoo 18)
  • E2E against real stack: 8/8 (login, partner, create, list, search, rejections)
  • CI gate: GitHub Actions on every branch/PR

📦 Assets

🔗 Links

⚠️ Upgrade note

  • Existing deployments: no breaking changes to configuration. The system prompt is
    now in English to align with the model's training base.
  • New deployments: run setup-local.sh (or the Doodba installer with Local AI option).
Assets 2
Loading