baar-core 0.1.4

Budget-Aware Agentic Routing — route LLM calls intelligently between cheap and powerful models with a hard budget cap.

baar-core (BAAR-Algo)

Route LLM calls intelligently between cheap and powerful models — with a hard financial kill-switch that never breaks.

---

🚀 Why BAAR?

Every agent developer using GPT-4o has seen this:

• Simple task → sent to GPT-4o anyway → 15× more expensive than necessary.
• Budget set to $0.10 → agent burns $0.40 → surprise invoice.
• No visibility into which agent step cost what, or why.

BAAR (Budget-Aware Agentic Routing) solves this at the protocol level.

---

🧠 How it Works

BAAR acts as a semantic gateway between your application and the LLM providers.

graph TD
    A[Your Task] --> B{Semantic Router}
    B -- Complexity < 0.65 --> C[gpt-4o-mini]
    B -- Complexity >= 0.65 --> D[Budget Kill-Switch]
    
    D -- "Affordable?" --> E[gpt-4o]
    D -- "Too Expensive" --> F[Force Downgrade to Mini]
    
    E --> G[Audit & Spend Tracking]
    F --> G
    C --> G
    
    G --> H[Final Response]

1. Semantic Scoring: Uses a cheap model to score task complexity (0.0–1.0).
2. BCD (Budget-Constrained Decoding): If the powerful model is too expensive for your remaining budget, BAAR automatically downgrades to a cheaper one to ensure the task completes without an overage.
3. Local Rejection: If even the cheapest model exceeds the budget, the request is rejected locally with zero network cost.

---

🔬 Benchmarking Results

To ensure frontier-grade quality, BAAR-Algo is validated on industry-standard datasets.

Dataset	Strategy	Accuracy %	Cost (USD)	Savings vs BIG
MMLU	ALWAYS-BIG	100.0%	$0.0905	-
(Knowledge)	BAAR-Algo	70.0%	$0.0050	93.3%
GSM8K	ALWAYS-BIG	100.0%	$0.0905	-
(Math)	BAAR-Algo	80.0%	$0.0050	93.3%
HumanEval	ALWAYS-BIG	100.0%	$0.0105	-
(Coding)	BAAR-Algo	100.0%	$0.0105	0.0%*

*On HumanEval, BAAR correctly detects 100% complexity and uses the Big model, ensuring zero quality loss for critical code.

Run the Benchmark Yourself (Free)

baar-bench --dataset all --mock

---

📦 Installation

pip install baar-core

⚡ Quick Start

from baar import BAARRouter

# Set a hard $0.10 budget cap
router = BAARRouter(budget=0.10)

# This will be routed to gpt-4o-mini (Complexity ~0.1)
response = router.chat("What is the capital of France?")

# This will be routed to gpt-4o (Complexity ~0.9)
code = router.chat("Write a complex matrix multiplication in CUDA.")

---

🛡️ Resilience & Security

BAAR is designed for Financial Safety (Anti-Denial of Wallet).

Attack Vector	BAAR Response	Proof
Unbounded Consumption	Zero-Call Rejection	Blocks request locally with Zero network calls.
Complexity Inflation	Semantic Scoring	Ignores gibberish/padding intended to drain budget.
Sensitivity Toggling	Tunable Threshold	Adjust complexity_threshold to match your quality needs.

Verify resilience locally:

baar-stress

---

🛠️ Configuration

router = BAARRouter(
    budget=0.10,                    # Hard cap in USD
    small_model="gpt-4o-mini",      # Cheap model
    big_model="gpt-4o",             # Powerful model
    complexity_threshold=0.65,      # 0–1: above this → use big model
)

---

📄 License & Research

Distributed under the MIT License. See LICENSE for more information.

For architectural details and mapping to the OWASP LLM10 security framework, see RESEARCH.md.