Upload examples/integration_example.py
Browse files- examples/integration_example.py +118 -0
examples/integration_example.py
ADDED
|
@@ -0,0 +1,118 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Example integration of Agent Cost Optimizer with a hypothetical agent harness."""
|
| 2 |
+
|
| 3 |
+
from aco import AgentCostOptimizer
|
| 4 |
+
from aco.config import ACOConfig
|
| 5 |
+
|
| 6 |
+
|
| 7 |
+
def example_agent_harness():
|
| 8 |
+
"""Example of how to bolt ACO onto any agent harness."""
|
| 9 |
+
|
| 10 |
+
# Initialize optimizer
|
| 11 |
+
config = ACOConfig.from_yaml("config.yaml")
|
| 12 |
+
optimizer = AgentCostOptimizer(config)
|
| 13 |
+
|
| 14 |
+
# Incoming user request
|
| 15 |
+
user_request = "Write a Python script to fetch data from an API and cache it in Redis"
|
| 16 |
+
|
| 17 |
+
# Build run state from current agent state
|
| 18 |
+
run_state = {
|
| 19 |
+
"trace_id": "agent-run-12345",
|
| 20 |
+
"current_cost": 0.0,
|
| 21 |
+
"planned_tools": [
|
| 22 |
+
("search", {"query": "redis python client"}),
|
| 23 |
+
("fetch", {"url": "https://api.example.com/docs"}),
|
| 24 |
+
("code_execution", {"code": "test script"}),
|
| 25 |
+
],
|
| 26 |
+
"previous_tool_calls": [],
|
| 27 |
+
"step_number": 1,
|
| 28 |
+
"total_steps": 3,
|
| 29 |
+
"is_irreversible": False,
|
| 30 |
+
"context_pieces": {
|
| 31 |
+
"system_rules": "You are a coding assistant.",
|
| 32 |
+
"tool_descriptions": "Available tools: search, fetch, code_execution",
|
| 33 |
+
"user_preferences": "Prefer Python 3.11+, type hints, async where possible",
|
| 34 |
+
"recent_messages": "User: Write a Python script...",
|
| 35 |
+
},
|
| 36 |
+
"retrieved_docs": [],
|
| 37 |
+
"routing_mode": "cascade",
|
| 38 |
+
}
|
| 39 |
+
|
| 40 |
+
# Call optimizer before executing
|
| 41 |
+
decision = optimizer.optimize(user_request, run_state)
|
| 42 |
+
|
| 43 |
+
print(f"Trace ID: {decision.trace_id}")
|
| 44 |
+
print(f"Selected Model: {decision.routing_decision.model_id} (tier {decision.routing_decision.tier})")
|
| 45 |
+
print(f"Estimated Cost: ${decision.estimated_cost:.4f}")
|
| 46 |
+
print(f"Estimated Latency: {decision.estimated_latency_ms:.0f}ms")
|
| 47 |
+
print(f"Confidence: {decision.confidence:.2f}")
|
| 48 |
+
print()
|
| 49 |
+
|
| 50 |
+
# Apply tool gate decisions
|
| 51 |
+
print("Tool Decisions:")
|
| 52 |
+
for td in decision.tool_decisions:
|
| 53 |
+
print(f" {td.tool_name}: {td.decision.value} (reason: {td.reasoning})")
|
| 54 |
+
|
| 55 |
+
# Apply context budget
|
| 56 |
+
if decision.context_budget:
|
| 57 |
+
print(f"\nContext Budget: {decision.context_budget.total_budget_tokens} tokens")
|
| 58 |
+
print(f" Cache prefix: {decision.context_budget.cache_prefix_tokens} tokens")
|
| 59 |
+
print(f" Dynamic suffix: {decision.context_budget.dynamic_suffix_tokens} tokens")
|
| 60 |
+
if decision.context_budget.omitted_sources:
|
| 61 |
+
print(f" Omitted: {[s.name for s in decision.context_budget.omitted_sources]}")
|
| 62 |
+
|
| 63 |
+
# Apply cache layout
|
| 64 |
+
if decision.prompt_layout:
|
| 65 |
+
print(f"\nCache Layout:")
|
| 66 |
+
print(f" Cold cost: ${decision.prompt_layout.estimated_cold_cost:.4f}")
|
| 67 |
+
print(f" Warm cost: ${decision.prompt_layout.estimated_warm_cost:.4f}")
|
| 68 |
+
print(f" Cache discount: ${decision.prompt_layout.cache_discount:.4f}")
|
| 69 |
+
|
| 70 |
+
# Check meta-tool
|
| 71 |
+
if decision.meta_tool_match:
|
| 72 |
+
print(f"\nMeta-Tool Match: {decision.meta_tool_match['meta_tool_id']}")
|
| 73 |
+
print(f" Estimated savings: ${decision.meta_tool_match['estimated_cost_savings']:.4f}")
|
| 74 |
+
|
| 75 |
+
# Check doom assessment
|
| 76 |
+
if decision.doom_assessment:
|
| 77 |
+
print(f"\nDoom Assessment: {decision.doom_assessment.action.value}")
|
| 78 |
+
print(f" Confidence: {decision.doom_assessment.confidence:.2f}")
|
| 79 |
+
print(f" Signals: {decision.doom_assessment.signals_triggered}")
|
| 80 |
+
|
| 81 |
+
# Check verifier
|
| 82 |
+
if decision.verifier_decision:
|
| 83 |
+
print(f"\nVerifier: {decision.verifier_decision.decision.value}")
|
| 84 |
+
print(f" Checks: {decision.verifier_decision.checks}")
|
| 85 |
+
print(f" Cost: ${decision.verifier_decision.estimated_verifier_cost:.4f}")
|
| 86 |
+
|
| 87 |
+
# After execution, record step and finalize
|
| 88 |
+
from aco.trace_schema import ModelCall, Outcome
|
| 89 |
+
|
| 90 |
+
model_call = ModelCall(
|
| 91 |
+
model_id=decision.routing_decision.model_id,
|
| 92 |
+
provider="cloud",
|
| 93 |
+
input_tokens=2048,
|
| 94 |
+
output_tokens=512,
|
| 95 |
+
cost_per_1k_input=0.003,
|
| 96 |
+
cost_per_1k_output=0.006,
|
| 97 |
+
)
|
| 98 |
+
|
| 99 |
+
optimizer.record_step(
|
| 100 |
+
trace_id=decision.trace_id,
|
| 101 |
+
model_call=model_call,
|
| 102 |
+
context_size_tokens=2048,
|
| 103 |
+
step_outcome=Outcome.SUCCESS,
|
| 104 |
+
)
|
| 105 |
+
|
| 106 |
+
# Finalize
|
| 107 |
+
trace = optimizer.finalize_trace(
|
| 108 |
+
trace_id=decision.trace_id,
|
| 109 |
+
outcome=Outcome.SUCCESS,
|
| 110 |
+
user_satisfaction=0.95,
|
| 111 |
+
)
|
| 112 |
+
|
| 113 |
+
print(f"\nTrace finalized. Total cost: ${trace.total_cost_computed:.4f}")
|
| 114 |
+
print(f"Cost saved vs frontier: ${trace.total_cost_saved_vs_frontier:.4f}")
|
| 115 |
+
|
| 116 |
+
|
| 117 |
+
if __name__ == "__main__":
|
| 118 |
+
example_agent_harness()
|