mirror of
https://github.com/open-jarvis/OpenJarvis.git
synced 2026-07-30 19:02:16 +00:00
Jon Saad-FalconandClaude Opus 4.6
301e9cd2d4
Complete implementation across six development phases (v0.1 through v1.0): - Core: Registry system, config, event bus, types (Phase 0) - Intelligence + Inference: Model routing, Ollama/vLLM/llama.cpp/Cloud engines (Phase 1) - Memory: SQLite/FAISS/ColBERT/BM25/Hybrid backends, document ingest, context injection (Phase 2) - Agents: Simple/Orchestrator/Custom/OpenClaw agents, tool system (Phase 3) - Learning: HeuristicRouter, reward functions, GRPO stub, telemetry aggregation (Phase 4) - SDK: Jarvis class, OpenClaw protocol/transport, benchmarks, Docker deployment (Phase 5) 520 tests passing, 8 skipped (optional deps). Ruff lint clean. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
110 lines
3.8 KiB
Python
110 lines
3.8 KiB
Python
"""Tests for HeuristicRewardFunction."""
|
|
|
|
from __future__ import annotations
|
|
|
|
import pytest
|
|
|
|
from openjarvis.learning._stubs import RoutingContext
|
|
from openjarvis.learning.heuristic_reward import HeuristicRewardFunction
|
|
|
|
|
|
class TestHeuristicRewardFunction:
|
|
def test_perfect_score(self) -> None:
|
|
rf = HeuristicRewardFunction()
|
|
score = rf.compute(
|
|
RoutingContext(), "model-a", "response",
|
|
latency_seconds=0.0, cost_usd=0.0,
|
|
prompt_tokens=10, completion_tokens=10,
|
|
)
|
|
# latency=1.0, cost=1.0, efficiency=0.5 → 0.4*1 + 0.3*1 + 0.3*0.5 = 0.85
|
|
assert score == pytest.approx(0.85)
|
|
|
|
def test_worst_score(self) -> None:
|
|
rf = HeuristicRewardFunction()
|
|
score = rf.compute(
|
|
RoutingContext(), "model-a", "response",
|
|
latency_seconds=60.0, cost_usd=0.1,
|
|
prompt_tokens=100, completion_tokens=0,
|
|
)
|
|
# latency clamped to 0, cost clamped to 0, efficiency=0/100=0
|
|
assert score == pytest.approx(0.0)
|
|
|
|
def test_latency_only_weight(self) -> None:
|
|
rf = HeuristicRewardFunction(
|
|
weight_latency=1.0, weight_cost=0.0, weight_efficiency=0.0,
|
|
)
|
|
score = rf.compute(
|
|
RoutingContext(), "m", "",
|
|
latency_seconds=15.0, cost_usd=0.0,
|
|
prompt_tokens=10, completion_tokens=10,
|
|
)
|
|
# 1 - 15/30 = 0.5
|
|
assert score == pytest.approx(0.5)
|
|
|
|
def test_cost_only_weight(self) -> None:
|
|
rf = HeuristicRewardFunction(
|
|
weight_latency=0.0, weight_cost=1.0, weight_efficiency=0.0,
|
|
)
|
|
score = rf.compute(
|
|
RoutingContext(), "m", "",
|
|
latency_seconds=0.0, cost_usd=0.005,
|
|
prompt_tokens=10, completion_tokens=10,
|
|
)
|
|
# 1 - 0.005/0.01 = 0.5
|
|
assert score == pytest.approx(0.5)
|
|
|
|
def test_efficiency_only_weight(self) -> None:
|
|
rf = HeuristicRewardFunction(
|
|
weight_latency=0.0, weight_cost=0.0, weight_efficiency=1.0,
|
|
)
|
|
score = rf.compute(
|
|
RoutingContext(), "m", "",
|
|
latency_seconds=0.0, cost_usd=0.0,
|
|
prompt_tokens=25, completion_tokens=75,
|
|
)
|
|
# 75/100 = 0.75
|
|
assert score == pytest.approx(0.75)
|
|
|
|
def test_clamped_to_unit_interval(self) -> None:
|
|
rf = HeuristicRewardFunction()
|
|
# Even with extreme values, score should be in [0, 1]
|
|
score = rf.compute(
|
|
RoutingContext(), "m", "",
|
|
latency_seconds=-10.0, cost_usd=-0.1,
|
|
prompt_tokens=0, completion_tokens=100,
|
|
)
|
|
assert 0.0 <= score <= 1.0
|
|
|
|
def test_custom_max_values(self) -> None:
|
|
rf = HeuristicRewardFunction(max_latency=10.0, max_cost=1.0)
|
|
score = rf.compute(
|
|
RoutingContext(), "m", "",
|
|
latency_seconds=5.0, cost_usd=0.5,
|
|
prompt_tokens=50, completion_tokens=50,
|
|
)
|
|
# latency: 1-5/10=0.5, cost: 1-0.5/1.0=0.5, efficiency: 50/100=0.5
|
|
# 0.4*0.5 + 0.3*0.5 + 0.3*0.5 = 0.5
|
|
assert score == pytest.approx(0.5)
|
|
|
|
def test_zero_total_tokens_default(self) -> None:
|
|
rf = HeuristicRewardFunction()
|
|
score = rf.compute(
|
|
RoutingContext(), "m", "",
|
|
latency_seconds=0.0, cost_usd=0.0,
|
|
prompt_tokens=0, completion_tokens=0,
|
|
)
|
|
# latency=1, cost=1, efficiency default=0.5
|
|
# 0.4*1 + 0.3*1 + 0.3*0.5 = 0.85
|
|
assert score == pytest.approx(0.85)
|
|
|
|
def test_extra_kwargs_ignored(self) -> None:
|
|
rf = HeuristicRewardFunction()
|
|
# Should not raise even with extra kwargs
|
|
score = rf.compute(
|
|
RoutingContext(), "m", "",
|
|
latency_seconds=1.0, cost_usd=0.001,
|
|
prompt_tokens=10, completion_tokens=10,
|
|
random_extra="ignored",
|
|
)
|
|
assert 0.0 <= score <= 1.0
|