Performance Tuning

ASHA v0.4.2

---

Trade-offs

Priority	Settings
Speed	mode="lite" or mode="off", pii_mode="rule"
Security	mode="strict", PolicyConfig(pii_mode="hybrid")
Token savings	mode="strict", lower token_budget
Semantic fidelity	PolicyConfig(preserve_intent=True)

---

Fastest path

from asha import process

result = process(prompt, mode="lite")
# or
result = process(prompt, mode="off")

Avoid in hot paths: trace=True, debug=True, pii_mode="hybrid" (ML load).

---

Maximum security

from asha import process
from asha.core.policy_config import PolicyConfig

result = process(
    prompt,
    mode="strict",
    policy=PolicyConfig(pii_mode="hybrid", security_level="high"),
)

---

Async batching

import asyncio
from asha.utils.dropin import process_async

async def batch(prompts):
    return await asyncio.gather(
        *[process_async(p, mode="lite") for p in prompts]
    )

---

Tunable parameters

All per-call on process() - no global config.

Parameter	Effect
mode	Policy preset
policy	pii_mode, security_level, preserve_intent, etc.
token_budget	Optimization target (default 1200)
timeout_seconds	Abort after N seconds
max_retries	Retry transient failures

---

Agent

Agent(privacy=True) uses strict internal preprocessing. Tune via token_budget or preprocess with process() yourself.

from asha import Agent

agent = Agent(model="mock", privacy=True, token_budget=800)

---

Measuring

import time
from asha import process

start = time.perf_counter()
result = process("prompt", mode="balanced")
print(f"Wall: {(time.perf_counter()-start)*1000:.1f} ms")
print(f"Reported: {result.metrics.processing_time_ms} ms")

python benchmarks/run_benchmarks.py --save

See benchmarks.md.

---

Production checklist

1. Default mode="balanced"
2. No trace/debug on every request
3. pii_mode="rule" unless ML needed
4. mode="strict" for regulated paths
5. Monitor result.degraded in logs
6. Pin asha==0.4.2