← Self-Healing Multi-Agent Systems
Response Repair ROI Calculator
Every failed validation that triggers a re-ask costs you another full inference. Response Repair fixes structural errors — wrong types, malformed JSON, bad enum casing — deterministically, in microseconds, before validation ever runs. This calculator estimates how much that saves at your pipeline volume.
Your pipeline parameters
Adjust to match your actual or estimated volumes. Defaults are calibrated for a mid-size company running heavy AI data pipelines on GPT-4 class models.
Total LLM calls per month across all agents and pipelines.
Average cost per single LLM response in US cents — for GPT-4o with 1,500 output tokens this is roughly 2–8¢.
Percentage of model responses that currently fail schema, type, or structural validation.
Of failing responses, what percentage currently trigger a full re-ask (another inference)?
Estimated share of failures that are structural — wrong types, bad JSON syntax, wrong enum casing, missing defaults — and can be auto-corrected without a retry.
Formula: Monthly savings = requests × failure rate × healable % × re-ask rate × cost per request.
Response Repair runs before validation and fixes structural issues deterministically. It does not call the model again — repair is microseconds, not another inference.
Estimated monthly savings
Monthly inference cost saved
$341.25
6,825 re-asks eliminated
Annual savings
$4,095
At current volume and cost
Additional gains
Latency hours saved / month
5.7 h
Assuming 3s per avoided re-ask
Pass-rate gain
+1.36 pp
Percentage-point lift in requests passing validation
Failures auto-healed / month
9,750
of 15,000 total validation failures
Note on scope
This calculator covers direct inference cost from eliminated re-asks. It does not model downstream savings from fewer SLA violations, reduced human review, or token overhead inside re-ask prompts — all of which typically add 20–40% to the real total.
Enable Response Repair in Agent Settings →What Response Repair actually does
Tier 1 — JSON shape fixes
Trims whitespace, removes trailing commas, fixes unquoted keys, normalizes smart quotes, and extracts the JSON body from markdown fences.
Tier 2 — Semantic normalization
Casts numeric strings to numbers, normalises boolean tokens, fills null or missing fields with configured defaults, and normalises enum casing.
Runs before validation
The repaired output hits your validators. If it passes, no re-ask needed. If it still fails, your normal block or re-ask logic fires — nothing is bypassed.
Configurable per-agent
Off by default. Set to Safe (Tier 1 only), Standard (Tier 1 + 2), or Custom to choose exactly which rules apply.
Auditable
Every repair is logged: which rule fired, which field was changed, before and after values. Check X-VG-Healed-Trace on the response header.
Conservative by design
Repair defaults are never injected into fields with existing values. Object and array defaults require explicit opt-in in the schema.