SHOULD YOUR AI AGENT ANSWER—OR ESCALATE?

Retrieval control plane for AI agents

Govern when agents search, when they answer, and what it cost.

One API between your agent and Tavily, Serper, or Exa — decides when to search, when to answer, when to escalate, and why.

>> WHO THIS IS FOR

  • LangGraph / agent eng with Tavily, Serper, or Exa in prod
  • Platform lead with a search bill and no per-customer attribution
  • Legal / compliance needing allow_answer + audit trail, not vibes
!! POLICY DECISIONS / examples
Question Decision Reason
Refund policy Answer 4 trusted sources
Legal liability Escalate Evidence threshold not met
HIPAA question Block Policy violation
Response JSON
{
  "allow_answer": false,
  "audit_id": "99c08409-28ec-4427-...",
  "policy": {
    "profile": "legal",
    "action": "block",
    "reasons": ["only 2 qualifying sources; minimum is 3"]
  },
  "overall_confidence": 0.80
}
>> INTEGRATION

Drop in before Tavily. One call from your agent loop.

integration.py
from query_fanout import RetrievalClient

client = RetrievalClient(preset="tavily", policy="legal", agent_id="support-bot")
report = await client.retrieve("How do refund policies work?")

if report.allow_answer:
    answer(report.synthesis)
else:
    escalate(report.audit_id, report.policy.reasons)
>> RETRIEVAL GOVERNANCE

Probe-first retrieval + allow_answer — the middleware layer your agent calls before search backends.

  1. 01 Should the agent search?

    Probe once. Fan out only when confidence is low.

  2. 02 Should the agent answer?

    allow_answer under legal, support, research policies.

  3. 03 Can we prove what happened?

    Sources, policy calls, blocks—each with an audit ID.

>> PIPELINE

Swipe chart → to read full pipeline

cost routing risk gate audit trail
>> BENCHMARK / tavily live

Across 125 production-style agent queries vs naive always-search on live Tavily — 125-task field study.

BLOCKED 75/125
legal policy stopped answers when evidence or authority fell short
625 > 173
72% fewer paid searches (625 → 173) — probe-first vs always-search, Tavily defaults
AUDIT OK
every decision traceable via audit_id + compliance export

Legal block rates are from a controlled web-search benchmark with default policy thresholds — illustrative of gating behavior, not a compliance certification or production SLA.

WITHOUT GOV

  • 5 searches per step
  • no allow_answer gate
  • token traces only
  • legal won't sign off

WITH FANOUT

  • 1 probe when confident
  • block / escalate + reasons
  • audit log + HTML report
  • ship with evidence
FREE · 48H · DESIGN PARTNER

FREE RETRIEVAL AUTOPSY

Send 3–7 days of agent JSONL — we return search spend, per-customer attribution, and allow_answer blocks in 48 hours. The wedge to validate the control plane on your traffic.

48h turnaround from validated receipt · limited design-partner slots

Your logs:

  • Retention: raw uploads deleted within 14 days of delivering your report
  • No resale: we do not sell or share your query data with third parties
  • Redact PII: send search queries and counts only; we flag common PII patterns on upload
  • Delete on request: email us anytime to purge your files early Full privacy policy.