Intermediate~12 minHazelJS

Retrieval Before Action

Require grounded lookup/RAG before write tools fire—via policy, graph edges, or tool allow-order discipline.

Authors: editorial-team
Published: Published 2026-08-13
Last reviewed: Reviewed 2026-08-13

Problem

Agents act on stale assumptions without checking systems of record.

Context

Ops and support agents that must read tickets/docs before mutating.

Forces and constraints

Writes without reads cause incidents
Forced retrieval adds latency
RAG alone may still be stale vs transactional reads

Recommended design

Structure graphs/pipelines so lookup tools precede writes; PolicyEngine or evals fail if write occurs without prior read tool in the trace; enableRAG for doc grounding but prefer transactional @Tool lookups for order/ticket state.

Minimal pseudocode

// Eval gate
expectTools(result, ['lookupOrder', 'processRefund']);
// Policy: deny processRefund if no prior lookup in run (app-level guard)
// Graph: lookup node → decide node → write node

Failure modes

RAG-only grounding for transactional truth

Security considerations

Read tools still need authz

Observability signals

Flag writes without preceding reads

Evaluation approach

Forbidden write-without-lookup cases in describeAgent

Trade-offs

Extra read cost vs safer writes

Optional · Build with HazelJS

HazelJS how-to

typescript

// 1) Graph/pipeline forces lookup before write
// 2) Eval gate:
expectTools(result, ['lookupOrder', 'processRefund']);

// 3) RAG for docs — not transactional truth
await runtime.execute('desk', goal, { enableRAG: true, ragTopK: 5 });

Docs: RAG, Testing.