Retrieval prompts: pack evidence and map claims

Last updated: 2 Oct 202611 min read

tutorial

A retrieval context packet should carry the request, authorized passage text, document and section identifiers, version, and a declared coverage limit. A small top-k list can omit a safety precondition that sits beside the selected chunk, so the packer should fetch necessary parent context or adjacent sections before generation. Reserve room for the actual answer and review instructions; filling the entire context window with search hits can truncate the tail or bury the decisive passage. Require each procedural claim to map to a passage that actually supports it. A valid evidence ID proves membership in the packet, not that the passage establishes the asserted fact.

Operational case

Meridian's active R-8 manual contains a reset step in section 4.2 and a lockout precondition in section 4.1. The top chunk starts at 4.2, which could produce a hazardous answer if packed alone. The context assembler adds 4.1 after confirming both sections are authorized and current. The answer names the precondition before the reset, then maps each claim to the right section ID. When the technician asks for an electrical isolation value absent from R-8, the assistant says the available packet does not establish it. It does not infer a value from a different pump or a retired revision.

Packet: R8/4.1 lockout prerequisite; R8/4.2 reset sequence.
Claim A: lockout required -> R8/4.1.
Claim B: reset sequence -> R8/4.2.
Electrical isolation value: absent -> do not invent.
Coverage: site M-12, pump P-47, active R-8 only.

Performance and operating cost

Packing P passages into a token budget is O(P) bookkeeping once token counts are known; expanding adjacent sections adds retrieval and token cost. A smaller packet lowers inference latency but raises omission risk. Test both answer correctness and precondition coverage, since an answer can cite a real passage yet omit the step that makes it safe. Claim mapping is O(C+P) with an indexed passage map for C claims. Keep references to packet IDs in a private trace so a reviewer can reproduce the exact evidence shown to the model.

Common Mistakes

Do not clip away a required precondition to meet a token limit.
Do not accept a real evidence ID as proof of an unrelated claim.
Do not fill a missing value from a different equipment model.

Connected lessons

prompt engineering

authorized retrieval