AI Research Concierge for the Library — Design Study

Design study — concept only. This page describes a system I designed on paper, growing out of my real library work. It has not been built, deployed, or approved by any employer, and it is not connected to any live library system.

The Idea

Today, a researcher's request reaches the library as an email or a conversation, and a librarian does the searching, assembling, and delivering by hand — work I do every day on the ORacLe team. This concept asks a focused question: how much of the assembly could an AI assistant draft, if the library kept its standards? The answer this design gives is strict — the assistant drafts, the librarian decides. Nothing reaches a researcher unverified.

The Workflow

1 · AskThe researcher describes what they need in plain language — no search syntax, no database names

→

2 · SearchThe assistant searches the library collection — the same collection the team curates

→

3 · AssembleIt builds a reading pack: the sources, a short summary of each, and full citations

→

4 · FlagRelevant regulatory changes touching the request are flagged and cited alongside the pack

→

5 · VerifyA librarian checks the pack — sources, citations, flags — and corrects or rejects anything weak

→

6 · DeliverOnly the verified pack goes to the researcher

Scoped to the collectionThe assistant answers from the library's collection only. If the collection does not hold it, the pack says so — the request routes to a librarian the traditional way.

Feedback loopEvery correction a librarian makes during verification is logged, so the same mistake is visible — and fixable — instead of silently repeated.

The Guardrails

Human verification, always: a librarian reviews and approves every pack before delivery. The assistant never sends anything to a researcher directly.
Citations on everything: every item in a pack carries a full source citation. A claim without a source does not ship — it gets cut at verification or never drafted.
No unsourced claims: the assistant may only state what a cited source in the collection supports. Thin evidence produces a shorter pack and a note to the librarian, not a confident guess.
Regulatory flags are pointers, not advice: flagged changes are cited and labeled for the researcher's own review — the pack surfaces them, it does not interpret them for the team.
Collection-scoped: no open-web answers, no content from outside the governed collection.

How Success Would Be Measured

Turnaround time

Time from a researcher's request to a verified delivery, compared against the manual process — the assistant only earns its place if verified packs move faster without getting sloppier.

Citation accuracy

Audited samples of delivered packs, checking that every citation is real, complete, and actually supports the item it is attached to. The target is unambiguous: no fabricated or mismatched citations, ever.

Correction rate at verification

How often librarians have to fix or reject drafted packs, and why. A falling correction rate shows the drafting is improving; a stubborn one shows where the concept breaks.

Researcher trust and reuse

Whether researchers come back with their next request — and whether systematic-review-scale work, the library's highest-volume support, gets faster without losing completeness.

Why This Design

The concept is deliberately conservative, because library work runs on trust: a reading pack is only useful if a researcher can rely on every line of it. So the AI does the legwork — searching, assembling, flagging — and the librarian keeps the judgment. That is the same division of labor that makes the manual service work today, with the drafting step sped up.