Private institutions are being sold an endless sequence of better models. The more durable capability is less glamorous: finding the right private fact with its authority intact. A stronger model cannot locate a contract that was never indexed, distinguish the signed version from a draft, understand that an adviser lost authority yesterday or explain why two reports disagree. It can produce a better sentence from the wrong context. Private search is the layer that makes intelligence possible.
The hard problem is institutional truth
Private information is fragmented across inboxes, document stores, portals, advisers, messaging applications and local files. The same entity has several names. Dates conflict. Permissions reflect old roles. Decisions are recorded in prose. The valuable system must resolve:
- Identity: which person, entity, account or asset is actually being referenced?
- Authority: which record governs and who may rely on it?
- Version: what was effective at the relevant time?
- Provenance: where did the fact originate and how was it transformed?
- Permission: may this user receive the source and the inference?
- Uncertainty: what is missing, disputed or stale?
Search should return an answer path
A list of documents transfers the burden to the principal. A fluent answer hides the burden. The useful middle is an answer path: a concise conclusion, exact supporting records, conflicting evidence, effective dates and the next decision the information enables.
Value object — The Private Search Contract
Define the search system through a contract for every important domain:
- Authoritative sources and the owner of each source.
- Freshness requirement and update mechanism.
- Identity and entity-resolution rules.
- Permission inheritance for sources, summaries and indexes.
- Citation requirements for factual answers.
- Treatment of contradiction and insufficient evidence.
- Retention, logging and deletion for queries and outputs.
The contract makes search quality testable without tying the institution to one model.
Index less, better
Do not ingest every private record because storage is cheap. Broad indexing increases exposure, noise and legal complexity. Start with domains where retrieval creates clear decision value: current mandates, entity records, approved policies, transaction materials and authoritative reporting. Exclude personal and restricted material until purpose and permission are designed.
Test questions that matter
Benchmark with real institutional questions, not generic knowledge tasks:
- Which document currently authorises this person to act?
- What changed in the exposure since the last board decision?
- Which source supports this ownership figure?
- What deadline depends on an adviser who no longer has access?
- Which two reports disagree, and why?
Score source correctness, permission, provenance and explicit uncertainty—not just whether the answer sounds useful.
Models become replaceable
When retrieval, provenance and permissions are independent, the institution can compare models and change them without rebuilding its memory. That is strategic leverage. The private knowledge layer compounds while model capability becomes a selectable component.
The system should know when it does not know
A serious search system returns “not established” and shows the gap. It does not merge stale drafts into confident prose. Build the ability to find and prove before buying another ability to write. Private search is where institutional intelligence stops being a demonstration and becomes a capability.