Precedent & Knowledge Finder | Real Minds AI
Professional Services /Retrieval (RAG) live field guide · 10 min

Precedent & Knowledge Finder

A fee-earner asks a question in plain language and gets the answer from your own precedents, advice, policies and matter files — with the exact document, version and clause quoted, and an honest "not on file" when it isn't there.

theater/demos/profsvc_precedent-knowledge-finder.html · sandbox · read-only
Open
FIG. 1

The live demo, running on fabricated data. Open it to step through the full flow — every output is shown for a person to approve before anything happens.

How it would work

Takes a fee-earner's plain-language question, retrieves the matching passages from your own indexed precedents and policies behind your existing access controls, and shows the cited answer — or an honest "not on file" routed to a person — for the fee-earner to judge before relying on it.

Input 01
A question + your corpus

A natural-language question from a fee-earner ("how long do we keep conflict-check records?"), answered against your indexed precedent library, engagement-terms and risk policy, conflicts manual and staff handbook.

Agent 02
Retrieves, grounds, or refuses

Filters candidate documents to what the asker is entitled to see, retrieves the matching passages, and either drafts a grounded answer with a citation per passage or refuses when nothing in the corpus matches.

Output 03
A cited answer, for the fee-earner

The answer with each source quoted — document, version, section — for the fee-earner to read and decide whether the precedent applies, or an "not in your approved documents" card routed to the Risk Partner or document owner.

Where it works well

It does the retrieval mechanically and never answers without a source the reader can open and check.

  • The biggest beneficiaries are the senior solicitors and partners who currently field "where's the rule on this?" interruptions, and the juniors who lose hours not knowing the firm answered the same question two years ago.
  • Best for a firm with a real corpus worth searching — a precedent library, an engagement-terms and risk policy, a conflicts and independence manual, a staff handbook with delegations — and a steady stream of lookup questions.
  • A grounded answer that names "Conflicts & Independence Manual v4, §7.4" recaptures the retrieval load from senior staff, so their judgement goes back to the questions only they can answer.

A firm's most valuable knowledge — the precedent that fits, the advice already given to a similar client, the clause that survived the last negotiation — lives in inboxes, drives and a few senior heads. The slow, invisible cost is the hunt: the answer exists, finding it on demand is the drudgery, and the hunt interrupts the people whose time is most expensive.

Where it works badly

It faithfully surfaces a messy corpus — and a superseded clause arrives with full citations, looking as authoritative as the current one.

  • Weak where the library is thin, contradictory or out of date — version metadata helps it prefer the current document, but it cannot know a precedent is stale if nothing in your library says so.
  • It is the wrong tool for judgements rather than lookups — "can we waive the liability cap to win this tender?" is a risk-and-delegations decision, and the honest behaviour there is refusal and escalation, not a generated opinion.
  • A precedent buried in a matter folder rather than filed as a discrete item, or a controlling policy living on a partner's desktop, simply won't be found — the answer is only as complete as what is indexed.
The honest test

Pick last week's five most-interrupting questions. If three or more could be answered by quoting an existing document, this fits. If most needed a partner to weigh competing considerations, it does not — yet.

The trap is a confident answer from a stale source. Index three versions of an engagement-terms policy with none marked current, and it can quote a superseded clause with a clean citation — correct that the words exist, wrong that they still govern. It is only as reliable as the corpus is curated.

What it doesn't do — and shouldn't

It surfaces and cites. A person decides whether the precedent applies. That boundary is deliberate.

WHAT IT DOES
Quotes the matching passage with its document, version and section reference
Shows a grounding indicator — how many passages matched, across how many of your documents
Refuses when nothing in the approved corpus matches, and routes the question to a person who can authorise
WHAT IT WON’T
Decide whether a particular conflict is live, or whether a precedent applies to the matter in front of you
Give legal advice, or judge whether a costs-disclosure obligation under the Legal Profession Uniform Law has been met
Act on, send or file anything — it presents, the person judges

A firm carries professional-indemnity exposure and confidentiality duties under the Australian Solicitors' Conduct Rules. A wrong answer relied on lands on the practitioner and the firm's PI cover, not the tool — so a tool that guesses is a liability, and one that says "not on file, here's who decides" is an asset. The accountable person stays on the decision because the consequence lands on them.

What your data has to look like

The right documents, identifiable by version, with the live one distinguishable from the superseded — behind permissions the index can honour.

32%
Typical readiness
across orgs we see, before the first job
A precedent library of discrete, findable items
Needs shaping
Policies and manuals carried with a version
Usual weak point
One identifiable live version per controlling document
Needs shaping
Document-level permissions in your DMS or SharePoint
Usual weak point
A trigger to re-index when a controlling document changes
Needs shaping
The real first job

The corpus is almost always the weak point — precedents emailed rather than filed, the controlling policy on a partner's desktop, three near-identical templates with no owner. Getting the source documents to a state where the current version is identifiable and access rules are clean is usually the real first job — a matter of how knowledge is captured and governed, not of buying a tool — and it is larger and more valuable than the AI layer on top.

Right fit if…
You have a real corpus worth searching — precedent library, risk policy, conflicts manual, staff handbook
A steady stream of "where's the rule on this?" lookup questions interrupting senior staff
Your precedents and policies carry a version, with the live one identifiable
Document-level permissions in your DMS or SharePoint the index can honour
Walk away if…
The library is thin, contradictory, or full of unversioned near-duplicates
Most of what your people need is judgement under uncertainty, not "find me the clause"
Precedents live in inboxes and matter folders, not as discrete filed items
You want a tool that decides whether a precedent applies so a fee-earner doesn't have to
Open questions

The worried-buyer questions, answered straight

The safeguard is that every answer is tied to a quoted passage with its document, version and section — there is no answer without a source the reader can open and check. When nothing in the corpus matches, it refuses rather than guessing and routes the question to the Risk Partner or the document owner. It retrieves and cites; the fee-earner still decides whether the precedent actually fits the matter in front of them. A wrong call on whether a precedent applies lands on the practitioner and the firm’s professional-indemnity cover, which is exactly why the tool never has the last word.
Then it will faithfully surface the mess, including superseded clauses, because it answers from what is indexed. Version metadata helps it prefer the current document, but it cannot know a precedent is stale if nothing in your library says so. Getting the corpus to a state where the live version is identifiable — precedents filed as discrete items, policies carried with a version, one owner per controlling document — is usually the real first job, and it is the work we help with before the AI layer earns its keep.
No. It takes the routine retrieval load off senior staff — the fifth of the day spent hunting for a clause that already exists — so their judgement goes to the questions only they can answer. Deciding whether a precedent applies, whether a conflict is live under the Australian Solicitors’ Conduct Rules, or whether to waive a liability cap stays with the person. It recaptures retrieval capacity; it does not replace the thinker.
As current as the decision it informs. For a retention rule under the Australian Solicitors’ Conduct Rules or a costs-disclosure obligation under the Legal Profession Uniform Law, an answer from a superseded policy can be wrong in a way that matters. Re-indexing is triggered when a controlling document is updated, and each answer shows the version it drew from so the reader can see how fresh it is. Stale source documents are the most common cause of a confidently wrong answer.
The index sits inside your environment behind your existing DMS or SharePoint permissions, so a user only ever retrieves from documents they are already entitled to see. Client-confidential and privileged material attract duties under the Australian Solicitors’ Conduct Rules, so we scope at the start exactly which sources are indexed and which model arrangement is used, with privileged matter content kept inside your tenancy rather than sent to a public service. The data path is part of the build, not an afterthought.
You don’t take it on faith. Each answer shows a grounding indicator — how many passages matched and across how many of your documents — and quotes each source so you can open it. But it only finds what is indexed: a precedent buried in a matter folder, or a controlling policy that was never added to the corpus, simply won’t surface. The cited passages are a starting point for the fee-earner, not a guarantee the firm has nothing better on file elsewhere, which is why a person stays responsible for the judgement.
What it takes to build
3–4 weeks · 4 phases
Reused from template~70%
Bespoke to this skin~30%
stack · Claude · private vector index · DMS/SharePoint · Teams
What it would cost

Fixed scope, fixed price, fixed dates.

01
Bite-sized first piece
Index one nominated corpus, one contained change
02
Pilot build
Most builds land here
03
Embedded support
Scale on proof

Considering this for your firm?

The honest place to start is the corpus — getting your precedents and policies to where the current version is identifiable and access rules are clean. Tell us where the hunting hurts; we'll play it back, scope it, and show you what's possible.

More in Professional Services
Evidence Chronology Builder
View →
Engagement Intake & Routing Desk
View →
Scope & Billing Guardian
View →
Contract Clause Review
View →
How We Work Proof Talk to us
How We Work Proof Talk to us
Ask us anything