Concept buildA worked example of how we would build this for a typical business. Not client work: no client, no client data, no achieved results.
Custom AI
Concept buildA concept build for an internal assistant that answers staff questions from a firm’s own documents, inside the firm’s cloud account. Every answer cites its sources, and retrieval is filtered by each person’s existing document permissions, so nobody gets an answer built from a file they could not open.
The brief
A typical US professional services firm keeps its know-how in documents that are hard to find. Precedents, templates, policies and past work sit across shared drives and a document management system, and finding the right one depends on knowing who wrote it. Public AI chat tools are quick, but the firm’s own policies keep client material out of them, and a generic assistant that indexes everything would show people documents behind ethical walls or matter-level restrictions. The firm needs search that answers in plain language, shows where each answer came from and respects the permissions it already has.
Explainer
Explainer · 7 slides · about 45 s
Concept buildConcept build: Private Knowledge Assistant on Your Own Cloud
Concept build, explained in 7 slides: a concept build for an internal assistant that answers staff questions from a firm’s own documents, inside the firm’s cloud account.
Slide 1 of 7
01 · Custom AI
Private Knowledge Assistant on Your Own Cloud
Concept build: Private Knowledge Assistant on Your Own Cloud.
02 · The problem
What a typical professional services firm is up against
A typical US professional services firm keeps its know-how in documents that are hard to find.
03 · The approach
How we would build it, in 5 steps
We would build it in 5 steps, from sources, permissions and the question set to evaluation gate and a staged rollout.
04 · Architecture
How the pieces connect
How the parts connect, from the first input to the person who reviews the result.
05 · Demo · sample run
Two associates ask the same question
A sample run, step by step, including where a person steps in.
06 · Typical targets
What we would measure it against
Typical targets, each with the basis it would be measured against.
07 · Next step
Scope a build like this for your business
Book a strategy call to scope a build like this.
Architecture
Architecture of the Private Knowledge Assistant on Your Own Cloud concept: Document stores → Permission sync (files + access lists); Permission sync → pgvector index (tagged chunks); Signed-in staff → Filtered retrieval (question + identity); pgvector index → Filtered retrieval (candidates); Filtered retrieval → Model in your VPC (permitted passages); Model in your VPC → Signed-in staff (cited answer); Model in your VPC → Content owner (no source found); Model in your VPC → Audit log (answer + sources).
Diagram with 8 parts: Signed-in staff (source), Document stores (source), Permission sync (processing), pgvector index (processing), Filtered retrieval (processing), Model in your VPC (model), Content owner (human), Audit log (system of record).
Connections:
Sample run
Associate asks a questionRoutine
Sample question: our usual limitation-of-liability wording for software deals.
Identity checked through SSORoutine
The signed-in user and their matter-team groups are attached to the query.
Filtered retrievalNeeds care
Six passages found; two from a restricted matter removed before ranking.
Cited answerRoutine
Answer built from four permitted passages, each linked to its source document.
Associate on the matter teamRoutine
Same question; the restricted precedent is included and cited with its matter tag.
Audit log writtenRoutine
Both questions, answers, citations and users logged in the firm’s account.
Build plan
We list the document stores in scope and how each one records access: groups, matter teams and individual exceptions. With practice leads we write an evaluation set of real questions, including some the documents cannot answer, which the assistant must decline. That set becomes the go-live gate.
Everything runs in the firm’s own cloud account: the index, the application and the model. The default is an open-weight model on a private GPU; where the firm prefers a commercial model, we use the provider’s zero-retention or enterprise data settings where offered and record the terms. Staff sign in through the firm’s existing single sign-on.
Documents are chunked and indexed in Postgres with pgvector, each chunk tagged with the access list of its source document. At question time, retrieval filters by the signed-in person’s identity before ranking, so a restricted document cannot reach the model for someone who could not open it. Permission changes re-sync on a schedule, and the delay is stated in the runbook.
The model answers only from retrieved passages and cites each one with a link to the source document. If retrieval finds nothing relevant, it says so and offers to send the question to the content owner for that area instead of answering from general knowledge. Every question, answer and citation is written to an audit log the firm controls.
The assistant opens to one practice group only after it passes the evaluation set, with refusals tested as carefully as correct answers. During the pilot, staff can flag an answer in one click, and flagged answers feed a weekly review with content owners before the next group is added.
Typical targets
Typical target
Answers drawn only from documents the asking user may open
Basis: Design property: permission filter at retrieval, tested with role-based test accounts before go-live
Typical target
Documents and the index stay in the firm’s cloud account
Basis: Design property: deployed in the firm’s account; queries leave it only if the firm picks a no-retention API
Typical target
85–95% of in-scope evaluation questions answered correctly, with a correct source
Basis: Graded on an evaluation set written from your real questions before launch. Questions your sources cannot answer must be declined or handed off, and never count as correct.
Typical target
Every answer logged with its sources for later review
Basis: Design property: question, answer, citations and user are written to an audit log the firm controls
Your project
Where each piece of data goes and which provider processes it; residency is checked per project, never assumed.
Candidate models compared on a held-out sample of your own documents or images before one is chosen.
On accounts in your name, self-hosted open-weight models, or on-premise as an Enterprise scope.
Drift checks, a review cadence and the steps to update the model or the index when your data changes.
Where this fits
FAQ
No. It is a concept build for a typical US professional services firm, and the associates, matter and precedent in the demo are fictional. The 85–95% range is a go-live bar we would set with a real firm’s practice leads on their own questions; it is not a score any deployment has achieved.
Not by design. Every indexed passage carries the access list of its source, and retrieval filters by the signed-in person before anything is ranked or sent to the model. Before launch we test this with accounts set up for different roles and matter teams, and permission changes re-sync on a stated schedule.
It shapes the design rather than ruling it out. The ABA’s Formal Opinion 512 (2024) asks lawyers who use generative AI to protect client information, understand the tool’s limits and review its output; private deployment, permission filtering, citations and audit logs are built to support those duties. The firm’s own counsel still signs off before anyone uses it.
Because it keeps every query and document inside the firm’s account, which some firms and their clients require. Open-weight models can trail the strongest commercial models on harder questions, so the evaluation set decides: if the private model misses the go-live bar, a commercial API under no-retention terms is the documented alternative.
Each practice area has a named content owner. Questions the assistant could not answer, and answers staff have flagged, are routed to that owner, who decides whether a document needs writing, updating or retiring. The index re-syncs from the source stores, so fixing the document fixes the answer.
Worked examples
Concept builds are worked examples, not client work.
Concept buildCustom AI
A concept build for a camera inspection step that sorts machined parts into pass, flagged or unsure with a detector trained on the plant’s own images. Unsure parts always go to a person, and the quality team keeps the final say on scrap and rework.
Next step
Book a free 30-minute strategy call. Bring one process you want to automate and we will talk through scope, integrations and a sensible first step.
Free 30-minute call · no commitment
