ServicesKnowledge assistants

Your team already has the answer. Finding it is the job

Plain language search across the documents your organisation actually runs on, with citations, permission awareness, and an honest answer when the material does not exist.

What it actually is

Search that returns an answer rather than a list of files, over content your team already wrote. It reads the documents, the prior work and the policies, and it tells you which document each sentence came from so you can check it.

01Ask in plain language instead of guessing filenames
02Every answer carries the source document and passage
03Respects who is allowed to see what, per document
04Says the material does not exist rather than inventing it
05Understands how entities relate, not just which words match

Why the off the shelf version does not hold up

01

Knowledge is spread across systems that do not talk

Templates in a document system, decisions in email, context in chat, and the rest in the head of whoever did it last time. The information exists; the retrieval path does not.

02

Keyword search fails the questions people actually ask

Nobody searches for the filename. They ask what did we agree with this client about renewal terms, and the answer is spread across three documents that never use that phrasing.

03

Permissions make naive AI dangerous internally

An assistant that indexes everything and answers everyone will eventually surface a salary band, a board paper or a matter an associate should not see. Permission has to be enforced at retrieval, not filtered afterwards.

04

Stale answers are indistinguishable from correct ones

A policy superseded eighteen months ago reads exactly as authoritative as the current one. Without freshness and versioning, the assistant confidently tells your team how things used to work.

The engineering behind it

What makes it work in production

The parts that decide whether this survives contact with real data, real volume and real edge cases.

People only get answers from documents they may open

Access control is applied at retrieval time and inherited from the source system, so two people asking the same question get answers drawn from different documents. Nothing is filtered after the fact, because by then it has already been read.

Permission aware retrieval

It understands that these two documents are about the same matter

We model the entities that matter to you, clients, matters, products, contracts, and the relationships between them, so a question about one surfaces the connected material rather than whatever shares vocabulary.

Knowledge graph

Every sentence points at its source

Answers are assembled only from retrieved passages, each linked to the document and location it came from. An unsupported claim cannot be produced, because there is nothing to attribute it to.

Citation enforced synthesis

Messy real documents, not a clean corpus

Scanned PDFs, tables, appendices, inconsistent templates and twenty years of drift. Parsing, chunking and layout handling is most of the work, and it is the part that decides whether answers are any good.

Document processing pipeline

The index tracks reality

Incremental sync as documents change, superseded versions demoted, and recency surfaced in the answer, so the assistant is not quietly quoting a policy that was replaced.

Freshness and versioning

Your documents stay yours

Data residency, retention limits, no training on your content, and encryption in transit and at rest. For regulated work we can keep retrieval and inference inside your own cloud tenancy.

AI data protection

What you end up owning

An assistant your team can query in plain language
Connectors to your document systems and wikis
A permission model mirrored from your source systems
Citations on every answer, linked to the source
A processing pipeline that handles your real documents
Retention, residency and access policy documented

Built into the tools you already run

Documents

  • SharePoint
  • NetDocuments
  • iManage
  • Google Drive
  • Dropbox
  • Box

Wikis and notes

  • Notion
  • Confluence
  • GitBook
  • Guru
  • Slite

Communication

  • Slack
  • Microsoft Teams
  • Gmail
  • Outlook

Vector and data

  • Postgres pgvector
  • Pinecone
  • Weaviate
  • Azure AI Search
  • Elastic

Start with one workflow

A 30 minute audit, no sales pitch. We map where this fits in your stack, what it would take to build, and whether it is worth doing at all.