Job
Answer from the canonical document
Not a paraphrase of something on the internet. Retrieval points at the version you actually rely on — the signed policy, the current procedure, the approved template — and quotes it with a link back.
Eastbourne · UK
/ AI Solutions / Document chat & searchAnswers grounded in the actual document, not the internet.
Retrieval over your policies, procedures, contracts, knowledge bases and inboxes — with citations, freshness and access controls built in from the first commit.
01/ What this actually is
Search and chat that quote your own documents back.
Most teams' first document AI is a chatbot answering from the public internet, or Copilot scraping a SharePoint library that was never organised for retrieval. The work we do begins where that stops — a properly-indexed corpus with permissions, citations and a retrieval layer tuned for how your people actually ask.
02/ Jobs it does
Six shapes of document-grounded work.
Narrow, well-scoped uses where grounded retrieval earns its keep — reliably enough that people stop opening SharePoint to double-check.
Job
Search the stuff nobody can find
SharePoint sprawl, shared drives, old tender responses, matter files, ticket history — indexed properly and searchable by intent, not just keyword.
Job
Cite every claim
Every answer comes back with sources. If the retrieval didn't find anything relevant, the assistant says so — rather than improvising a plausible-sounding reply.
Job
Respect who sees what
The user sees only documents they already have access to. Permissions are enforced at retrieval time, not filtered out of the answer after the fact.
Job
Handle the inbox as a corpus
Shared inboxes, support mailboxes, thread histories — indexed the same way as documents, so 'what did we tell this client last year' becomes one question, not an hour of scrolling.
Job
Stay current without a re-index every time
Incremental sync from the source systems, with freshness indicators on results — so nobody's quoting a policy that was superseded six months ago.
03/ How we'd approach it
One corpus, tuned retrieval, citations people can click.
Document AI fails when it tries to index everything at once. We'd rather get one collection right and grow from there.
Start with the corpus that matters
One well-scoped collection — HR policy, supplier contracts, the procedure library — indexed properly, beats a giant 'everything' index that returns noise.
Tune retrieval to how people actually ask
Chunking, embeddings and re-ranking tuned on real questions from your team — not the defaults a framework ships with.
Wire permissions at retrieval time
We resolve the user's identity against the source system before ranking results. If they can't open the document, they can't see it quoted back either.
Make citations first-class
Every answer links to the chunk it came from. Users can verify in one click — which is the thing that builds trust in the system over months of use.
For the governance side — retention, permissions, prompt logging and the identity perimeter that keeps a retrieval system honest — see the AI security perimeter.
04/ Copilot vs custom
Honest about which gets you there.
Sometimes Copilot over a properly-organised SharePoint site is the whole answer. Sometimes it isn't.
If your corpus lives in SharePoint and your permissions model is clean, Copilot configured properly will often do the job. Where it stops working is when the documents sit across a shared drive, a DMS, an inbox and a third-party archive — or when the retrieval needs to blend document content with live data from a CRM or PSA. We start every engagement by saying which applies.
05/ Where it lands
Useful in the tool people already have open.
A separate search portal gets forgotten. The ones that stick sit inside the apps people are already working in.
/ Into your systems
Retrieval reaches into the document stores, DMS, CRMs and inboxes that already hold the answers. Connecting those systems cleanly is the integration work documented on the systems integration page.
/ Into your assistants
The same retrieval layer powers the internal assistants we build — so the assistant and the search tool are quoting the same canonical source, not diverging copies.
/ Backed by
Delivered by M-Tech Labs with the compliance and security discipline of M-Tech Systems — Cyber Essentials certified, aligned to NCSC CAF 4.0 and progressing through the Assurix trustmark programme. Code is continuously scanned for quality and security with Aikido, and hosted software runs on our own Nutanix / Fortinet platform — continuously pen-tested, current-version, UK-based. See secure development for the full picture.
Back to AI Solutions/ Start a conversation
Start with one corpus, one set of questions.
A short discovery call on where grounded retrieval would earn its keep first — and whether the Copilot licence you already pay for gets you most of the way before anyone writes new code.