THE INSPECT & CONTROL DATABASE FOR AI
Control at
the retrieval layer.
No need to clean
No excessive control at the agentic layer
Better answers
Fewer tokens
That simple.
Aspected is a patented database you add to the stack you already run — inspect & control on your internal data, one layer down, in retrieval itself. Inspect it. Control it. Trust it. Instantly.
30-day free trial — the full database, up to 5 TB, in your own Azure or AWS tenant.
THE MOVE, IN ONE PICTURE
// before: control only in the agentic layer — after the fact, paid per query.
// why that matters (01) · the invention that made the move possible (02) · watch it live (03).
ASSUMED INDUSTRY WISDOM
“Clean or pay.”
For once, the entire industry agrees. Before you let AI near your content, clean it up — or pay for what happens next. Don’t take our word for it; take theirs:
SERVICENOW
“Content is the ultimate driver” of Now Assist’s answers — its own support guidance; readiness means auditing and retiring the knowledge base first.
MICROSOFT COPILOT
“Clean out redundant, outdated, and trivial (ROT) content… perform an extensive audit of all organizational content” — Microsoft’s own guidance for preparing your data for Copilot.
OPENTEXT
The prerequisite is sold as a service: an AI & Data Readiness workshop to clean the content before Aviator can answer from it.
SALESFORCE
Its own research: 26% of organizational data is untrustworthy, and only 7% of enterprises call their data AI-ready — fix the data before Agentforce.
SAP
AI success is concentrated in organizations that solved data quality before AI — everyone else waits.
ATLASSIAN
“Rovo works best when your Confluence site is full of detailed, complete, and up-to-date content” — Atlassian’s own best-practices documentation.
All of that wisdom exists because of one flaw: the cutoff. Today’s retrieval has to cut — a hard filter deletes whatever doesn’t match; a shortlist keeps only the top of one blended score. Your content has to make the cutoff, and in an estate full of drafts, duplicates and outdated versions, the chance that the right document survives it is low. So the industry compensates twice. You clean — so the cutoff happens to cut in the right place. And you patch — hard-coded filters, rerankers, GraphRAG builds, re-query loops — a fortune in tokens spent forcing the right answer to appear.
For AI on internal, unstructured content — the files, the mail, the documents — every solution on the market today is built on that single problem: the cutoff.
TODAY
cut off before scoring
ASPECTED
graded together — no cutoff
THE PATENTED ALTERNATIVE
Instead of metadata as a filter, Aspected uses metadata as a signal. Every signal — type, year, region, classification — is graded alongside meaning, in one pass. An aspect is a dial, not a gate.
No cutoff. Nothing is deleted before scoring — a wrong guess in the question can’t destroy the right answer.
ROT never pollutes the shortlist. Redundant, obsolete and trivial content demotes itself — on recency, on status, on every signal it fails.
Right at the first pass. The straight path: fewer checks above, fewer tokens spent.
Control in the retrieval layer makes cleaning a choice.
THE FREE TRIAL
Try the full database. Free, for 30 days.
The real product — Debug and Control included, full dials, your own internal content — deployed in your own Azure or AWS tenant, up to Database M: 5 TB. Debug your ten worst queries against your current retrieval, side by side, and see why every answer ranked. From September 7, start it straight from AWS Marketplace, in your own account — no sales call required.
01 · WHERE CONTROL LIVES TODAY
AI control happens one layer too late
Every enterprise AI stack runs the same chain — and hallucination control lives in the agentic layer, because retrieval, as it exists today, cannot be controlled. So every answer is checked after the fact; every failed check means another retrieval, another round, more tokens. If your AI feels slow, it isn’t broken — it’s checking. And you pay for every check.
Control in the agentic layer requires clean content in the retrieval layer. Control in the retrieval layer makes cleaning a choice.
THE MOVE
Control at the source — not after the fact
Aspected adds inspect & control where it wasn’t possible before: in retrieval itself. The right context, verified, the first time — so the answer is accurate at the first attempt, in one go, and the agentic layer keeps its own control with far less to do: fewer checks, fewer retries, dramatically fewer tokens. Add control one layer down, and everything above it gets faster and cheaper.
You saw the picture at the top of the page. The next section shows why adding control there needed an invention — and what it is.
02 · WHY THIS KEEPS HAPPENING
Two flaws. One invention.
Why couldn’t retrieval be controlled until now? Two structural flaws — the agentic layer’s checking exists to compensate for both. And both are removed by one patented invention.
FLAW 1
Two search methods fight each other
Vector search ranks by meaning; metadata filters cut by rules. Bolted together, they fight — lots of back and forth between the two, extra compute at every round, and often still the wrong answer.
FLAW 2
Observable — not inspectable
You can see it scored 0.96 on the vector. You cannot inspect why — there is nothing underneath the number to open, and nothing inside it to steer.
THE INVENTION
Turn metadata filters into aspects of a vector
Aspected is a multi-aspect query index: every metadata signal — type, year, region, classification — becomes an aspect of the query itself, graded alongside meaning. One pass that includes all the signals — instead of several steps that fight.
And because every aspect is scored separately, every answer can show why it scored — per signal, on demand. Control moves down a layer: accurate at the first attempt, in one go.
Dashed = what the question asked.
Solid = what this record offers.
The distance between the two shapes is the explanation.
US Patent · granted Jan 2026 · PCT filed · independent TNO evaluation underway
03 · SEE IT LIVE
One question, watched all the way through
// a realistic enterprise ask — five signals in ten words. Four states, from uncontrolled to controlled — the full anatomy is in the worked example (05).
Retrieval on meaning alone, like your AI today. Top result: an email thread — office gossip, stated as fact, with a source attached. The right document is buried at #4. And for the first time, you can see that: the ranking is open, per signal.
Document type becomes an aspect. Reports rise, the email sinks — but the French report wins. Differently wrong: one extra signal is not enough. Inspection shows exactly which signal is still missing.
Region joins — the right document takes #1, and you can verify why. And look at the year: the user asked 2023, but the truth lives in the 2022 report — the question itself was wrong. A hard filter would have obeyed the mistake and deleted the right answer; an aspect is a dial, not a gate. Control first — the better answer is the consequence.
The fundamental step: control, made permanent. The dial settings that produced the right, verified answer become the standing retrieval policy — every question from now on starts controlled. IT owns the policy, users and agents steer within it, an auditor can read it. Retrieval you can control, not just observe.
Now drive it yourself — real query, real data, real console
A live support use case on real ticket data: pick a question at the bottom left and compare a conventional vector database with Aspected, side by side — runtime console open, so you can watch both retrievals happen. The similar-sounding answer against the correct one; they’re often not the same document.
04 · WHAT IT GIVES YOU
Three advantages. Four use cases.
Metadata and meaning, fused into a single ranking pass — per signal, weight-controlled, patented. It’s the invention that adds control at the retrieval layer — everything below follows from that addition: three advantages it wins you, and four use cases it opens. None of it replaces what you run; all of it is added.
THE KILLER FEATURE
Inspectable, not just observable
Every retrieval stack can be observed — latency, recall, eval scores. Only Aspected can be inspected: open any answer and see why each document ranked, per signal. Observation tells you that it failed; inspection shows you where — and hands you the dial to fix it. This is the capability nothing else in the stack has — you just watched it in the three states above — and everything below follows from it.
Every end user asks the same question today: “Can I trust this answer?” And soon they’ll ask a second: “How do I control my token burn?” Both answers come from the same place — control at the retrieval layer. Three advantages: debug & control, costs, accuracy.
DEBUG & CONTROL
The core — possible nowhere else
Debug: every answer opens up per signal — see why it ranked, side by side against what you run today. Control: turn the dials and save what wins as the standing retrieval policy — IT owns it, users and agents steer within it, an auditor can read it. Debugging and controlling retrieval output is Aspected’s core — a blended vector score has nothing inside it to open.
ACCURACY
More signals, sharper answers
A question is never just meaning — it also says which year, which region, which type. Today’s retrieval ranks on meaning alone and filters whatever is left; Aspected ranks on all of it, in one ranking. Every signal you add removes a way to be wrong. And every answer arrives checkable — source and per-signal reasoning one click away, so the link is right, and visibly right.
COST
The straight path
Your AI token bill is a retrieval problem. Control at retrieval means the agentic layer has less to do: fewer checks, fewer retries. The patched pipeline pays three times: an oversized shortlist, a reranker model, and re-query loops. The straight path pays once — one query, one pass, the right records first time. Fewer tokens, fewer GPU cycles, and less irrelevant context stuffed into the model.
THE USE CASES
CONTROL
Hallucination control — at the source, as a dial
The control the agentic layer had to improvise, now native to retrieval. Weights are dials, not code. Put authority above similarity so the approved policy outranks seventeen near-identical drafts; constrain by classification; audit every answer’s grounding. Wrong document in means hallucination out — controlling retrieval is the most direct control over hallucinations a stack has.
NO CLEANUP
The everyday questions, on content as it is
“Latest HR policy?” works on day one, on the estate exactly as it stands: stale content demotes itself on recency, drafts lose to approved versions, the wrong region falls away. The “clean up your content first” project that structurally never finishes — no longer needed. Your estate is AI-ready as it stands.
ONE KNOWLEDGE LAYER
Across where your knowledge resides
Your knowledge lives in Microsoft 365 and the DMS, the file shares, the archive nobody dares to touch. Aspected is one layer across all of it: 60+ connectors, one index, unified aspects across the full estate — the same dials, the same inspectable ranking everywhere. Connectors delivered through Xillio as implementation partner. Internal data only: Aspected answers from the sources you connect — never from the public web — with your permissions applied on the index.
SOVEREIGN & INDEPENDENT
Copilot today. Tomorrow, wherever you go next.
MCP-native: Copilot routes to it in a Microsoft tenant; a Bedrock Agent reaches it exactly the same way. The knowledge layer is independent of the assistant on top of it — your retrieval policy, your dials, your index stay yours. Cloud or fully on-premises, in your VPC: sovereignty is a deployment choice.
Partners build and sell specific solutions on the Aspected Database — packaged, ready to deploy, running on your own sovereign database. And always AI-ready, without the cleanup: your content goes in as it stands.
BY XILLIO
AI-Ready Migration
Migrate network drives and legacy systems to SharePoint — and land AI-ready the day the migration completes. No cleanup project first: content moves as it stands.
BY XILLIO
Knowledge Layer
One knowledge layer across everywhere your knowledge resides — connectors, aspects and permissions delivered as a service. Nothing cleaned first: sources connect as they are.
BY XILLIO
AI-Ready SharePoint
Your existing SharePoint estate, made AI-ready as it stands — indexed, aspect-mapped, permission-aware. No migration, no cleanup.
05 · GO DEEPER - THE WORKED EXAMPLE
Every word is a signal
Read the question again — slowly. It isn’t one request. It’s five, stacked into ten words:
German office, 2023”
…and every file in your estate already carries the matching metadata — filename, created, modified, extension, author, department. Maintained by nobody. Accurate by default. Free.
What today’s retrieval does with those five signals
It compresses all of them into one similarity number — an “overall impression” of meaning. The other four signals are invisible, or bolted on afterwards as filters. That is not a flaw in your vector database; it is simply the nature of dense vectors. But it is why the confident wrong answer exists: similarity is not correctness.
What you’re looking at: one record, scored the Aspected way. Each axis is an aspect — one of the question’s signals. Dashed is what you asked; solid is what this record offers. Why it matters: the distance between the two shapes is the explanation. A ranking you can read — and, as you’ll see, dials you can turn.
06 · GO DEEPER - THE SAME RECORD, THREE VERDICTS
The same record. Three verdicts.
In our demo estate, no record matches all five signals — the realistic case. The best compromise is the DE office incident report (2022): right region, right type, right classification — one year off. Here is its fate under three retrieval methods:
The whole question, squeezed into one number. The report is shortlisted “on overall impression” — but an informal email thread outranks it. One blended score can’t weigh trade-offs. It picks by vibe.
RERANK
The industry’s fix. The hard filter on year = 2023 is binary: the best compromise is thrown away before the reranker ever sees it. Bolt-on filters can’t trade off — they can only kill. And every patch adds inference, latency and tokens.
Every signal graded, in one pass. The one-year miss is traded against perfect region, type and classification — ranked #1, the reason visible per signal. One query. No filter, no rerank, no loop.
What you’re looking at: the live demo’s own verdict panel for that record. Why it matters: the sophisticated pipeline performed worse than naive vector search — the structural cost of metadata bolted on next to a ranking instead of built into it. Open the live console and run it yourself →
07 · WHICH ONE ARE YOU?
Whoever sent you here — this is what they meant
You probably arrived with a sentence in your head. Find yours:
COPILOT & M365
Running Copilot on SharePoint?
“They add inspect & control to Copilot — no cleanup project first.”
Correct. Aspected deploys as an MCP server in your tenant; the Copilot agent routes questions to it from the first query — and answers accurately at the first attempt, burning far fewer tokens.
The Aspected Database →
MSPS & RESELLERS
Selling into managed M365 tenants?
“There’s a recurring product ladder on content you already manage.”
Correct. The Aspected Database — capacity-priced infrastructure running in the client’s own tenant, with margin for the partner and a demo that closes itself: show the customer why an answer ranked.
PLATFORMS & CLOUD
AWS, Nutanix, IONOS — sovereign by choice?
“It runs entirely inside your infrastructure.”
Correct. One container, MCP-native — a Bedrock Agent reaches it exactly the way Copilot does — in your VPC or fully on-premises. TLS, bearer tokens, ACL enforcement first-class. Sovereignty isn’t a special mode; it’s a deployment choice.
DEVELOPERS & ISVS
Building RAG, agents, or a product?
“It’s a database you add — and you can finally debug retrieval.”
Correct. Start the 30-day free trial — the full database in your own account, up to 5 TB, with a 30-minute quickstart. The prep pipeline handles the chunking, or push your own records through the ingestion API. LangChain & LlamaIndex retrievers arrive in September. docker pull aspected/engine
— OEM terms for embedding.
08 · HOW YOU ADOPT IT
Add a database. Don’t rebuild a stack.
Everything above might sound like an argument to rip out your vector database. It’s the opposite — nothing is migrated, nothing is decommissioned. You add one container, compare on your own content, and route only what wins:
And no cleanup project first: recency and freshness are aspects — ranked, not scrubbed by hand. Point Aspected at your content exactly as it is today. Clean up to save storage and reduce risk; not to make AI work.
09 · FROM THE FIELD
Don’t take our word for it
“Aspected has added the context Microsoft Copilot needed. The answers are now perfectly aligned with how we organize and use our information.”
“You guys are purely unique in this space… there’s your ROI, isn’t it?”
“Aspected stood up in 30 minutes and gave us something we didn't have: a way to actually see retrieval quality side by side instead of guessing, which immediately exposed a gap in our own baseline. What stands out isn't a single "it's better" number; it's that the aspect model gives you, and an agent, real levers to shape relevance per query that a plain vector store simply can't.”
Andy Copland · Interim CTO, Vertex Agility (CGA Management)
"Aspected includes all context directly in the search query upfront. This enables immediate governance, cuts search times in half, and delivers maximum accuracy at lower cost.”
Wim Verheij · Managing Director · Novadoc Netherlands
● US Patent · granted Jan 2026 · PCT filed ● Independent TNO evaluation underway
● A Xillio spin off · 22 years of content expertise behind it
10 · QUESTIONS, ANSWERED STRAIGHT
FAQ
No. Aspected works on your internal data — the sources you connect: Microsoft 365, the DMS, file shares, archives. Nothing is fetched from the public web, your permissions are applied on the index itself, and the whole layer runs where you choose: cloud, your VPC, or fully on-premises. Every answer is grounded in your own documents — source linked, reasoning open per signal.
Because it’s intelligent, not because it’s broken. Copilot checks its own answers; when retrieval feeds it the wrong context, it reformulates and retrieves again — and every retry costs time and tokens. Sometimes the loop converges late (slow); sometimes it gives up and answers from the wrong file (inaccurate). The control that causes it lives in the agentic layer — one layer too late. Aspected adds inspect & control at the retrieval layer, so Copilot is fed the right context the first time: accurate at the first attempt, in one go, and the agentic layer has far less to do.
Because your token bill is a retrieval problem. Control lives in the agentic layer, checking answers after the fact — and every failed check triggers another retrieval, another generation, another check, all paid in tokens. Aspected adds control at the retrieval layer, so the answer is grounded at the first attempt: right the first time, for a fraction of the tokens.
No — you add it. Aspected runs alongside your existing vector database from day one; nothing is migrated or decommissioned to start. Many teams later route primary retrieval to Aspected after comparing results side by side — but that’s a choice you make with evidence, not a prerequisite.
No — and after 22 years of content migrations, we say that with some authority. Outdated and near-duplicate documents demote themselves in the index, because recency and freshness are ranked aspects rather than problems to scrub away by hand. Cleaning was never really a data problem: when retrieval can’t weigh signals, a hard filter or a top-k on one blended score is a blind cutoff — and cleaning is reorganizing your content until that cutoff happens to cut in the right place. An aspect is a dial, not a gate, so the cutoff disappears — and with it the reason cleaning was mandatory. Cleanup is still worth doing to cut storage costs and reduce compliance risk — it’s just no longer the reason to postpone AI. Your content is AI-ready as it stands — aspects make it so.
No, and we won’t pretend otherwise. The core engine is patented and closed. The prep pipeline is free — not open source — client libraries and the aspect-schema format are freely available, and the 30-day free trial is the way in.
The Aspected Database — what you buy: priced by capacity, per month, running in your own Azure or AWS tenant — from €400/month for up to 1 TB of raw source content; 10 TB runs €3,000/month, and each additional terabyte adds a little — the rate per terabyte falls as volume grows (full table on the pricing page). Transactable on the Azure and AWS Marketplaces through your existing cloud agreement — with a 30-day free trial up to Database M (5 TB); cloud runtime and embedding run in your own tenant at your provider’s rates. And Cockpit, the chat window on top — licensed per user, on top of the database: €10 per user per month, humans and agents alike. Coming as of Q4 2026. From September 7, the listing is live on AWS Marketplace — Aspected Database — and the 30-day trial starts in your own account, no sales call required.
No. Aspected runs in your VPC or fully on-premises — TLS, bearer-token authentication, and ACL enforcement are first-class. Sovereignty isn’t a special mode; it’s a deployment choice.
11 · YOUR TURN
Your ten worst queries.
Thirty minutes. Side by side.
You’ve watched control being added at the retrieval layer, one question at a time. Now run yours. If adding Aspected doesn’t improve your results on your own content, tell us — we want to see where the aspect model breaks. If it does, you’ll have found out the way our best customers did.