AI Search
Search that finds the document even when you do not know its name.
What is AI search?
AI search indexes your internal documents and returns results based on meaning rather than exact word matching, searching across drives, email and shared folders at once while respecting the permissions already set on each file.
Ordinary search matches words. If the document says vehicle and you typed car, you get nothing, and the file you needed sits three folders away.
Meaning-based search closes that gap. Ask for the maintenance agreement from last year and it finds it, regardless of what somebody named the file when they saved it.
It searches your drives, your email and your shared folders together, and shows each person only what they already have permission to open.
Where we enforce permissions
Duara deploys internal search from Nairobi for organisations where finding a document has become a recognised time cost rather than an occasional annoyance.
We enforce permissions at the index rather than in the interface. Search that filters results after retrieval is search that can leak, and we would rather spend an extra week on the permission model than hand over something that occasionally shows the wrong person a salary letter.
Who this is for
- Staff waste time hunting for documents
- Files are spread across drives and email
- Knowledge left when a person did
- People ask colleagues rather than searching
Industries we serve
- Professional services firms
- Legal practices
- Engineering and construction
- Healthcare administration
- Financial services
- Public sector bodies
Problems this solves
These are the conditions where search becomes worth building rather than tolerating.
- Keyword search that finds nothing. The document uses different words from the ones you typed, so it may as well not exist.
- Files scattered across systems. Drives, email attachments and shared folders each need searching separately, so most people search one and give up.
- Asking a colleague instead. It is faster to interrupt somebody than to search, which means the same questions circulate endlessly.
- Knowledge that left with a person. When somebody resigns, what they knew about where things are and why they were done goes with them.
What we deliver
One search across every store, returning results by meaning, enforcing existing permissions at the index, and quoting the passage that answered with a link to the source.
One search across drives, email and shared folders
Results ranked by meaning rather than exact words
Existing file permissions enforced throughout
Answers quoted with a link to the source document
Reporting on what people search for and do not find
Coverage of scanned documents through text extraction
Features
- Meaning-based ranking. A search for car finds the document about vehicles, which is the whole point and what keyword search cannot do.
- Permissions enforced at the index. Access is applied when results are retrieved, not filtered afterwards, so nothing can leak through a ranking quirk.
- Quoted passages with sources. You see the sentence that matched and a link to the file, so verification takes seconds.
- Scanned document coverage. Text is extracted from scans and photographs so paper-origin documents are searchable too.
- Failed search reporting. What people looked for and did not find, which tells you exactly which documents are missing or misfiled.
Benefits
- Documents get found in seconds. The recurring cost of hunting for files disappears, and it is usually larger than anybody has measured.
- Fewer interruptions. Staff stop asking colleagues where things are, which returns time to both people.
- Institutional memory survives. What the business knows stops depending on who still works there.
- Misfiling becomes visible. Failed search reporting shows which documents people expect to exist and cannot locate.
Technologies we use
- Semantic indexing. Documents indexed by meaning so conceptually related material is found regardless of vocabulary.
- Permission mirroring. Access rules read from your file systems and enforced at retrieval rather than applied after the fact.
- Text extraction from scans. Optical reading so paper-origin documents join the searchable set rather than sitting outside it.
- Multi-source connectors. Google Drive, SharePoint, OneDrive, network shares and mail systems indexed together.
- Incremental indexing. New and changed files picked up continuously rather than through a nightly rebuild.
The permission model is the hard part, not the search quality. Getting retrieval right takes days; getting access right takes weeks and matters more.
How the process works, step by step
Internal search typically takes five to seven weeks, with the permission model consuming most of that time.
-
Step 1
Free discovery review
We establish where documents live and roughly how much time is currently lost finding them.
-
Step 2
Permission mapping
We map who may see what across every store, which is the slow part and the part that must be right.
-
Step 3
Indexing and extraction
Documents are indexed, including scans, with progress visible as coverage grows.
-
Step 4
Permission testing
We test with accounts at different access levels to confirm nothing surfaces that should not.
-
Step 5
Rollout with feedback
Search opens up, with failed search reporting driving what gets fixed or filed properly.
Quality assurance and how we work
The commitments attached to a system that reads all your documents.
- Permissions are enforced at retrieval. Access is applied when results are fetched, never filtered afterwards, because post-filtering can leak.
- We test with real access levels. Before launch we search as accounts at several permission levels to confirm nothing appears that should not.
- Nothing is indexed without agreement. Which stores and folders are included is your decision, and sensitive areas can be excluded entirely.
- Documents stay where they are. Nothing is copied out of your systems, and indexed content is never used to train public AI models.
- Failed searches are reported honestly. What people could not find is surfaced rather than hidden, because it is the most useful signal the system produces.
Frequently asked questions about AI search
Can staff see things they should not?
No. Permissions are enforced at retrieval rather than filtered after the fact, and we test with accounts at several access levels before launch.
Which storage systems does it work with?
The common ones including Google Drive, SharePoint and OneDrive, plus network shares and mail systems.
How is this different from the search we already have?
Existing search matches words. This matches meaning, so a search for car finds the document about vehicles.
Does it search email as well as files?
Yes, and searching both together is frequently where the value appears, since attachments often live only in mail.
Can it read scanned documents?
Yes, through text extraction, so paper-origin documents join the searchable set rather than remaining invisible.
How long until search is available to staff?
Five to seven weeks, with the permission model taking most of that. Getting access right is slower than getting search right.
Why does permission work take so long?
Because it must be exact. Search that occasionally shows the wrong person a salary letter is worse than no search at all.
Can we exclude certain folders?
Yes. Which stores and folders are indexed is entirely your decision, and sensitive areas can be left out completely.
Does anything we index leave our network?
Documents stay in your systems. Indexed content is never used to train public AI models.
What is a failed search report?
A record of what people looked for and did not find. It usually reveals misfiled documents and gaps in what you actually hold.
Does it show why a result matched?
Yes, by quoting the passage that matched and linking to the source, so verification takes seconds rather than opening the file.
How current are results?
The index updates continuously as files change rather than rebuilding overnight, so new documents are findable quickly.
Will it slow down our file systems?
No. Indexing runs separately and is throttled so it does not compete with normal use.
Can it search documents in Swahili?
Yes, and mixed language documents, which are common in internal correspondence here.
What happens when someone leaves?
Their documents remain findable within existing permissions, which is much of the point of building this.
Does it replace our document management system?
No. It sits over whatever you use and does not require moving anything.
Can it answer questions rather than return files?
That is retrieval augmented generation, which is a related but separate build. Search finds; RAG reads and answers.
How many documents can it handle?
Volume is rarely the constraint. Permission complexity is what determines difficulty far more than document count.
Who should we assign to this work?
Whoever administers your file systems and whoever owns the sensitive areas. Both, because the permission map needs their input.
What does search cost to run?
Document volume and query rate determine it. We project the monthly cost from your own figures before anything is agreed.
Glossary of terms
Search vocabulary as used here.
- Semantic search
- Finding documents by meaning rather than by exact word matching.
- Indexing
- Processing documents in advance so they can be searched quickly afterwards.
- Permission mirroring
- Reflecting the access rules already set on your files so search cannot reveal more than a person could open.
- Optical character recognition
- Reading text from scans and photographs so those documents become searchable.
- Relevance ranking
- The order results are returned in, based on how well each matches the intent of the query.
- Incremental index
- Updating the search index continuously as files change rather than rebuilding it nightly.
Related services in AI Implementation
Talk to us about AI search
Ask three people how long they spent looking for a document last week. The answers usually make the case better than we can.
Start your project