10decodersYour answers are already written down.
They're just impossible to find.
Most of what your team needs already lives in your documents, tickets, and wikis. We build a Retrieval-Augmented Generation knowledge base that reads all of it and answers in plain language, with a citation back to the source. Run it on its own, or inside Microsoft Copilot, Salesforce, or Shopify.
Trusted by leading enterprises and healthcare teams
Static documents are where knowledge goes to sit.
A search box that matches keywords will find a document. It won't read forty of them and give you the answer. So people ask a colleague, wait, or guess. Here is where that shows up.
The answer exists, nobody can reach it
Policies, contracts, past tickets and specs are spread across drives and tools. Finding the right passage means knowing where to look first.
Experts become the search engine
When search fails, people interrupt whoever knows. That person becomes a bottleneck, and their knowledge never gets written down for the next person.
New hires take months to get useful
Onboarding is mostly learning where things are kept and who to ask. A knowledge layer answers those questions on day one instead of month three.
A knowledge base that reads, then answers.
Retrieval-Augmented Generation does two jobs. It retrieves the passages that actually relate to a question, then it writes an answer grounded only in those passages, with a link back to the source so anyone can check the work. Nothing invented, nothing from outside your walls.
Connect your sources
Documents, wikis, ticket history, databases, and shared drives. We meet the content where it already lives.
Structure and index
We clean and chunk the content, then build a searchable index so retrieval returns the right passage, not the closest keyword.
Answer with citations
Ask a question in plain language. The system pulls the relevant sources and drafts a grounded answer you can trace back.
Deploy where people work
Ship it as a standalone co-pilot or embed it inside the tools your team already opens every day.
Everything the pilot needs to reach production.
A Foundation engagement isn't a demo you have to productionize later. It ships with the pieces that make a knowledge base safe to put in front of real users.
Source connectors
Ingestion from your drives, wikis, ticketing, and databases, with a repeatable pipeline for new content.
Data cleanup and structuring
We deduplicate, chunk, and tag content so retrieval stays accurate as the corpus grows.
Retrieval and grounding layer
The core RAG engine that pulls relevant passages and keeps every answer tied to a citation.
Access controls
Permissions follow your existing rules, so people only get answers from content they're allowed to see.
Deployment into your tools
A standalone co-pilot, or an embed inside Copilot, Salesforce, Shopify, or an internal app.
Evaluation and handover
We test answer quality against real questions, then hand over documentation your team can run with.
Same engine, different jobs.
A grounded knowledge layer pays off anywhere people burn time searching. These are the starting points we see most, including in the regulated industries we work in.
Answer desk for agents
Front-line teams get a cited answer from your product docs and past tickets instead of escalating or guessing. Faster resolution, fewer repeat questions.
Clinical and policy lookup
Staff query protocols, coverage rules, and prior documentation in plain language, with the source attached so the answer can be verified before it's used.
Policy, contract & compliance search
Analysts pull the right clause from long agreements and policy libraries in seconds, with a citation for the audit trail.
Onboarding co-pilot
New hires ask the knowledge base instead of interrupting a senior colleague, and get answers grounded in your real documents from day one.
Standalone, or inside the tools you already run.
You don't have to move your data or replace a platform to get value from it. We deploy the knowledge layer on its own, or wire it into the systems your team lives in.
The three questions every team asks first.
A knowledge base only gets used if people trust the answers and security signs off. We design for both from the start.
Grounded, with citations
Answers are built only from retrieved passages, and each one links back to its source. If the content isn't there, the system says so instead of inventing an answer.
Deployed into your environment
We architect deployments to keep your content inside your own environment and honor your access rules. Deployed onto dedicated enterprise cloud instances.
Permissions you already have
Retrieval respects your existing access controls, so an answer never surfaces content the person asking isn't cleared to read.
Time back, and answers people trust.
The point of a knowledge base is a shorter path from question to answer, measured in hours saved and questions that no longer land on a person's desk.
Five questions, answered in order.
A technical evaluation of a knowledge system comes down to five questions. Each one has a section below written to settle it rather than to describe a feature — so you can read only the ones you are being asked about internally.
- 01What we can read →
Can it read what we actually have?
Scans, tables and images alongside clean documents — and the list of pages we could not read, handed back rather than hidden.
- 02Who sees what →
Will it show the wrong person the wrong thing?
Permissions carried on every passage and applied before search runs, so restricted content is never a candidate in the first place.
- 03How we prove it →
How would we know it works?
A graded question set built with your experts, including questions the content cannot answer, and a scorecard your team can re-run without us.
- 04Keeping it current →
What happens on day ninety?
Detecting change without re-reading everything, retiring withdrawn material, and citations that stay pinned to the version that was live.
- 05Where it runs, what you own →
What do we hold at the end of it?
Where it runs, how the model and index get replaced later, and who owns the source when the engagement closes.
All five depend on one thing first: whether the right passage gets found at all. No configuration downstream recovers from a retrieval miss, so that is where we start — and it is the section immediately below. If a question you need answered is not on this list, it is a gap and we would rather hear about it.
Answer quality is a retrieval problem.
If the right passage never reaches the model, no amount of prompt work will rescue the answer. So the argument worth having is about retrieval — not about the chat window. Ours runs in five stages, and each one is measured separately.
Narrow the pool before ranking begins
Entitlements and metadata — document status, effective date, business unit, jurisdiction — are applied before search runs. Filtering first is what makes "only current policy" and "only what this person may see" reliable rather than likely.
Search for meaning and for exact terms, together
Two retrieval methods run in parallel and are combined: one that understands what a question means, one that matches literal strings — policy numbers, clause references, error codes, drug and product names. Meaning-only retrieval fails predictably on identifiers, and identifiers are usually what enterprise questions turn on.
Re-score the candidates against the actual question
The shortlist is reordered by how well each passage answers what was asked, rather than by how close it sits in vector space. This is consistently the largest measurable gain in answer quality — and the stage most demonstrations skip, because it changes the numbers without changing the interface.
Answer only from what was retrieved, and cite it precisely
Every claim carries a reference to the document, the version and the page it came from — not a footnote pointing at a file and leaving the reader to search it. A reviewer can open the source and land on the sentence.
Check the answer before returning it
The draft is verified against the retrieved passages. Claims without support are removed. Where support is too weak, the system returns what it did find and says it cannot answer — recorded as a measurable rate, not a silent guess.
The specific models, index configuration and scoring method are documented in the architecture pack we share during evaluation. We publish them to you rather than to the internet — and because they sit behind an interface, changing any of them later is a configuration decision, not a rebuild.
Most knowledge projects fail at the scanner, not the model.
A real corpus is never clean. It is current documents sitting beside fifteen-year-old scans, tables that carry the actual answer, and diagrams with no text in them at all. Here is what we do with each — and, more usefully, what still goes wrong.
| Content | What we do with it | Where it still breaks |
|---|---|---|
| Digital documents | Text extracted with reading order and structure intact, so a passage keeps track of which section it belongs to. | Dense multi-column layouts can interleave. Confidence is scored and weak pages are flagged rather than indexed quietly. |
| Scanned and image-only pages | Converted to text with a confidence score kept against every page and carried through to the citation, so you can see how the text was obtained. | Handwriting, faded generations and stamped overlays fall below usable confidence. Those pages are quarantined and listed. |
| Tables | Preserved as structured rows rather than flattened into prose, and indexed so both literal and meaning-based search can reach them. | Tables spanning pages with repeating headers and merged cells need a rule set per corpus. That is tuned during ingestion, not at query time. |
| Images and diagrams | Described, indexed on the description, and returned alongside the citation so a person can check what the description claims. | A diagram whose meaning lives in unlabelled geometry is not answerable. We say so rather than approximate it. |
| Spreadsheets | Headers and units preserved; values indexed so numeric ranges can be filtered. | Calculated values are indexed as results, not as the logic that produced them. |
| Mail and attachments | Threads reconstructed and attachments treated as documents in their own right, linked back to the conversation. | Repeated quoted text distorts relevance. Removed at the passage level. |
| Tickets and case history | Problem, resolution and outcome separated, so a case closed without a fix is not served as an answer. | Cases whose real resolution lives in a linked chat and never in the record remain a known gap. |
We run this against your documents in the first week and hand you the quarantine list. The count of what we could not read is a deliverable — not something you discover in month three.
Two people, same question, different answers.
A knowledge system that ignores permissions is a data breach with a friendly interface. Access control here is not a setting on the front end — it is a property of every passage in the index.
Permissions travel with the content
Every passage inherits the access rules of the document it came from, at the moment it is indexed. Nothing is re-derived later or maintained by hand.
Identity comes from your directory
People authenticate through the identity system you already run, and their group membership is resolved when they ask — not when the index was built.
Excluded before retrieval, not removed after
Content a person may not see never enters the candidate pool, so there is nothing to strip out of an answer. Filtering a generated response is redaction. Never retrieving it is access control.
Revocation takes effect on the next sync
When someone changes role or a document is reclassified, entitlements follow. Withdrawing access does not require rebuilding the index.
Every question leaves a record
Who asked, what was retrieved, what was withheld, what was answered and which version was cited. This is the artefact an auditor asks for and most pilots cannot produce.
Demonstrated the way it is tested
Two accounts with different entitlements, one question, and the difference visible on screen. Included in the evaluation engagement.
We hand over a score, not an impression.
"It works well" is not a finding. Before handover we build a graded question set with your subject-matter experts and run it as a regression suite against every change — to the content, the configuration or the model.
Between 50 and 100 questions, minimum
Written by people who already know the right answer. What goes into it matters as much as how many:
- Questions answerable from a single passage
- Questions needing several documents combined
- Questions whose answer sits in a table or on a scanned page
- Near misses, where a plausible but wrong document exists
- Questions the content cannot answer — usually a fifth of the set, because knowing when to decline is a behaviour that has to be tested
Scorecard deliverables
The scorecard is a deliverable. It ships with the system, it is re-run after every change, and your team can run it without us. We would rather you held a number you can re-verify than a testimonial you cannot.
The policy changed on Tuesday. The answer changes on Tuesday.
Every knowledge system is accurate on the day it goes live. What separates a pilot from a production system is what happens on day ninety.
Changes are detected, not re-crawled
Sources report what moved and we verify it, so nothing depends on scanning the entire library on a schedule.
Updates cost what the edit costs
Only the affected sections are reprocessed. A single amended clause does not trigger the bill for reprocessing the whole corpus.
Withdrawn content stops being an answer
Superseded material is removed from retrieval immediately — the failure mode regulated teams worry about most.
Citations pin the version
An answer given in March cites the March document, and the audit trail survives the April revision.
Deletion is honoured
Removal at source means removal from the system, including anything derived from it.
Refresh timing is contractual
Stated per source type as a commitment you can hold us to, rather than described in adjectives. Contractual SLA
You own the system. You can change what powers it.
The reason to engage an engineering firm rather than license a platform is that at the end of it, you hold the asset. We build to that outcome deliberately, and the commercial terms follow it.
Runs where your data is allowed to be
Inside your own cloud tenancy on the major providers, or on open models hosted in your private environment where content cannot leave the building at all. Residency, tenancy and retention are decisions you make and we implement — not defaults you inherit from us.
The engine is replaceable by design
The language model and the search index sit behind an interface, so replacing either is a configuration change followed by a re-run of your evaluation set. And we do re-run it, so you see exactly what moved rather than being told the two are equivalent.
Delivered into your repositories
Source, configuration, ingestion pipelines, evaluation sets and documentation. Intellectual property assigns to you on delivery. There is no runtime licence, no fee per person to keep it running, and no component only we can maintain. Full IP Assignment
Exit is part of the plan
Handover is complete when your engineers have run the evaluation suite themselves and reproduced the score — not when the invoice clears. If you take it in-house afterwards, that was always the intended ending.
Three steps, each with a price before it starts.
Nobody should sign a build on the strength of a demonstration. So the first commitment is small, fixed, and produces something useful whether or not you continue with us.
Assessment
Your content, your questions, your access rules. You end the engagement holding a working system on a slice of real data, the evaluation scorecard, the list of documents we could not read, and a running-cost model filled in with your volumes.
Build
Fixed scope, fixed price, agreed against the scorecard the assessment produced — so the target is a number both sides already understand rather than a description both sides can reinterpret. Delivered into your repositories.
Operate
Monthly, cancellable, and genuinely optional. Regression runs when content or models change, ingestion health, quality monitoring and a quarterly review of what the system is being asked and where it is declining to answer.
Four rules we hold ourselves to on price
Never per person, never per question
You own the system, so charging you for using it would contradict the thing we just sold you. Engagements are priced by the work, not by your success with the result.
Build cost and running cost quoted separately
They behave differently and they land on different budgets. Bundling them is how organisations get surprised in year two.
Cloud and model consumption goes to your account
Billed directly by your provider at your negotiated rate. We do not resell compute and we do not mark it up, which means our advice on what to run is not a function of what we earn on it.
Changes are priced before the work happens
Scope moves in every real project. It gets estimated and agreed first, so the invoice never contains a conversation you have not had.
What we will not do is quote a build price from a requirements document alone. Until someone has looked at your actual documents, any number is a guess dressed as a proposal — and the gap between the two is where projects fail.
Work out the return before you talk to anyone.
Most vendor calculators book every recovered minute as cash. That is not how organisations work, so this one asks you what share of the time you would actually convert — and defaults it low. Change the inputs; nothing is sent anywhere.
Salary plus employment costs, divided by working hours. Not the billing rate.
Count both sides: the person hunting, and the colleague interrupted.
Recovered minutes only become value if the time is redeployed or a hire is avoided. Scattered minutes across a week rarely are. We default this to 40%.
Working year assumed at 220 days.
Recovered time is only worth this if it is redeployed or a hire is avoided. Where it is not, treat the figure as capacity created rather than cost removed.
Pressure-test these numbers with us →Treat the output as a range, not a forecast. The two inputs that move it most — minutes lost and the share you convert — are the two nobody measures before starting. The assessment engagement measures both against your own content, which is usually the first honest number an organisation has on this.
Built by a team enterprises already trust.
Our engineers have shipped AI and platform work referenced by names your board will recognize, in regulated industries where a wrong answer has consequences.
Before you book the call.
Common questions about Retrieval-Augmented Generation (RAG), document ingestion, security, and setup timelines.
