Capability · Knowledge Bases

Your own reference repository, the source of truth Ceertia checks against.

Your files, your reference lists, your business registers: one reusable base, attached to the questions that need it. Ceertia confronts what it finds in a document with your own reference, and cites the exact row as proof.

The principle

A knowledge base does not only inform the analysis. It can settle it.

In many checks, the base is not there to give the AI context: it is the reference the answer must match. The list decides, the AI does not guess.

Your reference is authoritative

When the question is "does what I found appear in the list?", the lookup determines the answer. The AI extracts and normalises the value; it does not rule on its validity.

Attached where it matters

A base can serve a whole use case, or one precise question. Question-level attachment is what makes it usable: each question carries its own reference and its own matching rule.

Every answer cites its row

A verified answer shows the value extracted from the file and where it came from, the reference row retained, and the version and date of that reference.

Sources

Four kinds of source, one base each.

A base is created once and attached wherever it is useful. Its type determines what you can do with it.

Type 01Files

Reference documents: PDFs and other business files, ZIP archives holding several documents, or text typed straight into Ceertia.

  • Text extracted, split into segments, with a clear ready state
  • Extracted text and available segments can be inspected
  • Enrichment: doctrine, a procedure, a regulatory reference

Enrichment only · Version frozen at import

Type 02Reference lists

A structured list, imported or typed: certified people, approved companies, homologated equipment, authorised codes, applicable ceilings.

  • Usable columns: name, identifier, domain, validity dates, status
  • A declared matching key, and columns used as conditions
  • Verification: does what I found appear in the list?

Enrichment or verification · Dated version, re-imported manually or on a schedule

Type 03External registers

A structured source maintained outside Ceertia, queried live during the check. EPREL, the European register of energy-related products, is the concrete example.

  • Lookup by registration number, GTIN or EAN, or brand and model
  • Structured information pulled back and compared with the file
  • Enrichment or verification, depending on the attachment

Enrichment or verification · Real time, subject to source availability

Type 04Workflows

A workflow produces dynamic context when plain files are not enough: call an API, query a register, search documents, transform data, ask the AI.

  • Its result becomes an additional context source
  • Or the reference of a verification, when its output is structured
  • Recomputed at every run, with its own log

Enrichment, or verification if the output is structured · Recomputed at every run

Modes

Enrichment or verification, decided by the attachment.

The mode belongs to the attachment, not to the base. The same base can enrich one question and verify another.

Mode 01Enrichment

The base adds material to the analysed context. The AI reads it alongside the documents of the file and uses it to form its answer. The content informs; it imposes nothing.

  • For text, doctrine, a procedure, a history
  • Attached at use case level by default
  • The answer reads as reasoning, with its sources
Mode 02Verification

The base is authoritative. Ceertia extracts a precise value from the file, looks it up in the reference, and the result of that lookup determines the answer.

  • For certification, approval, homologation, membership of a closed list
  • Attached at question level by default
  • The answer reads as a confrontation, not as a paragraph
Attachment

Two levels, one precedence rule.

A base attached to a use case serves all of its questions. A base attached to a question serves that question alone, and wins over the use case attachment for it. A base attached at both levels is never counted twice in the context.

Question-level attachment carries its own configuration: the value to extract, the matching key, the conditions to check and the behaviour when nothing matches. Two questions can use the same list with different keys.

LevelScopeDefault mode
Use caseEvery question of the use caseEnrichment
QuestionThat question onlyVerification
Mechanics

From document to verdict, six steps.

This is the heart of the capability, and what the interface makes readable at every step.

01Extraction

The question declares the value to look for in the file: a name, a registration number, a certification number, an equipment reference. Ceertia keeps its exact source: document, page, passage.

02Normalisation

Case, accents, spacing, punctuation, number and date formats are brought to a comparable form. Without it, "DUPONT Jean-Marc" and "Jean Marc Dupont" never meet.

03Lookup

The normalised value is searched in the reference according to the key declared on the attachment: exact match, normalised match, or approximate match with a threshold.

04Conditions

The row found must still satisfy the declared conditions: certification valid at the date of the works, domain matching the operation checked, active status.

05Verdict

Compliant, non-compliant, or undetermined. Undetermined is a result in its own right, not a hidden failure.

06Evidence

The extracted value and its location, the reference row retained, the version and date of the reference, and the reason for the verdict. The final decision stays under human supervision.

Verdicts

Every case has a named outcome.

No situation produces a silent verdict, and no technical failure produces a non-compliant one.

SituationVerdictWhat the user sees
One match, conditions satisfiedCompliantThe reference row, the extracted value, its source in the file
One match, one condition failedNon-compliantThe reason, named explicitly: certification expired on a given date, domain not covered
No matchNon-compliant or to be checkedConfigured on the attachment. Default: to be checked, not non-compliant
Several matchesUndeterminedThe list of candidates, with human arbitration requested
Approximate match below the thresholdUndeterminedThe closest candidate, its distance, and a confirmation to give
Value absent from the documentUndeterminedData not found in the file: absence of proof is not proof of absence
External source unavailableUndeterminedAn explicit incident, never a fallback verdict
Reference past its declared freshnessVerdict returned, flaggedA visible warning on the answer and on the base
One non-negotiable rule

A technical failure never produces a non-compliant verdict. A file is not penalised because an API did not answer.

Worked example

Is this installer certified for this operation?

A base of certified installers is imported as a reference list, with company name, registration number, certification number, domain and validity dates. The registration number is declared as the primary key, the company name as the secondary key, and the base is attached to this one question in verification mode.

At check time, Ceertia extracts the company name, the registration number and the date of the works from the quote, normalises them, finds the row, and checks that the works fall inside the validity period and that the domain matches. If the registration number is missing from the quote, it falls back to the company name; if that only gives an approximate match, the question becomes undetermined with the candidate shown.

Verified answerlist v4 · 02/2026
Registration number found in the reference listquote p.1, zone B2 → row 1 482
Certification valid at the date of the works01/2024 to 12/2026 · works 14/03/2026
!Domain to confirm on a second operationapproximate match below threshold · arbitration requested
Stored with the answer
extracted valuesource zonereference rowlist versionarbitration
External registers

Live sources, in the same analysis.

An external register is queried during the check, from the references present in the documents. EPREL is the concrete case: the European register of regulatory information on energy-related products.

Lookup, several ways

From a registration number, a GTIN or EAN, or a reference and model with its brand.

Structured information, pulled back

What the register returns is added to the analysed context, alongside the documents of the file.

Field-by-field confrontation

Attached to a question in verification mode, the register produces a comparison between the quote and the register record.

Your rules still decide

The register informs the analysis. The final outcome stays determined by your use case and by Ceertia’s check rules.

Product rules

What the capability guarantees.

No automatic decision

The verdict is a proposal. Human supervision stays systematic.

Traceable to a row

Any answer from a verification traces back to a reference row and to a dated version of that reference.

Replayable

A re-run check can state which version of the reference was used.

Undetermined is first class

Undetermined is a state of its own in the interface, not a discreet variant of non-compliant.

An unavailable source is visible

It is shown, not guessed. A stale reference is flagged on the answer and on the base.

Frequently asked questions

Common questions, straight answers.

Does a knowledge base decide on its own?+
No. A verification produces a proposed verdict with its evidence: the extracted value, its location in the file, the reference row and the reference version. Human supervision stays systematic.
What is the difference between enrichment and verification?+
Enrichment adds material to the analysis: the content informs the answer without imposing it. Verification makes the base authoritative: Ceertia extracts a value, looks it up in your reference and the lookup determines the answer.
What happens when an external source is unavailable?+
The question returns undetermined with an explicit incident. A technical failure never produces a non-compliant verdict, so a file is never penalised because an API did not answer.
Can two questions use the same list differently?+
Yes. The configuration lives on the attachment, not on the base: each question declares its own value to extract, its own matching key, its own conditions and its own behaviour when there is no match.
Your reference data, cited row by row.

Create a base once, attach it to a use case or to a single question, and every answer it produces carries the row, the version and the date behind it.

Schedule a demo

Bring one of your reference lists.

In a 30-minute demo we import it, attach it to a real question, and run the check on one of your files.