Mapping links the same person, company, policy, vehicle or asset across your whole document estate, reconciles the values those documents disagree on, and builds a relationship graph you can query from your own systems. Every link keeps the evidence that justified it.
Resolution thresholds are yours to set. The precision and recall achieved on your own entity population is measured during the Proof of Value, on your documents, against a human-adjudicated sample.
Extraction tells you what a document says. Mapping tells you whether forty documents are talking about the same thing — which is the question most operational decisions actually depend on.
The claimant on a form, the holder on a policy schedule and the payee on an invoice are resolved to one entity, with each mention retained as a separate observation rather than merged away.
Mohammed, Muhammad, Mohamad and محمد are treated as candidate renderings of one name, then confirmed or rejected against corroborating identifiers rather than accepted on spelling alone.
A claim is checked against the policy that should cover it and the invoice that should support it, field by field, with the source of each value retained on both sides.
Conflicting values, near-duplicate submissions and resubmitted documents with one changed digit are surfaced as findings with both versions shown, not resolved silently.
Entities and the relationships between them form a graph spanning every document you hold, so you can ask which policies touch one company, or which claims share a repairer.
Each link records its creator, timestamp, active rules and thresholds, source documents and the coordinates of the values that supported it, including human confirmations and rejections.
Three documents, produced by three organisations, describing one event. Nothing in them is formatted the same way.
Identification has produced typed values with page numbers, bounding boxes and confidence scores from the claim form, the policy schedule and the repair invoice. Mapping never works from loose text.
Names, identifiers, plate numbers, IBANs and addresses are normalised into comparable forms — including Arabic-Indic numerals and three transliterations of one family name — and blocked into candidate sets.
A name match alone is not enough. A matching Emirates ID number, a shared date of birth and a consistent address push the combined score past your threshold; a name match with a conflicting identifier does not.
Coverage dates against the incident date, the vehicle on the schedule against the vehicle repaired, the invoice total against the policy limit and against the sum of its own line items.
The invoice quotes a chassis number that differs from the schedule by two characters. That becomes a mismatch record naming both documents, both values and both coordinate sets, routed to the adjuster.
Entities, links, findings and evidence land in your own database and are available over your own API, ready for the case management system, the fraud team and the data warehouse.
One person in a UAE document estate is routinely spelled four ways in Latin script and once in Arabic script, across an Emirates ID, a bank statement, a policy schedule and a handwritten form. A resolution engine tuned on English names treats them as four customers, and every downstream count is wrong.
This is the difference between an entity count you can report and an entity count you have to apologise for.
The two things organisations reach for instead are manual reconciliation and a vector index. One does not scale, and the other has no idea what an entity is.
What the system reasons about
Arabic name variants
Documents comparable at once
Conflicting values between documents
Near-duplicate detection
Auditability of a decision
Where the result lives
Where processing happens
Retrieval-augmented generation is genuinely useful, and DocxIntel ships it as the Ask capability. The point is narrower: a similarity index is not an entity model, and asking it to reconcile records is asking the wrong component. Verify entity-resolution claims against current vendor documentation before making a decision.
What the mapping service models, what it resolves on, and what it hands back to your systems.
Entity model, resolution thresholds and integration points for a specific deployment are agreed in writing during the Proof of Value, against your own entity population rather than a generic schema.
What gets asked once someone realises the duplicate rate in the archive is not what the report said.
Resolution combines several signals rather than trusting string similarity alone. Arabic transliteration variants, honorifics, name-order differences and abbreviated middle names are normalised, then weighed against corroborating identifiers such as an Emirates ID number, a date of birth, an IBAN or a shared address. A link is only asserted when the combined evidence clears the threshold you set, and the evidence is stored with the link.
