TECHNOLOGY · ONTOLOGIES
Not just search.A model of your finance.
An ontology is a machine-readable map of a domain: the things that exist, their attributes and the rules that connect them. Amaril builds one for finance, so every answer is grounded, connected and cited.
01 · DEFINITION
What an ontology is, simply.
A formal, shared model that describes the things in a domain and the links between them. Not just a list of words, but their meaning and the logical rules that govern them. It answers: what kinds of things exist? What attributes do they have? How are they connected? Different people, and different software, read it the same way.
02 · ANATOMY
Four building blocks
Four elements describe almost anything.
Classes
The categories of things: Issuer, Counterparty, Instrument, Covenant. The "types" of the domain.
Instances
The concrete cases of a class: "Intesa Sanpaolo" is an Issuer, a specific loan is an Instrument.
Relations
The links between things: "is-counterparty-of", "secures", "is-governed-by". The real value.
Axioms
Logical rules: a loan has at least one covenant; a covenant binds exactly one borrower.
03 · THE FORMAT
From a triple to a graph
Knowledge is stored as minimal three-part facts: subject, predicate, object. Millions of connected triples form a graph you can navigate.
SUBJECT
Loan A
PREDICATE
is-secured-by
OBJECT
Pledge B
From one node you can hop to others: it is the multi-hop reasoning a flat list of paragraphs cannot do.
04 · WHY IT MATTERS
An LLM aloneisn't enough on your data.
Powerful on language, fragile on specific facts.
/ 01
It hallucinates
Plausible but invented answers, with no anchor to real facts.
/ 02
No sources
Hard to know where a statement comes from, or to verify it.
/ 03
Weak on links
Multi-hop questions over your internal data trip it up.
The turning point: LLMs remove the friction, the ontology keeps the structure. They are complementary, not competitors. Fluid statistics plus verifiable logic, the core of neuro-symbolic AI.
05 · HOW AMARIL USES THEM
A financial ontology,built from your documents.
Four ways the ontology and the LLM reinforce each other inside Amaril.
01
Amaril builds the ontology
The LLM reads financial statements, filings, contracts and data rooms and extracts entities (issuers, counterparties, instruments, covenants, KPIs) and relations. It reconciles name variants ("Intesa Sanpaolo", "ISP", the counterparty) into one entity. A human validates: the LLM accelerates, it does not replace judgement.
02
The ontology grounds the LLM
Every term maps to a concept with a unique identifier. Generation is constrained to entities and relations that actually exist in the graph: fewer hallucinations, unambiguous terms, and every answer cited to document, page and paragraph.
03
GraphRAG, not just similarity
Retrieval runs over the financial graph, not only over "similar" paragraphs. So Amaril answers multi-hop questions, deducing links that are never written explicitly, e.g. "which counterparties are exposed to client X through the covenants of loan Y?"
04
Neuro-symbolic reasoning
The LLM interprets the question and orchestrates queries over the graph; logical reasoners verify the ontology constraints (covenants, accounting rules). The answer joins the fluency of language with the verifiable solidity of symbols.
06 · ARCHITECTURE
Three layers that reinforce each other.
Amaril does not rely on a single LLM: it combines the fluency of language, the solidity of structure and the predictability of search.
01
Neural
LLMs read the documents, extract entities and answer in natural language.
02
Symbolic
Canonical entity types and their relations form the financial domain ontology.
03
Deterministic
A query planner resolves many questions without inference, where the LLM is not needed: fast, predictable, low cost.
Neuro-symbolic AI, in production.
07 · THE PIPELINE
From document to entities, with sources.
A dedicated worker processes each file in the background, in two phases.
01
Document
PDF, Office, text or image.
02
Chunks
Passages sampled from the file.
03
Phase 1 · Regex
VAT, IBAN, ISIN, EUR amounts, dates, with real validators (Luhn, mod-97).
04
Phase 2 · LLM
One call, JSON, temperature 0.1: issuers, counterparties, roles, deadlines, values. It also checks Phase 1 false positives.
05
Graph
Entities and relations stored, ready for search and the agent.
The entities enter the search index: "contracts over EUR 50M" becomes a filter, not just "similar". A reranker puts the best passages on top before the agent.
08 · WHAT YOU GAIN
The sum is worth more than the parts.
01
Fewer hallucinations
Answers are bound to facts verified in the graph.
02
Multi-hop reasoning
Complex questions answered by following chains of relations.
03
Consistency & explainability
Reasoners and constraints keep data consistent and answers traceable to source.
04
Fast to build
LLMs cut construction from months of manual work to days.
90%+
GraphRAG accuracy on schema-bound queries where vector-only RAG can fail.
Days
Build time. What took months of NLP experts now starts in days.
Cited
Every answer carries document, page and paragraph of origin.
Indicative figures from 2024-2025 industry reviews and benchmarks; they depend on the implementation.