Capabilities

Information, semantics and knowledge

Turning scattered sources into information that is structured, connected and usable by people and systems.

On one page

We work with ingestion, structured extraction, entities, relationships, taxonomies, graphs, semantics, hybrid search, RAG, embeddings, reranking and memory. We design representations that make it possible to find, relate and reuse knowledge while preserving context, provenance and meaning.

01Ingestion and structure. We build pipelines for documents, PDFs, spreadsheets, databases, APIs, web pages, messages, repositories, code, configurations, images, audio, video and events.
02Entities and relationships. We identify people, organisations, products, documents, operations, systems, concepts, events, figures, rules, decisions, states and questions.
03Domain models. A domain model describes vocabulary, entities, relationships, hierarchies, states, rules, constraints, invariants and exceptions.

Systems receive information in different forms: documents, tables, messages, pages, databases, code, images, audio, video, events and explanations from people.

Each form carries its own structure and context.

Semantics connects data with the domain. It allows «customer», «case», «status», «completion», «risk» or «current version» to have a meaning shared widely enough to be able to query, decide and act.

Ingestion and structure

We build pipelines for documents, PDFs, spreadsheets, databases, APIs, web pages, messages, repositories, code, configurations, images, audio, video and events.

The originals and their metadata are preserved. Extraction produces objects with their source location and validation status.

Entities and relationships

We identify people, organisations, products, documents, operations, systems, concepts, events, figures, rules, decisions, states and questions.

Resolution relates aliases and versions, and keeps apart those objects whose meaning or scope differs.

3.14 — EVIDENCE BEFORE ELOQUENCE

FIG. 02 — GROUNDED CLAIMS

FORM 10-K · FISCAL YEARp.47
total revenue $412.3M · +12.1% YoY
SEC EDGAR

HEADCOUNT REPORT · Q4p.3
FTE 1,284 → 1,291 · +0.5%
HRIS

query: churn rate
Ø no matching source in corpus

Revenue grew 12% year over year,while headcount stayed flat,UNSUPPORTED — NO SOURCE THREADand churn fell to a record low.— no public evidence.

PRINCIPLE 01 — EVERY CLAIM KEEPS A THREAD TO ITS SOURCE, OR SAYS SO

Domain models

A domain model describes vocabulary, entities, relationships, hierarchies, states, rules, constraints, invariants and exceptions.

It can be expressed through schemas, taxonomies, graphs and KIR, the typed intermediate representation from which it is projected towards each use.

The representation is developed with specialists and evolves with the evidence.

Graphs and logic

Graphs make it possible to navigate relationships that linear text hides.

For example:

SOURCE → CLAIM → ENTITY → RULE → DECISION → OPERATION

or:

QUESTION → EVIDENCE → HYPOTHESIS → DECISION → WORK

Rules and deductions are kept explicit when they need to be reproducible and reviewable.

Hybrid retrieval and RAG

A RAG system relates retrieval, context, evidence, permissions and composition in order to answer in an examinable way.

We combine exact, lexical, semantic, structured and relational search. The system can retrieve an identifier, a passage, an entity, a version, a relationship, a rule, a code dependency, a decision or a set of pieces of evidence.

Reranking orders the shortlist. Expansion adds related context. Composition uses only permitted evidence and preserves citations.

Memory

Memory organises decisions, states, questions, results, relationships, learning and temporal context.

We design policies for scope, version, currency and forgetting. Memory preserves objectives, decisions, states and dependencies with explicit policies, so that it helps to carry on without dragging along obsolete context.

Applications

informational intelligence;
research;
reporting;
business logic;
localisation;
project memory;
agent context;
documentation;
corporate search;
code analysis;
decision support.

Evaluation

coverage;
extraction accuracy;
entity resolution;
retrieval;
citation support;
contradiction detection;
abstention;
updating;
usefulness.

Closing

Information becomes knowledge when it preserves meaning, relationships, provenance and a clear way of entering the work.

What this capability brings together

01preserving sources
02extracting facts and relationships
03resolving entities
04defining concepts
05representing logic
06maintaining versions
07retrieving evidence
08detecting contradictions
09delivering knowledge to people, models, agents and applications