Knowledge Graph
The rulebooks every agent follows
The Knowledge Graph is the foundation of Fontana's agents. It grounds each agent decision in your firm's established policies, procedures, standards, and prior decisions. Each action is evidenced by auditable provenance: a citation to the file, the section, and the passage.
It allows you to:
Investigate faster
Find the relevant policy and see how similar cases were resolved.
Protect sensitive information
Agents can only retrieve information they have permission to access.
Verify answers at source
Trace each figure back to the document it came from.
How data is parsed and stored
Fontana can parse most document types used in financial services, from clean data files to messy MS Excel and Word documents, even scanned PDFs. Documents become multi-dimensional graph data, so an agent can explore and find what it needs by moving between those representations.
- Data files
- Excel
- Word
- Scanned PDF
Structural
How folders, files, and file contents are organised, including the folder structure, page index, headings, and subheadings inside each document.
Semantic
What the documents say about the world: the entities and the typed relationships between them. People, organisations, rules, products, and the links that join them.
Ontology
The schema that governs that world. Every entity rolls up to the type it belongs to, under a policy you set: strict, logged, or free to evolve as the corpus grows.
Vector
How passages are related by similarity. Embeddings show which documents read alike, so related material surfaces as a set.
Provenance
Where each fact came from. Every entity traces back to the exact section that evidences it, with a citation an agent can open.
How data is retrieved
The knowledge graph gives an agent a structured way to navigate, query, and request computations on the data.
Hybrid search
One call fuses full-text search, vector similarity, and graph expansion, then re-ranks the hits into a single result. Co-occurrence joins that fusion when the question depends on entities that appear together.
OpenCypher
Queries traverse the graph and return a fully auditable, immutable result, so the evidence and citations can be recorded in each conversation.
SQL
SELECT statements let the agent run computations, filters, and aggregations at query time, recorded as immutable proof behind each answer.
Python and JavaScript
The agent can run sandboxed code against the knowledge graph to perform complex computations. Every call is recorded as immutable evidence.
Entity search
Ontology entities can be found by name or by similarity, so the agent can jump between documents and follow connected data.
Structural search
Navigate a file's outline, and move between sections, neighbouring passages, and similar files.
