--- title: "Architecture — How the Certant brain actually works" description: "Knowledge graph + hybrid retrieval, adaptive chunking, citations on every answer, deployable on cloud or air-gapped. The technical overview." url: "https://certant.ai/architecture" --- # How the brain actually works A knowledge graph and hybrid retrieval over your own documents, running anywhere from our cloud to a room with no network. [Talk to an architect](https://certant.ai/contact-sales) [Read the security overview](https://certant.ai/security) About eight minutes Written for CIOs and security architects Five layers, raw bytes to an answer The stack ## Five layers, one platform The Certant stack from raw bytes on disk to an answer in someone's hands. Hover a layer to see where it sits. 1. Ingestion ### Reads everything, keeps the structure PDFs, Word, slides, scans, images. We parse with the structure preserved: tables, sections and headings stay intact. OCR runs on scans. 2. The graph ### The graph builds itself Your entities, such as employer, member, contract and vendor, and the relationships between them, learned from your own vocabulary. No manual ontology. 3. Retrieval ### Hybrid by default Vector search finds answers by meaning. GraphRAG follows the relationships. Structured query hits your warehouse. We pick what is best per question. 4. Reasoning ### Every answer is grounded The citation engine attaches the source paragraph, and guardrails refuse rather than hallucinate. 5. Surfaces ### One brain, four surfaces Agent Builder, Analytics, Chatbots and Strata all read the same graph with the same retrieval and the same citations. What differs is the interface. Retrieval ## How retrieval actually works Vector retrieval ### Finds answers by meaning Every paragraph in your document gets converted into a numerical fingerprint. When someone asks a question, we fingerprint the question and find paragraphs that mean something similar, not just paragraphs containing the same words. Vectors work well on open-ended language: "what does our policy say about overtime?" They struggle with questions about how things relate, and most questions in an organisation are about how things relate. GraphRAG ### Walks the relationships Start from a named entity in the graph: a vendor, a contract, a person. Walk the edges and pull back the connected sub-graph as context. The model gets not just the relevant text, but everything connected to it. A question a vector index cannot answer alone: "which contracts does our largest vendor have with our highest-risk member entities?" That's a graph traversal, not a similarity search. Hybrid ### Combined at query time For every question, the model picks the strategy: vector for prose-heavy lookups, graph traversal for relationship questions, structured query for numbers, often a blend. You can see the path the answer took: ask why, and the sub-graph and the cited passages come back. Ingestion ## How Certant reads documents ### Adaptive chunking It reads a document the way a person does. Tables, sections and headings survive intact, and nothing is cut mid-clause. ### Multimodal native PDFs, Word documents, slides, scans and images. OCR runs where text isn't extractable, and page-image references survive into the citation engine. ### Auto-ontology It learns your vocabulary as it goes: "Worker", "Staff" and "Employee" become one entity, mapped to your schema, with no manual setup. The graph ## No black boxes Every answer traces back through this graph. Drag to pan, scroll to zoom, click any node. _[Diagram: Example knowledge graph]_ Every node is something in your business. Every edge is a relationship we learned from your data. Edges are typed (e.g. governed_by, employs, requires) and clickable. Models ## Bring your own, or use ours We don't lock you to a model. They change every few months, and a lock-in now is a lock-out later. For Cloud, you can pick from AWS Bedrock, Azure AI Foundry, GCP Vertex AI, or our defaults. We route per workload: reasoning models for hard questions, fast models for short answers, embedding models for retrieval. For Sovereign, you bring your own, open-source weights or a licensed vendor, and run them on your hardware. We support local GPUs (≥80 GB VRAM) and offline embedding models. | Where it runs | Model | Via | |---|---|---| | Cloud · reasoning | Claude Sonnet 4.5 | Bedrock · default | | Cloud · reasoning | GPT-5 | Azure AI Foundry | | Cloud · fast | Claude Haiku 4.5 | Bedrock | | Cloud · fast | Gemini 2.5 Flash | Vertex AI | | Sovereign | Llama 3.3 70B | On-prem · GPU | | Sovereign | Mistral Large 2 | On-prem · GPU | | Sovereign | Qwen 2.5 72B | On-prem · GPU | | Embedding | BGE-M3 / Cohere v3 | On-prem or cloud | Deployment ## Cloud to air-gapped, one product Managed ### Cloud - Hosted in our AU, EU and US regions - Tenant isolation at the data layer - SOC 2-aligned controls - Live in minutes Your cloud ### Inside your VPC - Deploys into your AWS, Azure or GCP account - Your VPC, your IAM, your KMS keys - Certant ships the software, you operate the cloud - Available via cloud marketplaces Air-gapped ### Inside your facility - No outbound connectivity required - Models, indexes, audit logs all local - Signed update bundles via removable media - Designed for IRAP / DEFCON / sovereign workloads Docker images and Kubernetes manifests are available for any deployment mode. This page is not a manual. If you need the YAML, ask deployment. [Talk to deployment](https://certant.ai/contact-sales) Citations ## Every answer cites its source When Certant answers a question, the answer carries the paragraph it came from. Click and you see the original document with the cited passage highlighted. Print it. Send it to your auditor. Nothing here asks you to take our word for it. AnswerExample **When does the new EBA require overtime to be paid at double time?** Under the 2024 EBA, overtime is paid at double time after the first three hours of overtime worked, or for any overtime performed on a Sunday or public holiday. EBA 2024 · page 14 · clause 7.2 "…ordinary overtime shall be paid at time-and-a-half for the first three hours and at double time thereafter. Overtime in excess of three hours, or worked on a Sunday or public holiday, shall be paid at double the ordinary hourly rate." Honesty ## What we don't do Foundation models We don't train base models. We make the ones you already trust useful on your own data. BI suites Analytics is one surface among several. If a pivot table is all you need, buy a pivot table. App builders Agent Builder builds workflows. It does not build customer-facing apps. ## Convinced or sceptical, talk to us We answer technical questions in plain English. [Talk to an architect](https://certant.ai/contact-sales) [Read the security overview](https://certant.ai/security) AnswerExample Under the 2024 EBA, overtime is paid at double time after the first three hours worked, or for any overtime on a Sunday or public holiday. EBA 2024 · page 14 · clause 7.2